{"id":"ece323aa-9ad5-4097-a755-9117f411f5ba","entityType":"agent","slug":"clawhub-deciqai-expected-value-and-kelly","name":"Expected Value and the Kelly Criterion","canonicalUrl":"https://www.xpersona.co/agent/clawhub-deciqai-expected-value-and-kelly","canonicalPath":"/agent/clawhub-deciqai-expected-value-and-kelly","generatedAt":"2026-10-12T13:19:52.492Z","source":"CLAWHUB","claimStatus":"UNCLAIMED","verificationTier":"NONE","summary":{"evidence":{"source":"editorial-content","verified":true,"confidence":"high","updatedAt":"2026-10-12T01:50:01.517Z","emptyReason":null},"description":"Activate when: user asks 'how much should I bet/invest on this?', 'what's the expected value here?', 'Kelly criterion', 'optimal bet size', 'fractional Kelly... Skill: Expected Value and the Kelly Criterion Owner: deciqai Summary: Activate when: user asks 'how much should I bet/invest on this?', 'what's the expected value here?', 'Kelly criterion', 'optimal bet size', 'fractional Kelly... Tags: latest:1.0.4 Version history: v1.0.4 | 2026-07-16T17:59:26.940Z | user Description tail link + agents machine-readable metadata line (deciqai.com/s/expected-value-and-kelly.json) v1.0","descriptionLabel":"Technical summary","evidenceSummary":"Capability contract not published. No trust telemetry is available yet. 959 downloads reported by the source. Last updated 10/12/2026.","installCommand":"clawhub skill install s17a4mqcnk515kvaca5ze55d0x88pfpx:expected-value-and-kelly","sourceUrl":"https://clawhub.ai/deciqai/expected-value-and-kelly","homepage":"https://clawhub.ai/deciqai/skills/expected-value-and-kelly","primaryLinks":[{"label":"View on ClawHub","url":"https://clawhub.ai/deciqai/expected-value-and-kelly","kind":"source"},{"label":"Homepage","url":"https://clawhub.ai/deciqai/skills/expected-value-and-kelly","kind":"homepage"}],"safetyScore":84,"overallRank":62,"popularityScore":60,"trustScore":null,"claimedByName":null,"isOwner":false,"seoDescription":"Activate when: user asks 'how much should I bet/invest on this?', 'what's the expected value here?', 'Kelly criterion', 'optimal bet size', 'fractional Kelly..."},"coverage":{"evidence":{"source":"public-profile","verified":false,"confidence":"medium","updatedAt":"2026-10-12T01:50:01.517Z","emptyReason":null},"protocols":[{"protocol":"OPENCLEW","label":"OpenClaw","status":"self-declared","notes":"Declared in the public agent profile."}],"capabilities":[],"verifiedCount":0,"selfDeclaredCount":1,"capabilityMatrix":{"rows":[{"key":"OPENCLEW","type":"protocol","support":"unknown","confidenceSource":"profile","notes":"Listed on profile"}],"flattenedTokens":"protocol:OPENCLEW|unknown|profile"}},"adoption":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"medium","updatedAt":"2026-10-12T01:50:01.517Z","emptyReason":null},"stars":null,"forks":null,"downloads":959,"likes":null,"task":null,"library":null,"packageName":null,"latestVersion":"1.0.4","tractionLabel":"959 downloads"},"release":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"medium","updatedAt":"2026-10-12T01:50:01.453Z","emptyReason":null},"lastUpdatedAt":"2026-10-12T01:50:01.517Z","lastCrawledAt":"2026-10-12T01:50:01.453Z","lastIndexedAt":null,"nextCrawlAt":"2026-10-13T01:50:01.453Z","lastVerifiedAt":null,"highlights":[{"version":"1.0.4","createdAt":"2026-07-16T17:59:26.940Z","changelog":"Description tail link + agents machine-readable metadata line (deciqai.com/s/expected-value-and-kelly.json)","fileCount":7,"zipByteSize":17979},{"version":"1.0.3","createdAt":"2026-07-08T11:02:05.705Z","changelog":"Footer now uses /c/ short link (fixes UTM truncation when SKILL.md is read in a terminal)","fileCount":7,"zipByteSize":17670},{"version":"1.0.2","createdAt":"2026-07-08T00:47:28.602Z","changelog":"Refreshed content + GitHub star link in footer","fileCount":6,"zipByteSize":13140},{"version":"1.0.1","createdAt":"2026-07-07T20:33:10.053Z","changelog":"Add catalog categories and topics","fileCount":5,"zipByteSize":10423},{"version":"1.0.0","createdAt":"2026-06-28T07:18:15.841Z","changelog":"Initial publish","fileCount":5,"zipByteSize":10186}]},"execution":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"low","updatedAt":null,"emptyReason":"No published capability contract is available yet."},"installCommand":"clawhub skill install s17a4mqcnk515kvaca5ze55d0x88pfpx:expected-value-and-kelly","setupComplexity":"low","setupSteps":["Setup complexity is LOW. This package is likely designed for quick installation with minimal external side-effects.","Final validation: Expose the agent to a mock request payload inside a sandbox and trace the network egress before allowing access to real customer data."],"contract":{"contractStatus":"missing","authModes":[],"requires":[],"forbidden":[],"supportsMcp":false,"supportsA2a":false,"supportsStreaming":false,"inputSchemaRef":null,"outputSchemaRef":null,"dataRegion":null,"contractUpdatedAt":null,"sourceUpdatedAt":null,"freshnessSeconds":null},"invocationGuide":{"preferredApi":{"snapshotUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-expected-value-and-kelly/snapshot","contractUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-expected-value-and-kelly/contract","trustUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-expected-value-and-kelly/trust"},"curlExamples":["curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-expected-value-and-kelly/snapshot\"","curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-expected-value-and-kelly/contract\"","curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-expected-value-and-kelly/trust\""],"jsonRequestTemplate":{"query":"summarize this repo","constraints":{"maxLatencyMs":2000,"protocolPreference":["OPENCLEW"]}},"jsonResponseTemplate":{"ok":true,"result":{"summary":"...","confidence":0.9},"meta":{"source":"CLAWHUB","generatedAt":"2026-10-12T13:19:52.486Z"}},"retryPolicy":{"maxAttempts":3,"backoffMs":[500,1500,3500],"retryableConditions":["HTTP_429","HTTP_503","NETWORK_TIMEOUT"]}},"endpoints":{"dossierUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-expected-value-and-kelly/dossier","snapshotUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-expected-value-and-kelly/snapshot","contractUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-expected-value-and-kelly/contract","trustUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-expected-value-and-kelly/trust"}},"reliability":{"evidence":{"source":"runtime-metrics","verified":false,"confidence":"low","updatedAt":null,"emptyReason":"No trust, reliability, or runtime telemetry is available."},"trust":{"status":"unavailable","handshakeStatus":"UNKNOWN","verificationFreshnessHours":null,"reputationScore":null,"p95LatencyMs":null,"successRate30d":null,"fallbackRate":null,"attempts30d":null,"trustUpdatedAt":null,"trustConfidence":"unknown","sourceUpdatedAt":null,"freshnessSeconds":null},"decisionGuardrails":{"doNotUseIf":["Contract metadata is missing or unavailable for deterministic execution."],"safeUseWhen":[],"riskFlags":["missing_or_unavailable_contract","trust_data_unavailable","schema_references_missing"],"operationalConfidence":"low"},"executionMetrics":{"observedLatencyMsP50":null,"observedLatencyMsP95":null,"estimatedCostUsd":null,"uptime30d":null,"rateLimitRpm":null,"rateLimitBurst":null,"lastVerifiedAt":null,"verificationSource":null},"runtimeMetrics":{"successRate":null,"avgLatencyMs":null,"avgCostUsd":null,"hallucinationRate":null,"retryRate":null,"disputeRate":null,"p50Latency":null,"p95Latency":null,"lastUpdated":null}},"benchmarks":{"evidence":{"source":"no-benchmark-data","verified":false,"confidence":"low","updatedAt":null,"emptyReason":"No benchmark suites or observed failure patterns are available."},"suites":[],"failurePatterns":[]},"artifacts":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"high","updatedAt":"2026-10-12T01:50:01.517Z","emptyReason":null},"readme":"Skill: Expected Value and the Kelly Criterion\n\nOwner: deciqai\n\nSummary: Activate when: user asks 'how much should I bet/invest on this?', 'what's the expected value here?', 'Kelly criterion', 'optimal bet size', 'fractional Kelly...\n\nTags: latest:1.0.4\n\nVersion history:\n\nv1.0.4 | 2026-07-16T17:59:26.940Z | user\n\nDescription tail link + agents machine-readable metadata line (deciqai.com/s/expected-value-and-kelly.json)\n\nv1.0.3 | 2026-07-08T11:02:05.705Z | user\n\nFooter now uses /c/ short link (fixes UTM truncation when SKILL.md is read in a terminal)\n\nv1.0.2 | 2026-07-08T00:47:28.602Z | user\n\nRefreshed content + GitHub star link in footer\n\nv1.0.1 | 2026-07-07T20:33:10.053Z | user\n\nAdd catalog categories and topics\n\nv1.0.0 | 2026-06-28T07:18:15.841Z | user\n\nInitial publish\n\nArchive index:\n\nArchive v1.0.4: 7 files, 17979 bytes\n\nFiles: examples/bill-benter-hong-kong-horse-racing-1985-2001.md (4108b), examples/ed-thorp-blackjack-and-princeton-newport-1961-1988.md (5794b), examples/hyperscaler-gpu-capex-kelly-2023-2026.md (7330b), references/sources.md (4329b), skill-card.md (3402b), SKILL.md (9691b), _meta.json (143b)\n\nFile v1.0.4:SKILL.md\n\n---\nname: expected-value-and-kelly\ndescription: \"Activate when: user asks 'how much should I bet/invest on this?', 'what's the expected value here?', 'Kelly criterion', 'optimal bet size', 'fractional Kelly', 'how big a position should I take?', or is allocating capital across repeated decisions (ad spend by segment, VC portfolio construction, position sizing, A/B test ramp).\n Do NOT activate when: the decision is one-shot and non-repeating (career change, marriage) — use regret-minimization instead; or when the user cannot estimate probabilities or payoffs even roughly. More: deciqai.com/c/expected-value-and-kelly\"\n---\n\n# Expected Value and the Kelly Criterion\n\n## Overview\n\nTwo questions decide most repeated bets: **is this bet good?** (EV) and **how big?** (Kelly). Most professional ruin comes from positive-EV bets sized wrong. **EV = p · W − q · L.** If EV ≤ 0, do not bet. **Kelly f\\* = (bp − q) / b** maximizes long-term geometric growth (Kelly, Bell Labs, 1956). Full Kelly requires casino-grade certainty; default to half- or quarter-Kelly for estimated edges.\n\nNeighbors: first-principles · occams-razor · second-order-thinking · inversion · regret-minimization (for non-repeating life decisions).\n\n## When to Use\n\n- Decision **repeats many times** — capital allocation, position sizing, VC portfolio, ad spend, A/B test budget\n- *How big to bet* matters as much as *whether to bet*; you have a **measurable or estimable edge**\n- Someone says: \"expected value,\" \"EV,\" \"Kelly,\" \"optimal bet size,\" \"how much should we put on this?\"\n- Sizing bets in a boom with power-law payoffs and possible ruin — **how much to allocate to AI startups / GPU-compute capex / AI-exposed equities** given frothy AI valuations, uncertain AI adoption, and correlated bets\n\n**When NOT to use:** one-shot life decisions → regret-minimization; negative-EV bets (don't bet); unestimable probabilities; correlated bets without portfolio adjustment.\n\n## Coaching Novices (Adaptive Front Door)\n\n**Engine mode:** user has a concrete repeated bet → run The Process directly. **Coach mode:** user is unfamiliar → guide step by step.\n\nIn Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.\n\n1. One-line what-it-is: EV tells you **whether the bet is worth taking**; Kelly tells you **what fraction of bankroll to stake** — sized to maximize long-term growth without ruin.\n2. Check fit against When to Use / When NOT to use. If one-shot life decision, redirect to regret-minimization. If EV is negative, say \"don't bet\" and stop.\n3. Elicit their real bet. Ask for a concrete repeated decision with measurable inputs. > **[WAIT — do not advance until user responds]**\n4. Walk The Process one step per turn: outcomes → probabilities → payoffs → EV → Kelly → fractional Kelly. > **[WAIT — do not advance until user responds]**\n5. Close by naming their sizing rule: \"bet f\\* × bankroll, use half-Kelly given estimation uncertainty\" — and the trigger that would change it. > **[WAIT — do not advance until user responds]**\n\n## The Process\n\nRun the **EV-Kelly Sizing** (EV → Kelly → fractional Kelly → stop trigger):\n\n1. **Confirm the decision is repeated.** If \"once,\" stop → use regret-minimization.\n2. **Map the bet.** Win prob *p*, loss prob *q = 1−p*, payoff on win *W*, loss *L*, odds *b = W/L*.\n3. **Compute EV.** EV = Σ(pᵢ · payoffᵢ). State per unit staked.\n4. **Edge gate.** EV > 0? If no, stop — do not bet.\n5. **Estimate input uncertainty.** Are *p* and *W* measured or estimated? Write 80% CI on edge.\n6. **Compute Kelly fraction.** f\\* = (bp − q) / b. For continuous: f\\* ≈ μ/σ².\n7. **Apply fractional Kelly.** Half-Kelly under modest uncertainty; quarter-Kelly under serious uncertainty.\n8. **Set stop trigger.** \"I will re-estimate if: (a) drawdown > X%, (b) outcomes diverge N σ over Y trials, (c) regime change invalidates edge model.\"\n\n### Output: EV-Kelly Sizing\n\n```markdown\n# EV-Kelly Sizing: \n## Repeatability: — if one-shot, STOP.\n## Bet map: p=<>, q=<>, W=<>, L=<>, b=<>\n## EV: p·W − q·L = — Edge: \n## Estimation uncertainty: ; 80% CI on edge: \n## Kelly fraction: f* = → practitioner: = \n## Stop trigger: \"I will re-estimate if .\"\n## Correlation check: bets independent? \n```\n*→ Method in Action: [Ed Thorp, Blackjack, and Princeton-Newport (1961 → 1988)](examples/ed-thorp-blackjack-and-princeton-newport-1961-1988.md) · [Bill Benter, Hong Kong Horse Racing (1985 → 2001)](examples/bill-benter-hong-kong-horse-racing-1985-2001.md)*\n*→ 2026 lens: [Sizing Hyperscaler GPU/Compute Capex (2023–2026)](examples/hyperscaler-gpu-capex-kelly-2023-2026.md) — one CFO's repeated capex bet: wide CI, stranded-capital ruin tail, and fractional Kelly as staged, survivable capex.*\n## Sizing Packs\n\n| Domain | Fractional-Kelly | Stop trigger |\n|---|---|---|\n| Active equity | quarter-Kelly or less | drawdown > 2× expected annual vol |\n| Venture capital | portfolio-level quarter-Kelly | hit rate diverges from model by vintage |\n| Ad spend by segment | half- to full Kelly | ROAS falls >2σ over N conversions |\n| A/B test ramp | fractional Kelly on traffic % | regression in primary metric |\n\n## Applying It Well\n\n- **Positive EV is necessary but not sufficient.** A positive-EV bet sized too large still wrecks you.\n- **Default to fractional Kelly.** Full Kelly is for measured edges. Half- or quarter-Kelly for estimated edges — growth cost is small, protection is large.\n- **Correlation eats Kelly fast.** All positions long tech, all VC bets in one vintage — adjust portfolio Kelly down.\n- **Kelly is for bankroll, not for life.** Use regret-minimization for career, marriage, time.\n- **Update continuously.** Static Kelly on a stale edge is how winning strategies ride into ruin.\n\n*→ Primary sources: [references/sources.md](references/sources.md)*\n## Common Rationalizations\n\n**[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.**\n\n| Fake move | Reality |\n|---|---|\n| [D] Applying Kelly to **one-shot** decisions | Kelly maximizes long-run geometric growth rate across many bets. For a one-time bet, EV is the right concern; for life decisions, use regret-minimization. |\n| [D] Treating **estimated *p*** as **known *p*** | Kelly is brutal when probability estimates are wrong. Use fractional Kelly to compensate for estimated inputs. |\n| [D] Using **full Kelly with subjective probabilities** | Full Kelly is for casino-grade certainty. Half-Kelly costs ~25% of growth and cuts ruin risk by far more (Thorp 1997). |\n| [D] Confusing **positive EV** with **\"should bet\"** | EV ignores variance, bankroll, path-dependency. A +1% EV bet that ruins you 1% of the time is not equivalent to one that ruins you 0%. |\n| [D] **Ignoring correlation** across bets | Kelly assumes independent bets. Correlated bets (same sector, same vintage) require a lower portfolio-level Kelly. |\n| [D] Using Kelly on **negative-EV** bets | Kelly returns zero or negative fraction when EV ≤ 0. No sizing rescues a -EV bet. Don't bet. |\n| [D] Treating EV as the **only number that matters** | Bernoulli 1738: utility of money is non-linear. A +$1000 EV bet that risks your rent ≠ one that risks your rounding error. |\n| [D] Forgetting Kelly's **brutal variance** | Even correct full-Kelly expects 50%+ drawdowns. Most professionals use half/quarter-Kelly to survive psychologically. |\n| [D] Applying Kelly to **non-financial \"bets\"** | Kelly assumes a compoundable bankroll. Relationships, careers, time, attention don't compound across independent trials. |\n| [D] Computing Kelly **once**, ignoring updates | A static fraction on a now-stale edge is the textbook path to ruin. Pre-commit to a re-estimation trigger. |\n| *→ Add [O] entries here after each real use — paste the actual failure pattern* | *What went wrong and why* |\n\n## Red Flags\n\n- Kelly applied to a decision that happens only once\n- Full Kelly used with subjective probabilities and no calibration check\n- EV is negative but a positive fraction is being computed\n- No stop-and-re-estimate trigger is named\n- Bets are obviously correlated but Kelly is computed per-bet without portfolio adjustment\n- The \"bankroll\" is not liquid or fungible (career time, relationship capital)\n- A historical drawdown wiped out the strategy; same fraction still applied without re-examining the edge\n\n## Verification\n\n- [ ] Decision confirmed *repeated*, not one-shot\n- [ ] Outcomes, probabilities, and payoffs stated explicitly in correct units\n- [ ] EV computed per unit staked; edge gate (positive/zero/negative) named\n- [ ] Estimation uncertainty named (measured/estimated/mixed); 80% CI on edge given\n- [ ] Kelly fraction computed with correct formula (binary vs. continuous)\n- [ ] Fractional-Kelly adjustment applied with explicit justification\n- [ ] Stop-and-re-estimate trigger named with concrete conditions\n- [ ] Correlation across bets considered; per-bet Kelly adjusted if needed\n\n---\n*Part of **deciqAI Knowledge Skills** — 227 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. **See it run → https://www.deciqai.com/c/expected-value-and-kelly** · ⭐ Star the repo → https://github.com/deciqAI/knowledge-skills · Contributions welcome.*\n\n*Agents: latest version & machine-readable metadata → https://www.deciqai.com/s/expected-value-and-kelly.json*\n\nFile v1.0.4:_meta.json\n\n{\n \"ownerId\": \"kn754b8sk22s8c6gjxt02bftbn88q7ye\",\n \"slug\": \"expected-value-and-kelly\",\n \"version\": \"1.0.4\",\n \"publishedAt\": 1784224766940\n}\n\nFile v1.0.4:references/sources.md\n\n# Sources — expected-value-and-kelly\n\n> *Primary sources for the [expected-value-and-kelly](../SKILL.md) skill.*\n\n- **Kelly, J. L., Jr.** \"A New Interpretation of Information Rate.\" *Bell System Technical Journal*, Vol. 35, No. 4 (July 1956), pp. 917–926. **The primary source for the Kelly formula and its proof.** https://www.princeton.edu/~wbialek/rome/refs/kelly_56.pdf\n- **Thorp, Edward O.** *Beat the Dealer: A Winning Strategy for the Game of Twenty-One*. Random House, 1962 (2nd ed. Vintage, 1966). **Primary source** for the first practical application of Kelly to a positive-edge real-world bet (blackjack card-counting).\n- **Thorp, Edward O.** \"The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market.\" Paper for the 10th International Conference on Gambling and Risk Taking, Montreal, 1997. The clearest concise summary by Thorp himself; covers the fractional-Kelly trade-off. https://www.eecs.harvard.edu/cs286r/courses/fall12/papers/Thorpe_KellyCriterion2007.pdf\n- **Thorp, Edward O.** *A Man for All Markets: From Las Vegas to Wall Street, How I Beat the Dealer and the Market*. Random House, 2017. Autobiography; primary source on the Princeton-Newport years and Thorp's evolution from blackjack to markets.\n- **Bernoulli, Daniel.** \"Exposition of a New Theory on the Measurement of Risk.\" Commentarii Academiae Scientiarum Imperialis Petropolitanae, Tomus V (St. Petersburg, 1738). English translation by Louise Sommer in *Econometrica*, Vol. 22, No. 1 (January 1954), pp. 23–36. **Primary source** for the utility-versus-EV distinction.\n- **Benter, William.** \"Computer Based Horse Race Handicapping and Wagering Systems: A Report.\" In *Efficiency of Racetrack Betting Markets*, edited by Donald B. Hausch, Victor S. Y. Lo, and William T. Ziemba, pp. 183–198. Academic Press, 1994 (reprinted World Scientific, 2008). **Primary source** for combining a fitted probability model with public odds and sizing pari-mutuel bets by fractional Kelly under estimated inputs and market impact.\n- **Poundstone, William.** *Fortune's Formula: The Untold Story of the Scientific Betting System That Beat the Casinos and Wall Street*. Hill & Wang, 2005. The popular history; useful as a secondary reference and for the Bell Labs / Shannon / Kelly intellectual lineage.\n- **U.S. Securities and Exchange Commission filings and quarterly earnings materials of Microsoft, Alphabet, Amazon, Meta, and Nvidia (2023–2025).** Durable **primary sources** for AI-driven data-center capital-expenditure guidance and Nvidia's data-center-segment revenue growth during the 2023–2026 AI investment boom — used in preference to fast-moving press figures, which should be verified against these filings. https://www.sec.gov/cgi-bin/browse-edgar\n- **Contemporaneous business-press coverage of frontier-AI-lab financing rounds and valuations (e.g., The Wall Street Journal, Financial Times, The Information, 2023–2026).** **Secondary** reference for reported startup valuations (OpenAI, Anthropic, and others); treat specific figures as approximate and time-stamped, since boom-era valuations are reported inconsistently and change rapidly.\n- **NVIDIA Corporation, Annual Report on Form 10-K and quarterly Form 10-Q / earnings materials (fiscal years 2024–2025).** **Primary source** for the growth of the Data Center reporting segment during the 2023–2026 AI compute boom; used in preference to press summaries of \"Nvidia's AI revenue,\" which should be verified against the segment disclosures in these filings. https://www.sec.gov/cgi-bin/browse-edgar\n- **MacLean, Leonard C., Edward O. Thorp, and William T. Ziemba (eds.).** *The Kelly Capital Growth Investment Criterion: Theory and Practice.* World Scientific, 2011. **Authoritative collected reference** on Kelly and fractional-Kelly investing, including the growth-vs-drawdown trade-off that justifies defaulting to a fraction of full Kelly when the edge is estimated rather than measured — directly relevant to sizing correlated, power-law AI bets.\n- The popular framing \"the house always wins\" is **not** cited here as a source — by this skill's own rule, an aphorism is not evidence. The Kelly criterion is precisely the formula for *the conditions under which the house does not always win* — when the player has a measurable positive edge.\n\nFile v1.0.4:examples/bill-benter-hong-kong-horse-racing-1985-2001.md\n\n# Method in Action: Bill Benter and Computer-Model Horse Betting in Hong Kong (1985 → 2001)\n\n> *Example for the [expected-value-and-kelly](../SKILL.md) skill.*\n\nA worked example in a domain far from a card table: **pari-mutuel horse racing**, where the odds are set by the crowd's own money and your bet moves the payout against you. Primary-source documented in Benter's own 1994 technical paper.\n\nIn the mid-1980s **William \"Bill\" Benter** — who had learned card-counting from Thorp's *Beat the Dealer* and been barred from casinos — turned to the Hong Kong Jockey Club, then the largest pari-mutuel racing pool in the world. His insight was statistical, not tipster's intuition: build a **multinomial logit model** that estimates each horse's true win probability from fundamental factors (past times, weight, jockey, barrier, going), then combine that estimate with the **public's implied probability** encoded in the tote odds. The crowd is a strong predictor; a model that ignores it loses, and a model that only mimics it has no edge. Benter's edge came from the *residual* — the systematic gap between his model's probabilities and the crowd's.\n\nThat residual defined the bet. On any race, a horse is a positive-EV wager only when Benter's estimated win probability *p* exceeds the break-even probability implied by its tote odds after the Jockey Club's ~17–19% takeout. Most horses in most races failed that gate. The discipline was to bet only the small fraction that cleared it — and to size each qualifying bet by Kelly.\n\nPari-mutuel racing punishes overbetting twice. First, Kelly's ordinary lesson: a positive-EV bet staked too large courts ruin through variance. Second, a mechanical trap absent from casino play — **your own money enters the pool and lowers the odds you receive.** A full-Kelly stake computed against the pre-bet odds is, after your bet lands, an *overbet* against the true post-bet payout. Benter's paper is explicit that they used **fractional Kelly**, both to absorb estimation error in the probability model and to blunt this self-impact on the tote. Run over hundreds of thousands of bets across many seasons, the syndicate is widely reported to have won on the order of a billion dollars — the integral effect of a small per-race edge, correctly sized and relentlessly repeated.\n\nWalk the EV-Kelly Sizing through Benter's case:\n\n- **Repeatability (Step 1):** Hundreds of races per season, many seasons, thousands of qualifying bets. Emphatically repeated. ✓\n- **Bet map (Step 2):** For a candidate horse, model win probability *p*; payoff *W* set by the tote (dividend minus stake) after ~17–19% takeout; loss *L* = the stake.\n- **EV (Step 3):** EV = p·W − q·L per unit staked. Computed per horse, per race.\n- **Edge gate (Step 4):** Bet only when *p* exceeds the takeout-adjusted break-even implied by the odds. The overwhelming majority of horses fail this gate and are not bet.\n- **Estimation uncertainty (Step 5):** High and irreducible — *p* comes from a fitted statistical model, not from a countable deck. This is estimated, not measured, edge.\n- **Kelly fraction (Step 6):** f\\* = (bp − q) / b per qualifying bet, with *b* from the tote odds.\n- **Fractional Kelly (Step 7):** A fraction of full Kelly — for model estimation error *and* to limit the bet's own depressive effect on the pari-mutuel payout.\n- **Stop trigger (Step 8):** Continuous model re-estimation each season; degrade or halt when out-of-sample calibration drifts or the residual edge shrinks as the pool grows more efficient.\n\nThe lesson Benter's case adds to Thorp's: when *p* is a model output and *W* is a payout your own bet degrades, full Kelly is doubly wrong. Fractional Kelly is not caution — it is the correct sizing under estimated inputs and market impact.\n\nPrimary source: Benter, William. \"Computer Based Horse Race Handicapping and Wagering Systems: A Report.\" In *Efficiency of Racetrack Betting Markets*, edited by Donald B. Hausch, Victor S. Y. Lo, and William T. Ziemba, pp. 183–198. Academic Press, 1994 (reprinted World Scientific, 2008).\n\nFile v1.0.4:examples/ed-thorp-blackjack-and-princeton-newport-1961-1988.md\n\n# Method in Action: Ed Thorp, Blackjack, and Princeton-Newport (1961 → 1988)\n\n> *Example for the [expected-value-and-kelly](../SKILL.md) skill.*\n\nA worked example. Not a casino fable — primary-source documented, including the Bell Labs origin.\n\nIn 1960, **Ed Thorp**, a young mathematics professor at MIT, was reading Baldwin et al.'s 1956 paper on basic-strategy blackjack and realized something the casino-gambling literature had missed: because cards in blackjack are drawn *without replacement* from a finite deck, the composition of the remaining deck shifts the player's edge from hand to hand. A player who **tracks** the cards already played (a \"counter\") sees a small but real positive edge on certain deck compositions — typically when the remaining deck is rich in tens and aces.\n\nThorp's contribution had two parts. First, he proved the math: a counting system (his \"Ten-Count,\" later refined to the \"High-Low\" by Harvey Dubner) yields an average edge of roughly **+1% over the house** when applied disciplinedly. Second — and this is what most blackjack popularizers miss — he applied **Kelly sizing** to the bet: vary the wager up and down with the count, betting more when the count favored the player and the minimum when it didn't.\n\nTo prove it in the field, Thorp went to Las Vegas and Reno casinos in spring 1961, funded by **Manny Kimmel**, a New York businessman with $10,000 in seed capital. Over a series of trips, they confirmed the math at the tables. The result was published in **Thorp's 1962 book *Beat the Dealer*** (Random House) — a primary source for everything in this Method-in-Action section. The casinos' second-order response was rapid: they changed rules, introduced multiple decks and shoe shuffles, and trained dealers to spot counters. Within a few years, Thorp's specific edge in single-deck games had been engineered out of most casinos.\n\nBut the *framework* — combine a measurable edge with disciplined Kelly sizing across a repeated series of independent bets — translated. In 1969, Thorp co-founded **Princeton-Newport Partners**, one of the first quantitative hedge funds. The fund's strategy was convertible-bond arbitrage, statistical arbitrage, and other measurable-edge market trades, sized using Kelly principles (Thorp typically used a fractional Kelly given the higher estimation uncertainty in markets vs. casinos). **From 1969 to 1988, the fund returned 19.1% annually before fees with no down year** (Thorp, *A Man for All Markets*, 2017, ch. 14). The Princeton-Newport partnership ended for unrelated legal reasons in 1988.\n\nWalk through the EV-Kelly Sizing on Thorp's blackjack case:\n\n- **Repeatability (Step 1):** Tens of thousands of hands per casino trip. Strongly repeated. ✓\n- **Bet map (Step 2):** Per hand at a favorable count, win probability *p ≈ 0.505*, loss *q ≈ 0.495* (the +1% edge); payoff approximately 1:1 (with some variance from doubles, splits, blackjack 3:2).\n- **EV (Step 3):** ≈ +0.01 per unit staked. *Positive but small.*\n- **Edge gate (Step 4):** Positive. Proceed.\n- **Estimation uncertainty (Step 5):** In 1961, with a measurable card-counting system and verifiable math, this was as close to **casino-grade certainty** as bets get. Estimation uncertainty was unusually low (one of the few real-world cases where near-full Kelly is defensible).\n- **Kelly fraction (Step 6):** With *p = 0.505, q = 0.495, b = 1*: f\\* = (1 × 0.505 − 0.495) / 1 = **0.01 = 1% of bankroll per hand at a favorable count**. (Thorp varied the bet *up* to roughly this level at the highest-positive counts and dropped to the table minimum at negative counts.)\n- **Fractional Kelly (Step 7):** Because the edge math was unusually well-measured, Thorp used closer to full Kelly than is typical for market trades. In later markets work at Princeton-Newport, he scaled to fractional Kelly given larger estimation uncertainty.\n- **Stop trigger (Step 8):** Casino rule changes (multiple decks, frequent shuffles) — *the casinos' second-order response* — eventually made the edge structure obsolete in most venues. Thorp re-estimated repeatedly and pivoted out.\n\nThe inversion-style lesson Thorp himself emphasized later: **the most expensive mistake in EV-Kelly is using full Kelly with an estimated edge.** Half-Kelly costs roughly 25% of growth rate but cuts variance and ruin risk by far more (Thorp 1997 conference paper, §4).\n\n**Sources:** Kelly, J. L. **\"A New Interpretation of Information Rate.\"** *Bell System Technical Journal*, Vol. 35, No. 4 (July 1956), pp. 917–926 — the primary source for the formula and its derivation. https://www.princeton.edu/~wbialek/rome/refs/kelly_56.pdf . Thorp, Edward O. ***Beat the Dealer: A Winning Strategy for the Game of Twenty-One*** (Random House, 1962; 2nd ed. Vintage 1966) — primary source for the blackjack application. Thorp, Edward O. **\"The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market.\"** Paper for the 10th International Conference on Gambling and Risk Taking, Montreal, 1997 — the clearest concise summary by Thorp himself. https://www.eecs.harvard.edu/cs286r/courses/fall12/papers/Thorpe_KellyCriterion2007.pdf . Thorp, Edward O. ***A Man for All Markets*** (Random House, 2017) — autobiography; primary source for the Princeton-Newport years. Bernoulli, Daniel. **\"Exposition of a New Theory on the Measurement of Risk\"** (St. Petersburg, 1738; English tr. Louise Sommer, *Econometrica*, Vol. 22, No. 1, January 1954, pp. 23–36) — primary source for the utility-vs-EV distinction. Poundstone, William. ***Fortune's Formula: The Untold Story of the Scientific Betting System That Beat the Casinos and Wall Street*** (Hill & Wang, 2005) — the popular history of Kelly + Thorp, useful as a secondary reference.\n\nFile v1.0.4:examples/hyperscaler-gpu-capex-kelly-2023-2026.md\n\n# Method in Action: Sizing Hyperscaler GPU/Compute Capex in the AI Boom (2023–2026)\n\n> *Example for the [expected-value-and-kelly](../SKILL.md) skill.*\n\nA worked example on the *capex* side of the 2023–2026 AI gold rush, told as a single decision-maker's repeated allocation problem. Where the companion example ([Bet-Sizing the 2023–2026 AI Gold Rush](ai-gold-rush-vc-bet-sizing-2023-2026.md)) treats a venture fund's whole portfolio, this one narrows to one recurring board-level choice: **how much of annual free cash flow should a large cloud provider commit to AI compute (GPUs, data centers, power) each planning cycle?** It is a positive-expected-value bet with a fat upside, a wide confidence interval, and a genuine ruin tail (stranded, hard-to-repurpose capital). That combination is exactly what fractional Kelly exists for.\n\nThe setting is durable and well documented. From roughly 2023 onward — after ChatGPT's late-2022 public release — the large hyperscalers (Microsoft, Alphabet, Amazon, Meta) each raised AI-related capital expenditure sharply, guiding through 2024–2025 to tens of billions of dollars per year, much of it tied to AI compute; Nvidia's data-center GPU demand drove its revenue and market capitalization to among the largest of any public company. The *structure* of the capex decision — not the specific dollar figures, which move fast and are reported inconsistently — is what the EV-Kelly lens clarifies.\n\nWalk the **EV-Kelly Sizing** through one hyperscaler's annual capex-allocation policy:\n\n- **Confirm the decision is repeated (Step 1):** This is not a one-shot. Capex is re-set every planning cycle — quarterly guidance, annual budgets, multi-year build-outs — and the *sizing policy* recurs across cycles even as any single data-center build is one-off. You are sizing the recurring allocation rule, not a single facility. Kelly applies. ✓\n- **Map the bet (Step 2):** Stake = the fraction of free cash flow (the \"bankroll\") committed to AI compute this cycle. On a **win** (AI demand monetizes: cloud consumption, ads, subscriptions, enterprise seats), payoff *W* is large and compounding — you own the scarce capacity when demand arrives. On a **loss** (demand disappoints or a cheaper compute regime arrives), loss *L* is the stranded portion of the capex: GPUs and buildings that can't be fully repurposed, depreciating on a short useful-life schedule. This is a **skewed, power-law-flavored** payoff, not a symmetric coin flip — see [power-law-distribution](../power-law-distribution/SKILL.md).\n- **Compute EV (Step 3):** EV = p·W − q·L per dollar committed. Because *W* (owning scarce AI capacity in a demand surge) is large and the marginal loss on a miss is a fraction of the stake (some capacity is reusable for general cloud), portfolio EV can be clearly positive even if the probability of the *full* upside is well under one-half. The math tolerates a modest hit rate because the win is big and the downside is bounded, not total.\n- **Edge gate (Step 4):** Positive EV is necessary but not sufficient. The edge is real only if the firm has a genuine advantage — distribution, existing cloud demand, model or data assets — over merely paying boom-era prices for capacity everyone else is also buying. If the \"edge\" is just fear of missing out, EV is illusory and no sizing rescues it.\n- **Estimate input uncertainty (Step 5):** Wide and largely irreducible. *p* (that demand monetizes before the capex is written off) and *W* are forward guesses about an unprecedented technology mid-hype-cycle — not measured frequencies (a deck) or decades-fitted (Benter's races). The 80% CI on the edge is broad and could approach zero if a capability plateau or a compute-cost collapse arrives. This alone forces a small Kelly fraction.\n- **Compute Kelly fraction (Step 6):** In the continuous framing, f\\* ≈ μ/σ². The dispersion σ of AI-capex outcomes is enormous — the range from \"generational franchise\" to \"expensive stranded silicon\" is wide — so the Kelly-optimal fraction of the free-cash-flow bankroll on *any single* cycle's bet is modest even for a genuinely positive edge. High variance shrinks Kelly.\n- **Apply fractional Kelly (Step 7):** Given both the estimation uncertainty (Step 5) and the fat-tailed variance (Step 6), the discipline is **well under full Kelly** — quarter-Kelly territory. In capex language: size the AI build so that a *wrong* bet is survivable, not fatal — retain balance-sheet slack, stage commitments so they can be paused, and prefer capacity that can fall back to general-purpose cloud. Bounded downside, retained exposure to the upside. That *is* fractional Kelly, implemented as staged, survivable capex.\n- **Set stop trigger (Step 8):** Pre-commit to re-estimating the edge when AI revenue attach-rates diverge from the model, when utilization of deployed capacity runs persistently low, when GPU price/performance shifts enough to strand the current generation, or when a capability plateau or funding-market freeze breaks the demand thesis.\n\n**Where correlation bites.** Kelly assumes independent bets; hyperscaler AI capex is anything but. The capex bet, the firm's own frontier-model bet, its cloud customers' AI startups, and the public GPU complex all rest on the same load-bearing assumptions — that capability keeps improving, that compute stays available, and that demand monetizes before write-off. When those co-move, a \"diversified\" AI position behaves like one concentrated bet. The skill's rule is blunt: **correlation eats Kelly fast** — same narrative, same cycle, same underlying assumption means a *lower* portfolio-level Kelly, not the naive sum of per-project Kellys.\n\nThe lesson this case adds to Thorp and Benter: their edges were *measured* or *fitted* on stationary processes. Hyperscaler AI capex is neither — probabilities are guesses, payoffs are skewed, ruin (stranded capital) is possible, and the bets are secretly one bet. Under those conditions full Kelly is not merely aggressive; it is a category error. **Fractional Kelly, expressed as staged and survivable capex, is the correct response to a genuinely positive but genuinely uncertain edge.** The question is not \"how big can we bet on AI?\" — it is \"how do we stay solvent and flexible long enough to still own the capacity if and when the demand actually shows up?\"\n\n*Sources: Kelly, J. L., Jr., \"A New Interpretation of Information Rate,\" Bell System Technical Journal 35(4), 1956, pp. 917–926 (the Kelly formula). Thorp, Edward O., \"The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market,\" 1997 (the fractional-Kelly trade-off under estimated edges). For the 2023–2026 hyperscaler AI-capex environment — annual capital-expenditure guidance tied to AI compute and Nvidia's data-center-driven revenue and market capitalization — see the companies' own public filings and earnings materials (Microsoft, Alphabet, Amazon, Meta, and Nvidia SEC filings and quarterly reports, 2023–2025) as durable primary sources; specific capex and valuation figures cited in contemporaneous business press (e.g., The Wall Street Journal, Financial Times, The Information, 2023–2026) should be checked against those filings, as boom-era figures move quickly and are reported inconsistently.*\n\nFile v1.0.4:skill-card.md\n\n## Description:\n\nHelps agents evaluate repeated bets and capital-allocation decisions by computing expected value, Kelly sizing, fractional Kelly adjustments, and stop-and-reestimate triggers.\n\nThis skill is ready for commercial/non-commercial use.\n\n## Publisher:\n\n[deciqai](https://clawhub.ai/user/deciqai)\n\n### License/Terms of Use:\n\nMIT-0\n\n## Use Case:\n\nExternal users and agents use this skill to decide whether a repeated bet, investment, budget allocation, A/B-test ramp, or position-sizing problem has positive expected value and how large the stake should be. It guides the agent through probability, payoff, uncertainty, correlation, fractional-Kelly, and stop-trigger checks before producing a sizing recommendation.\n\n### Deployment Geography for Use:\n\nGlobal\n\n## Known Risks and Mitigations:\n\nRisk: Users may treat EV or Kelly outputs as financial, gambling, or legal advice.\n\nMitigation: Present outputs as educational decision support and require independent review of probabilities, payoffs, bankroll constraints, and applicable legal or policy limits before action.\n\nRisk: Incorrect probability, payoff, or correlation assumptions can produce misleading stake sizes even when the formulas are applied correctly.\n\nMitigation: Require explicit assumptions, uncertainty ranges, fractional-Kelly sizing for estimated edges, correlation checks, and stop-and-reestimate triggers.\n\nRisk: Applying Kelly to one-shot, non-repeatable, or non-compoundable life decisions can give inappropriate recommendations.\n\nMitigation: Gate use to repeated decisions with estimable probabilities and redirect one-shot life decisions to another decision framework.\n\n## Reference(s):\n\n- [Expected Value and Kelly Source List](references/sources.md)\n- [Ed Thorp Blackjack and Princeton-Newport Example](examples/ed-thorp-blackjack-and-princeton-newport-1961-1988.md)\n- [Bill Benter Hong Kong Horse Racing Example](examples/bill-benter-hong-kong-horse-racing-1985-2001.md)\n- [Hyperscaler GPU Capex Kelly Example](examples/hyperscaler-gpu-capex-kelly-2023-2026.md)\n- [Kelly, A New Interpretation of Information Rate](https://www.princeton.edu/~wbialek/rome/refs/kelly_56.pdf)\n- [Thorp, The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market](https://www.eecs.harvard.edu/cs286r/courses/fall12/papers/Thorpe_KellyCriterion2007.pdf)\n- [SEC EDGAR Company Filings](https://www.sec.gov/cgi-bin/browse-edgar)\n- [ClawHub Skill Page](https://clawhub.ai/deciqai/skills/expected-value-and-kelly)\n- [deciqAI Expected Value and Kelly Page](https://www.deciqai.com/c/expected-value-and-kelly)\n- [deciqAI Machine-Readable Skill Metadata](https://www.deciqai.com/s/expected-value-and-kelly.json)\n\n## Skill Output:\n\n**Output Type(s):** [Analysis, Markdown, Guidance]\n\n**Output Format:** [Markdown EV-Kelly sizing analysis with formulas, assumptions, uncertainty, fractional-Kelly recommendation, stop trigger, and correlation check]\n\n**Output Parameters:** [1D]\n\n**Other Properties Related to Output:** [Educational decision support; does not execute code, access data, or place bets]\n\n## Skill Version(s):\n\n1.0.4 (source: server release evidence)\n\n## Ethical Considerations:\n\nUsers should evaluate whether this skill is appropriate for their environment, review any generated or modified files before relying on them, and apply their organization's safety, security, and compliance requirements before deployment.\n\nArchive v1.0.3: 7 files, 17670 bytes\n\nFiles: examples/bill-benter-hong-kong-horse-racing-1985-2001.md (4108b), examples/ed-thorp-blackjack-and-princeton-newport-1961-1988.md (5794b), examples/hyperscaler-gpu-capex-kelly-2023-2026.md (7330b), references/sources.md (4329b), skill-card.md (2768b), SKILL.md (9532b), _meta.json (143b)\n\nFile v1.0.3:SKILL.md\n\n---\nname: expected-value-and-kelly\ndescription: \"Activate when: user asks 'how much should I bet/invest on this?', 'what's the expected value here?', 'Kelly criterion', 'optimal bet size', 'fractional Kelly', 'how big a position should I take?', or is allocating capital across repeated decisions (ad spend by segment, VC portfolio construction, position sizing, A/B test ramp).\n Do NOT activate when: the decision is one-shot and non-repeating (career change, marriage) — use regret-minimization instead; or when the user cannot estimate probabilities or payoffs even roughly.\"\n---\n\n# Expected Value and the Kelly Criterion\n\n## Overview\n\nTwo questions decide most repeated bets: **is this bet good?** (EV) and **how big?** (Kelly). Most professional ruin comes from positive-EV bets sized wrong. **EV = p · W − q · L.** If EV ≤ 0, do not bet. **Kelly f\\* = (bp − q) / b** maximizes long-term geometric growth (Kelly, Bell Labs, 1956). Full Kelly requires casino-grade certainty; default to half- or quarter-Kelly for estimated edges.\n\nNeighbors: first-principles · occams-razor · second-order-thinking · inversion · regret-minimization (for non-repeating life decisions).\n\n## When to Use\n\n- Decision **repeats many times** — capital allocation, position sizing, VC portfolio, ad spend, A/B test budget\n- *How big to bet* matters as much as *whether to bet*; you have a **measurable or estimable edge**\n- Someone says: \"expected value,\" \"EV,\" \"Kelly,\" \"optimal bet size,\" \"how much should we put on this?\"\n- Sizing bets in a boom with power-law payoffs and possible ruin — **how much to allocate to AI startups / GPU-compute capex / AI-exposed equities** given frothy AI valuations, uncertain AI adoption, and correlated bets\n\n**When NOT to use:** one-shot life decisions → regret-minimization; negative-EV bets (don't bet); unestimable probabilities; correlated bets without portfolio adjustment.\n\n## Coaching Novices (Adaptive Front Door)\n\n**Engine mode:** user has a concrete repeated bet → run The Process directly. **Coach mode:** user is unfamiliar → guide step by step.\n\nIn Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.\n\n1. One-line what-it-is: EV tells you **whether the bet is worth taking**; Kelly tells you **what fraction of bankroll to stake** — sized to maximize long-term growth without ruin.\n2. Check fit against When to Use / When NOT to use. If one-shot life decision, redirect to regret-minimization. If EV is negative, say \"don't bet\" and stop.\n3. Elicit their real bet. Ask for a concrete repeated decision with measurable inputs. > **[WAIT — do not advance until user responds]**\n4. Walk The Process one step per turn: outcomes → probabilities → payoffs → EV → Kelly → fractional Kelly. > **[WAIT — do not advance until user responds]**\n5. Close by naming their sizing rule: \"bet f\\* × bankroll, use half-Kelly given estimation uncertainty\" — and the trigger that would change it. > **[WAIT — do not advance until user responds]**\n\n## The Process\n\nRun the **EV-Kelly Sizing** (EV → Kelly → fractional Kelly → stop trigger):\n\n1. **Confirm the decision is repeated.** If \"once,\" stop → use regret-minimization.\n2. **Map the bet.** Win prob *p*, loss prob *q = 1−p*, payoff on win *W*, loss *L*, odds *b = W/L*.\n3. **Compute EV.** EV = Σ(pᵢ · payoffᵢ). State per unit staked.\n4. **Edge gate.** EV > 0? If no, stop — do not bet.\n5. **Estimate input uncertainty.** Are *p* and *W* measured or estimated? Write 80% CI on edge.\n6. **Compute Kelly fraction.** f\\* = (bp − q) / b. For continuous: f\\* ≈ μ/σ².\n7. **Apply fractional Kelly.** Half-Kelly under modest uncertainty; quarter-Kelly under serious uncertainty.\n8. **Set stop trigger.** \"I will re-estimate if: (a) drawdown > X%, (b) outcomes diverge N σ over Y trials, (c) regime change invalidates edge model.\"\n\n### Output: EV-Kelly Sizing\n\n```markdown\n# EV-Kelly Sizing: \n## Repeatability: — if one-shot, STOP.\n## Bet map: p=<>, q=<>, W=<>, L=<>, b=<>\n## EV: p·W − q·L = — Edge: \n## Estimation uncertainty: ; 80% CI on edge: \n## Kelly fraction: f* = → practitioner: = \n## Stop trigger: \"I will re-estimate if .\"\n## Correlation check: bets independent? \n```\n*→ Method in Action: [Ed Thorp, Blackjack, and Princeton-Newport (1961 → 1988)](examples/ed-thorp-blackjack-and-princeton-newport-1961-1988.md) · [Bill Benter, Hong Kong Horse Racing (1985 → 2001)](examples/bill-benter-hong-kong-horse-racing-1985-2001.md)*\n*→ 2026 lens: [Sizing Hyperscaler GPU/Compute Capex (2023–2026)](examples/hyperscaler-gpu-capex-kelly-2023-2026.md) — one CFO's repeated capex bet: wide CI, stranded-capital ruin tail, and fractional Kelly as staged, survivable capex.*\n## Sizing Packs\n\n| Domain | Fractional-Kelly | Stop trigger |\n|---|---|---|\n| Active equity | quarter-Kelly or less | drawdown > 2× expected annual vol |\n| Venture capital | portfolio-level quarter-Kelly | hit rate diverges from model by vintage |\n| Ad spend by segment | half- to full Kelly | ROAS falls >2σ over N conversions |\n| A/B test ramp | fractional Kelly on traffic % | regression in primary metric |\n\n## Applying It Well\n\n- **Positive EV is necessary but not sufficient.** A positive-EV bet sized too large still wrecks you.\n- **Default to fractional Kelly.** Full Kelly is for measured edges. Half- or quarter-Kelly for estimated edges — growth cost is small, protection is large.\n- **Correlation eats Kelly fast.** All positions long tech, all VC bets in one vintage — adjust portfolio Kelly down.\n- **Kelly is for bankroll, not for life.** Use regret-minimization for career, marriage, time.\n- **Update continuously.** Static Kelly on a stale edge is how winning strategies ride into ruin.\n\n*→ Primary sources: [references/sources.md](references/sources.md)*\n## Common Rationalizations\n\n**[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.**\n\n| Fake move | Reality |\n|---|---|\n| [D] Applying Kelly to **one-shot** decisions | Kelly maximizes long-run geometric growth rate across many bets. For a one-time bet, EV is the right concern; for life decisions, use regret-minimization. |\n| [D] Treating **estimated *p*** as **known *p*** | Kelly is brutal when probability estimates are wrong. Use fractional Kelly to compensate for estimated inputs. |\n| [D] Using **full Kelly with subjective probabilities** | Full Kelly is for casino-grade certainty. Half-Kelly costs ~25% of growth and cuts ruin risk by far more (Thorp 1997). |\n| [D] Confusing **positive EV** with **\"should bet\"** | EV ignores variance, bankroll, path-dependency. A +1% EV bet that ruins you 1% of the time is not equivalent to one that ruins you 0%. |\n| [D] **Ignoring correlation** across bets | Kelly assumes independent bets. Correlated bets (same sector, same vintage) require a lower portfolio-level Kelly. |\n| [D] Using Kelly on **negative-EV** bets | Kelly returns zero or negative fraction when EV ≤ 0. No sizing rescues a -EV bet. Don't bet. |\n| [D] Treating EV as the **only number that matters** | Bernoulli 1738: utility of money is non-linear. A +$1000 EV bet that risks your rent ≠ one that risks your rounding error. |\n| [D] Forgetting Kelly's **brutal variance** | Even correct full-Kelly expects 50%+ drawdowns. Most professionals use half/quarter-Kelly to survive psychologically. |\n| [D] Applying Kelly to **non-financial \"bets\"** | Kelly assumes a compoundable bankroll. Relationships, careers, time, attention don't compound across independent trials. |\n| [D] Computing Kelly **once**, ignoring updates | A static fraction on a now-stale edge is the textbook path to ruin. Pre-commit to a re-estimation trigger. |\n| *→ Add [O] entries here after each real use — paste the actual failure pattern* | *What went wrong and why* |\n\n## Red Flags\n\n- Kelly applied to a decision that happens only once\n- Full Kelly used with subjective probabilities and no calibration check\n- EV is negative but a positive fraction is being computed\n- No stop-and-re-estimate trigger is named\n- Bets are obviously correlated but Kelly is computed per-bet without portfolio adjustment\n- The \"bankroll\" is not liquid or fungible (career time, relationship capital)\n- A historical drawdown wiped out the strategy; same fraction still applied without re-examining the edge\n\n## Verification\n\n- [ ] Decision confirmed *repeated*, not one-shot\n- [ ] Outcomes, probabilities, and payoffs stated explicitly in correct units\n- [ ] EV computed per unit staked; edge gate (positive/zero/negative) named\n- [ ] Estimation uncertainty named (measured/estimated/mixed); 80% CI on edge given\n- [ ] Kelly fraction computed with correct formula (binary vs. continuous)\n- [ ] Fractional-Kelly adjustment applied with explicit justification\n- [ ] Stop-and-re-estimate trigger named with concrete conditions\n- [ ] Correlation across bets considered; per-bet Kelly adjusted if needed\n\n---\n*Part of **deciqAI Knowledge Skills** — 164 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. **See it run → https://www.deciqai.com/c/expected-value-and-kelly** · ⭐ Star the repo → https://github.com/deciqAI/knowledge-skills · Contributions welcome.*\n\nFile v1.0.3:_meta.json\n\n{\n \"ownerId\": \"kn754b8sk22s8c6gjxt02bftbn88q7ye\",\n \"slug\": \"expected-value-and-kelly\",\n \"version\": \"1.0.3\",\n \"publishedAt\": 1783508525705\n}\n\nFile v1.0.3:references/sources.md\n\n# Sources — expected-value-and-kelly\n\n> *Primary sources for the [expected-value-and-kelly](../SKILL.md) skill.*\n\n- **Kelly, J. L., Jr.** \"A New Interpretation of Information Rate.\" *Bell System Technical Journal*, Vol. 35, No. 4 (July 1956), pp. 917–926. **The primary source for the Kelly formula and its proof.** https://www.princeton.edu/~wbialek/rome/refs/kelly_56.pdf\n- **Thorp, Edward O.** *Beat the Dealer: A Winning Strategy for the Game of Twenty-One*. Random House, 1962 (2nd ed. Vintage, 1966). **Primary source** for the first practical application of Kelly to a positive-edge real-world bet (blackjack card-counting).\n- **Thorp, Edward O.** \"The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market.\" Paper for the 10th International Conference on Gambling and Risk Taking, Montreal, 1997. The clearest concise summary by Thorp himself; covers the fractional-Kelly trade-off. https://www.eecs.harvard.edu/cs286r/courses/fall12/papers/Thorpe_KellyCriterion2007.pdf\n- **Thorp, Edward O.** *A Man for All Markets: From Las Vegas to Wall Street, How I Beat the Dealer and the Market*. Random House, 2017. Autobiography; primary source on the Princeton-Newport years and Thorp's evolution from blackjack to markets.\n- **Bernoulli, Daniel.** \"Exposition of a New Theory on the Measurement of Risk.\" Commentarii Academiae Scientiarum Imperialis Petropolitanae, Tomus V (St. Petersburg, 1738). English translation by Louise Sommer in *Econometrica*, Vol. 22, No. 1 (January 1954), pp. 23–36. **Primary source** for the utility-versus-EV distinction.\n- **Benter, William.** \"Computer Based Horse Race Handicapping and Wagering Systems: A Report.\" In *Efficiency of Racetrack Betting Markets*, edited by Donald B. Hausch, Victor S. Y. Lo, and William T. Ziemba, pp. 183–198. Academic Press, 1994 (reprinted World Scientific, 2008). **Primary source** for combining a fitted probability model with public odds and sizing pari-mutuel bets by fractional Kelly under estimated inputs and market impact.\n- **Poundstone, William.** *Fortune's Formula: The Untold Story of the Scientific Betting System That Beat the Casinos and Wall Street*. Hill & Wang, 2005. The popular history; useful as a secondary reference and for the Bell Labs / Shannon / Kelly intellectual lineage.\n- **U.S. Securities and Exchange Commission filings and quarterly earnings materials of Microsoft, Alphabet, Amazon, Meta, and Nvidia (2023–2025).** Durable **primary sources** for AI-driven data-center capital-expenditure guidance and Nvidia's data-center-segment revenue growth during the 2023–2026 AI investment boom — used in preference to fast-moving press figures, which should be verified against these filings. https://www.sec.gov/cgi-bin/browse-edgar\n- **Contemporaneous business-press coverage of frontier-AI-lab financing rounds and valuations (e.g., The Wall Street Journal, Financial Times, The Information, 2023–2026).** **Secondary** reference for reported startup valuations (OpenAI, Anthropic, and others); treat specific figures as approximate and time-stamped, since boom-era valuations are reported inconsistently and change rapidly.\n- **NVIDIA Corporation, Annual Report on Form 10-K and quarterly Form 10-Q / earnings materials (fiscal years 2024–2025).** **Primary source** for the growth of the Data Center reporting segment during the 2023–2026 AI compute boom; used in preference to press summaries of \"Nvidia's AI revenue,\" which should be verified against the segment disclosures in these filings. https://www.sec.gov/cgi-bin/browse-edgar\n- **MacLean, Leonard C., Edward O. Thorp, and William T. Ziemba (eds.).** *The Kelly Capital Growth Investment Criterion: Theory and Practice.* World Scientific, 2011. **Authoritative collected reference** on Kelly and fractional-Kelly investing, including the growth-vs-drawdown trade-off that justifies defaulting to a fraction of full Kelly when the edge is estimated rather than measured — directly relevant to sizing correlated, power-law AI bets.\n- The popular framing \"the house always wins\" is **not** cited here as a source — by this skill's own rule, an aphorism is not evidence. The Kelly criterion is precisely the formula for *the conditions under which the house does not always win* — when the player has a measurable positive edge.\n\nFile v1.0.3:examples/bill-benter-hong-kong-horse-racing-1985-2001.md\n\n# Method in Action: Bill Benter and Computer-Model Horse Betting in Hong Kong (1985 → 2001)\n\n> *Example for the [expected-value-and-kelly](../SKILL.md) skill.*\n\nA worked example in a domain far from a card table: **pari-mutuel horse racing**, where the odds are set by the crowd's own money and your bet moves the payout against you. Primary-source documented in Benter's own 1994 technical paper.\n\nIn the mid-1980s **William \"Bill\" Benter** — who had learned card-counting from Thorp's *Beat the Dealer* and been barred from casinos — turned to the Hong Kong Jockey Club, then the largest pari-mutuel racing pool in the world. His insight was statistical, not tipster's intuition: build a **multinomial logit model** that estimates each horse's true win probability from fundamental factors (past times, weight, jockey, barrier, going), then combine that estimate with the **public's implied probability** encoded in the tote odds. The crowd is a strong predictor; a model that ignores it loses, and a model that only mimics it has no edge. Benter's edge came from the *residual* — the systematic gap between his model's probabilities and the crowd's.\n\nThat residual defined the bet. On any race, a horse is a positive-EV wager only when Benter's estimated win probability *p* exceeds the break-even probability implied by its tote odds after the Jockey Club's ~17–19% takeout. Most horses in most races failed that gate. The discipline was to bet only the small fraction that cleared it — and to size each qualifying bet by Kelly.\n\nPari-mutuel racing punishes overbetting twice. First, Kelly's ordinary lesson: a positive-EV bet staked too large courts ruin through variance. Second, a mechanical trap absent from casino play — **your own money enters the pool and lowers the odds you receive.** A full-Kelly stake computed against the pre-bet odds is, after your bet lands, an *overbet* against the true post-bet payout. Benter's paper is explicit that they used **fractional Kelly**, both to absorb estimation error in the probability model and to blunt this self-impact on the tote. Run over hundreds of thousands of bets across many seasons, the syndicate is widely reported to have won on the order of a billion dollars — the integral effect of a small per-race edge, correctly sized and relentlessly repeated.\n\nWalk the EV-Kelly Sizing through Benter's case:\n\n- **Repeatability (Step 1):** Hundreds of races per season, many seasons, thousands of qualifying bets. Emphatically repeated. ✓\n- **Bet map (Step 2):** For a candidate horse, model win probability *p*; payoff *W* set by the tote (dividend minus stake) after ~17–19% takeout; loss *L* = the stake.\n- **EV (Step 3):** EV = p·W − q·L per unit staked. Computed per horse, per race.\n- **Edge gate (Step 4):** Bet only when *p* exceeds the takeout-adjusted break-even implied by the odds. The overwhelming majority of horses fail this gate and are not bet.\n- **Estimation uncertainty (Step 5):** High and irreducible — *p* comes from a fitted statistical model, not from a countable deck. This is estimated, not measured, edge.\n- **Kelly fraction (Step 6):** f\\* = (bp − q) / b per qualifying bet, with *b* from the tote odds.\n- **Fractional Kelly (Step 7):** A fraction of full Kelly — for model estimation error *and* to limit the bet's own depressive effect on the pari-mutuel payout.\n- **Stop trigger (Step 8):** Continuous model re-estimation each season; degrade or halt when out-of-sample calibration drifts or the residual edge shrinks as the pool grows more efficient.\n\nThe lesson Benter's case adds to Thorp's: when *p* is a model output and *W* is a payout your own bet degrades, full Kelly is doubly wrong. Fractional Kelly is not caution — it is the correct sizing under estimated inputs and market impact.\n\nPrimary source: Benter, William. \"Computer Based Horse Race Handicapping and Wagering Systems: A Report.\" In *Efficiency of Racetrack Betting Markets*, edited by Donald B. Hausch, Victor S. Y. Lo, and William T. Ziemba, pp. 183–198. Academic Press, 1994 (reprinted World Scientific, 2008).\n\nFile v1.0.3:examples/ed-thorp-blackjack-and-princeton-newport-1961-1988.md\n\n# Method in Action: Ed Thorp, Blackjack, and Princeton-Newport (1961 → 1988)\n\n> *Example for the [expected-value-and-kelly](../SKILL.md) skill.*\n\nA worked example. Not a casino fable — primary-source documented, including the Bell Labs origin.\n\nIn 1960, **Ed Thorp**, a young mathematics professor at MIT, was reading Baldwin et al.'s 1956 paper on basic-strategy blackjack and realized something the casino-gambling literature had missed: because cards in blackjack are drawn *without replacement* from a finite deck, the composition of the remaining deck shifts the player's edge from hand to hand. A player who **tracks** the cards already played (a \"counter\") sees a small but real positive edge on certain deck compositions — typically when the remaining deck is rich in tens and aces.\n\nThorp's contribution had two parts. First, he proved the math: a counting system (his \"Ten-Count,\" later refined to the \"High-Low\" by Harvey Dubner) yields an average edge of roughly **+1% over the house** when applied disciplinedly. Second — and this is what most blackjack popularizers miss — he applied **Kelly sizing** to the bet: vary the wager up and down with the count, betting more when the count favored the player and the minimum when it didn't.\n\nTo prove it in the field, Thorp went to Las Vegas and Reno casinos in spring 1961, funded by **Manny Kimmel**, a New York businessman with $10,000 in seed capital. Over a series of trips, they confirmed the math at the tables. The result was published in **Thorp's 1962 book *Beat the Dealer*** (Random House) — a primary source for everything in this Method-in-Action section. The casinos' second-order response was rapid: they changed rules, introduced multiple decks and shoe shuffles, and trained dealers to spot counters. Within a few years, Thorp's specific edge in single-deck games had been engineered out of most casinos.\n\nBut the *framework* — combine a measurable edge with disciplined Kelly sizing across a repeated series of independent bets — translated. In 1969, Thorp co-founded **Princeton-Newport Partners**, one of the first quantitative hedge funds. The fund's strategy was convertible-bond arbitrage, statistical arbitrage, and other measurable-edge market trades, sized using Kelly principles (Thorp typically used a fractional Kelly given the higher estimation uncertainty in markets vs. casinos). **From 1969 to 1988, the fund returned 19.1% annually before fees with no down year** (Thorp, *A Man for All Markets*, 2017, ch. 14). The Princeton-Newport partnership ended for unrelated legal reasons in 1988.\n\nWalk through the EV-Kelly Sizing on Thorp's blackjack case:\n\n- **Repeatability (Step 1):** Tens of thousands of hands per casino trip. Strongly repeated. ✓\n- **Bet map (Step 2):** Per hand at a favorable count, win probability *p ≈ 0.505*, loss *q ≈ 0.495* (the +1% edge); payoff approximately 1:1 (with some variance from doubles, splits, blackjack 3:2).\n- **EV (Step 3):** ≈ +0.01 per unit staked. *Positive but small.*\n- **Edge gate (Step 4):** Positive. Proceed.\n- **Estimation uncertainty (Step 5):** In 1961, with a measurable card-counting system and verifiable math, this was as close to **casino-grade certainty** as bets get. Estimation uncertainty was unusually low (one of the few real-world cases where near-full Kelly is defensible).\n- **Kelly fraction (Step 6):** With *p = 0.505, q = 0.495, b = 1*: f\\* = (1 × 0.505 − 0.495) / 1 = **0.01 = 1% of bankroll per hand at a favorable count**. (Thorp varied the bet *up* to roughly this level at the highest-positive counts and dropped to the table minimum at negative counts.)\n- **Fractional Kelly (Step 7):** Because the edge math was unusually well-measured, Thorp used closer to full Kelly than is typical for market trades. In later markets work at Princeton-Newport, he scaled to fractional Kelly given larger estimation uncertainty.\n- **Stop trigger (Step 8):** Casino rule changes (multiple decks, frequent shuffles) — *the casinos' second-order response* — eventually made the edge structure obsolete in most venues. Thorp re-estimated repeatedly and pivoted out.\n\nThe inversion-style lesson Thorp himself emphasized later: **the most expensive mistake in EV-Kelly is using full Kelly with an estimated edge.** Half-Kelly costs roughly 25% of growth rate but cuts variance and ruin risk by far more (Thorp 1997 conference paper, §4).\n\n**Sources:** Kelly, J. L. **\"A New Interpretation of Information Rate.\"** *Bell System Technical Journal*, Vol. 35, No. 4 (July 1956), pp. 917–926 — the primary source for the formula and its derivation. https://www.princeton.edu/~wbialek/rome/refs/kelly_56.pdf . Thorp, Edward O. ***Beat the Dealer: A Winning Strategy for the Game of Twenty-One*** (Random House, 1962; 2nd ed. Vintage 1966) — primary source for the blackjack application. Thorp, Edward O. **\"The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market.\"** Paper for the 10th International Conference on Gambling and Risk Taking, Montreal, 1997 — the clearest concise summary by Thorp himself. https://www.eecs.harvard.edu/cs286r/courses/fall12/papers/Thorpe_KellyCriterion2007.pdf . Thorp, Edward O. ***A Man for All Markets*** (Random House, 2017) — autobiography; primary source for the Princeton-Newport years. Bernoulli, Daniel. **\"Exposition of a New Theory on the Measurement of Risk\"** (St. Petersburg, 1738; English tr. Louise Sommer, *Econometrica*, Vol. 22, No. 1, January 1954, pp. 23–36) — primary source for the utility-vs-EV distinction. Poundstone, William. ***Fortune's Formula: The Untold Story of the Scientific Betting System That Beat the Casinos and Wall Street*** (Hill & Wang, 2005) — the popular history of Kelly + Thorp, useful as a secondary reference.\n\nFile v1.0.3:examples/hyperscaler-gpu-capex-kelly-2023-2026.md\n\n# Method in Action: Sizing Hyperscaler GPU/Compute Capex in the AI Boom (2023–2026)\n\n> *Example for the [expected-value-and-kelly](../SKILL.md) skill.*\n\nA worked example on the *capex* side of the 2023–2026 AI gold rush, told as a single decision-maker's repeated allocation problem. Where the companion example ([Bet-Sizing the 2023–2026 AI Gold Rush](ai-gold-rush-vc-bet-sizing-2023-2026.md)) treats a venture fund's whole portfolio, this one narrows to one recurring board-level choice: **how much of annual free cash flow should a large cloud provider commit to AI compute (GPUs, data centers, power) each planning cycle?** It is a positive-expected-value bet with a fat upside, a wide confidence interval, and a genuine ruin tail (stranded, hard-to-repurpose capital). That combination is exactly what fractional Kelly exists for.\n\nThe setting is durable and well documented. From roughly 2023 onward — after ChatGPT's late-2022 public release — the large hyperscalers (Microsoft, Alphabet, Amazon, Meta) each raised AI-related capital expenditure sharply, guiding through 2024–2025 to tens of billions of dollars per year, much of it tied to AI compute; Nvidia's data-center GPU demand drove its revenue and market capitalization to among the largest of any public company. The *structure* of the capex decision — not the specific dollar figures, which move fast and are reported inconsistently — is what the EV-Kelly lens clarifies.\n\nWalk the **EV-Kelly Sizing** through one hyperscaler's annual capex-allocation policy:\n\n- **Confirm the decision is repeated (Step 1):** This is not a one-shot. Capex is re-set every planning cycle — quarterly guidance, annual budgets, multi-year build-outs — and the *sizing policy* recurs across cycles even as any single data-center build is one-off. You are sizing the recurring allocation rule, not a single facility. Kelly applies. ✓\n- **Map the bet (Step 2):** Stake = the fraction of free cash flow (the \"bankroll\") committed to AI compute this cycle. On a **win** (AI demand monetizes: cloud consumption, ads, subscriptions, enterprise seats), payoff *W* is large and compounding — you own the scarce capacity when demand arrives. On a **loss** (demand disappoints or a cheaper compute regime arrives), loss *L* is the stranded portion of the capex: GPUs and buildings that can't be fully repurposed, depreciating on a short useful-life schedule. This is a **skewed, power-law-flavored** payoff, not a symmetric coin flip — see [power-law-distribution](../power-law-distribution/SKILL.md).\n- **Compute EV (Step 3):** EV = p·W − q·L per dollar committed. Because *W* (owning scarce AI capacity in a demand surge) is large and the marginal loss on a miss is a fraction of the stake (some capacity is reusable for general cloud), portfolio EV can be clearly positive even if the probability of the *full* upside is well under one-half. The math tolerates a modest hit rate because the win is big and the downside is bounded, not total.\n- **Edge gate (Step 4):** Positive EV is necessary but not sufficient. The edge is real only if the firm has a genuine advantage — distribution, existing cloud demand, model or data assets — over merely paying boom-era prices for capacity everyone else is also buying. If the \"edge\" is just fear of missing out, EV is illusory and no sizing rescues it.\n- **Estimate input uncertainty (Step 5):** Wide and largely irreducible. *p* (that demand monetizes before the capex is written off) and *W* are forward guesses about an unprecedented technology mid-hype-cycle — not measured frequencies (a deck) or decades-fitted (Benter's races). The 80% CI on the edge is broad and could approach zero if a capability plateau or a compute-cost collapse arrives. This alone forces a small Kelly fraction.\n- **Compute Kelly fraction (Step 6):** In the continuous framing, f\\* ≈ μ/σ². The dispersion σ of AI-capex outcomes is enormous — the range from \"generational franchise\" to \"expensive stranded silicon\" is wide — so the Kelly-optimal fraction of the free-cash-flow bankroll on *any single* cycle's bet is modest even for a genuinely positive edge. High variance shrinks Kelly.\n- **Apply fractional Kelly (Step 7):** Given both the estimation uncertainty (Step 5) and the fat-tailed variance (Step 6), the discipline is **well under full Kelly** — quarter-Kelly territory. In capex language: size the AI build so that a *wrong* bet is survivable, not fatal — retain balance-sheet slack, stage commitments so they can be paused, and prefer capacity that can fall back to general-purpose cloud. Bounded downside, retained exposure to the upside. That *is* fractional Kelly, implemented as staged, survivable capex.\n- **Set stop trigger (Step 8):** Pre-commit to re-estimating the edge when AI revenue attach-rates diverge from the model, when utilization of deployed capacity runs persistently low, when GPU price/performance shifts enough to strand the current generation, or when a capability plateau or funding-market freeze breaks the demand thesis.\n\n**Where correlation bites.** Kelly assumes independent bets; hyperscaler AI capex is anything but. The capex bet, the firm's own frontier-model bet, its cloud customers' AI startups, and the public GPU complex all rest on the same load-bearing assumptions — that capability keeps improving, that compute stays available, and that demand monetizes before write-off. When those co-move, a \"diversified\" AI position behaves like one concentrated bet. The skill's rule is blunt: **correlation eats Kelly fast** — same narrative, same cycle, same underlying assumption means a *lower* portfolio-level Kelly, not the naive sum of per-project Kellys.\n\nThe lesson this case adds to Thorp and Benter: their edges were *measured* or *fitted* on stationary processes. Hyperscaler AI capex is neither — probabilities are guesses, payoffs are skewed, ruin (stranded capital) is possible, and the bets are secretly one bet. Under those conditions full Kelly is not merely aggressive; it is a category error. **Fractional Kelly, expressed as staged and survivable capex, is the correct response to a genuinely positive but genuinely uncertain edge.** The question is not \"how big can we bet on AI?\" — it is \"how do we stay solvent and flexible long enough to still own the capacity if and when the demand actually shows up?\"\n\n*Sources: Kelly, J. L., Jr., \"A New Interpretation of Information Rate,\" Bell System Technical Journal 35(4), 1956, pp. 917–926 (the Kelly formula). Thorp, Edward O., \"The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market,\" 1997 (the fractional-Kelly trade-off under estimated edges). For the 2023–2026 hyperscaler AI-capex environment — annual capital-expenditure guidance tied to AI compute and Nvidia's data-center-driven revenue and market capitalization — see the companies' own public filings and earnings materials (Microsoft, Alphabet, Amazon, Meta, and Nvidia SEC filings and quarterly reports, 2023–2025) as durable primary sources; specific capex and valuation figures cited in contemporaneous business press (e.g., The Wall Street Journal, Financial Times, The Information, 2023–2026) should be checked against those filings, as boom-era figures move quickly and are reported inconsistently.*\n\nFile v1.0.3:skill-card.md\n\n## Description:
\nHelps agents evaluate repeated bets or capital-allocation decisions by calculating expected value, Kelly sizing, fractional Kelly adjustments, and stop triggers.
\n\nThis skill is ready for commercial/non-commercial use.
\n\n## Publisher:
\n[deciqai](https://clawhub.ai/user/deciqai)
\n\n### License/Terms of Use:
\nMIT-0
\n\n\n## Use Case:
\nExternal users and operators use this skill to decide whether a repeated financial, betting, marketing, testing, or capital-allocation decision has positive expected value and how to size it without taking ruinous risk.
\n\n### Deployment Geography for Use:
\nGlobal
\n\n## Known Risks and Mitigations:
\nRisk: Users may treat EV or Kelly outputs as guaranteed financial or betting advice.
\nMitigation: Present outputs as decision support, require calibrated assumptions, and use professional review when real money or business capital is at risk.
\nRisk: Full Kelly or stale probability estimates can oversize uncertain bets and increase drawdown or ruin risk.
\nMitigation: Default to fractional Kelly, document input uncertainty, and predefine stop-and-reestimate triggers.
\nRisk: Applying Kelly to one-shot, unestimable, illegal, or highly correlated decisions can produce misleading sizing guidance.
\nMitigation: Confirm repeatability, estimable probabilities and payoffs, legal/compliance constraints, and portfolio-level correlation before sizing.
\n\n\n## Reference(s):
\n- [ClawHub skill page](https://clawhub.ai/deciqai/skills/expected-value-and-kelly)
\n- [Sources - expected-value-and-kelly](references/sources.md)
\n- [Kelly, A New Interpretation of Information Rate](https://www.princeton.edu/~wbialek/rome/refs/kelly_56.pdf)
\n- [Thorp, The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market](https://www.eecs.harvard.edu/cs286r/courses/fall12/papers/Thorpe_KellyCriterion2007.pdf)
\n- [SEC EDGAR company filings](https://www.sec.gov/cgi-bin/browse-edgar)
\n\n\n## Skill Output:
\n**Output Type(s):** [text, markdown, guidance]
\n**Output Format:** [Markdown EV-Kelly sizing worksheet with calculations, uncertainty notes, fractional sizing, and stop triggers]
\n**Output Parameters:** [1D]
\n**Other Properties Related to Output:** [Requires user-provided probability, payoff, bankroll, and correlation assumptions; produces no executable code.]
\n\n## Skill Version(s):
\n1.0.3 (source: server release evidence)
\n\n## Ethical Considerations:
\nUsers should evaluate whether this skill is appropriate for their environment, review any generated or modified files before relying on them, and apply their organization's safety, security, and compliance requirements before deployment.
\n\nArchive v1.0.2: 6 files, 13140 bytes\n\nFiles: examples/bill-benter-hong-kong-horse-racing-1985-2001.md (4108b), examples/ed-thorp-blackjack-and-princeton-newport-1961-1988.md (5794b), references/sources.md (2595b), skill-card.md (3233b), SKILL.md (9182b), _meta.json (143b)\n\nFile v1.0.2:SKILL.md\n\n---\nname: expected-value-and-kelly\ndescription: \"Activate when: user asks 'how much should I bet/invest on this?', 'what's the expected value here?', 'Kelly criterion', 'optimal bet size', 'fractional Kelly', 'how big a position should I take?', or is allocating capital across repeated decisions (ad spend by segment, VC portfolio construction, position sizing, A/B test ramp).\n Do NOT activate when: the decision is one-shot and non-repeating (career change, marriage) — use regret-minimization instead; or when the user cannot estimate probabilities or payoffs even roughly.\"\n---\n\n# Expected Value and the Kelly Criterion\n\n## Overview\n\nTwo questions decide most repeated bets: **is this bet good?** (EV) and **how big?** (Kelly). Most professional ruin comes from positive-EV bets sized wrong. **EV = p · W − q · L.** If EV ≤ 0, do not bet. **Kelly f\\* = (bp − q) / b** maximizes long-term geometric growth (Kelly, Bell Labs, 1956). Full Kelly requires casino-grade certainty; default to half- or quarter-Kelly for estimated edges.\n\nNeighbors: first-principles · occams-razor · second-order-thinking · inversion · regret-minimization (for non-repeating life decisions).\n\n## When to Use\n\n- Decision **repeats many times** — capital allocation, position sizing, VC portfolio, ad spend, A/B test budget\n- *How big to bet* matters as much as *whether to bet*; you have a **measurable or estimable edge**\n- Someone says: \"expected value,\" \"EV,\" \"Kelly,\" \"optimal bet size,\" \"how much should we put on this?\"\n\n**When NOT to use:** one-shot life decisions → regret-minimization; negative-EV bets (don't bet); unestimable probabilities; correlated bets without portfolio adjustment.\n\n## Coaching Novices (Adaptive Front Door)\n\n**Engine mode:** user has a concrete repeated bet → run The Process directly. **Coach mode:** user is unfamiliar → guide step by step.\n\nIn Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.\n\n1. One-line what-it-is: EV tells you **whether the bet is worth taking**; Kelly tells you **what fraction of bankroll to stake** — sized to maximize long-term growth without ruin.\n2. Check fit against When to Use / When NOT to use. If one-shot life decision, redirect to regret-minimization. If EV is negative, say \"don't bet\" and stop.\n3. Elicit their real bet. Ask for a concrete repeated decision with measurable inputs. > **[WAIT — do not advance until user responds]**\n4. Walk The Process one step per turn: outcomes → probabilities → payoffs → EV → Kelly → fractional Kelly. > **[WAIT — do not advance until user responds]**\n5. Close by naming their sizing rule: \"bet f\\* × bankroll, use half-Kelly given estimation uncertainty\" — and the trigger that would change it. > **[WAIT — do not advance until user responds]**\n\n## The Process\n\nRun the **EV-Kelly Sizing** (EV → Kelly → fractional Kelly → stop trigger):\n\n1. **Confirm the decision is repeated.** If \"once,\" stop → use regret-minimization.\n2. **Map the bet.** Win prob *p*, loss prob *q = 1−p*, payoff on win *W*, loss *L*, odds *b = W/L*.\n3. **Compute EV.** EV = Σ(pᵢ · payoffᵢ). State per unit staked.\n4. **Edge gate.** EV > 0? If no, stop — do not bet.\n5. **Estimate input uncertainty.** Are *p* and *W* measured or estimated? Write 80% CI on edge.\n6. **Compute Kelly fraction.** f\\* = (bp − q) / b. For continuous: f\\* ≈ μ/σ².\n7. **Apply fractional Kelly.** Half-Kelly under modest uncertainty; quarter-Kelly under serious uncertainty.\n8. **Set stop trigger.** \"I will re-estimate if: (a) drawdown > X%, (b) outcomes diverge N σ over Y trials, (c) regime change invalidates edge model.\"\n\n### Output: EV-Kelly Sizing\n\n```markdown\n# EV-Kelly Sizing: \n## Repeatability: — if one-shot, STOP.\n## Bet map: p=<>, q=<>, W=<>, L=<>, b=<>\n## EV: p·W − q·L = — Edge: \n## Estimation uncertainty: ; 80% CI on edge: \n## Kelly fraction: f* = → practitioner: = \n## Stop trigger: \"I will re-estimate if .\"\n## Correlation check: bets independent? \n```\n*→ Method in Action: [Ed Thorp, Blackjack, and Princeton-Newport (1961 → 1988)](examples/ed-thorp-blackjack-and-princeton-newport-1961-1988.md) · [Bill Benter, Hong Kong Horse Racing (1985 → 2001)](examples/bill-benter-hong-kong-horse-racing-1985-2001.md)*\n## Sizing Packs\n\n| Domain | Fractional-Kelly | Stop trigger |\n|---|---|---|\n| Active equity | quarter-Kelly or less | drawdown > 2× expected annual vol |\n| Venture capital | portfolio-level quarter-Kelly | hit rate diverges from model by vintage |\n| Ad spend by segment | half- to full Kelly | ROAS falls >2σ over N conversions |\n| A/B test ramp | fractional Kelly on traffic % | regression in primary metric |\n\n## Applying It Well\n\n- **Positive EV is necessary but not sufficient.** A positive-EV bet sized too large still wrecks you.\n- **Default to fractional Kelly.** Full Kelly is for measured edges. Half- or quarter-Kelly for estimated edges — growth cost is small, protection is large.\n- **Correlation eats Kelly fast.** All positions long tech, all VC bets in one vintage — adjust portfolio Kelly down.\n- **Kelly is for bankroll, not for life.** Use regret-minimization for career, marriage, time.\n- **Update continuously.** Static Kelly on a stale edge is how winning strategies ride into ruin.\n\n*→ Primary sources: [references/sources.md](references/sources.md)*\n## Common Rationalizations\n\n**[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.**\n\n| Fake move | Reality |\n|---|---|\n| [D] Applying Kelly to **one-shot** decisions | Kelly maximizes long-run geometric growth rate across many bets. For a one-time bet, EV is the right concern; for life decisions, use regret-minimization. |\n| [D] Treating **estimated *p*** as **known *p*** | Kelly is brutal when probability estimates are wrong. Use fractional Kelly to compensate for estimated inputs. |\n| [D] Using **full Kelly with subjective probabilities** | Full Kelly is for casino-grade certainty. Half-Kelly costs ~25% of growth and cuts ruin risk by far more (Thorp 1997). |\n| [D] Confusing **positive EV** with **\"should bet\"** | EV ignores variance, bankroll, path-dependency. A +1% EV bet that ruins you 1% of the time is not equivalent to one that ruins you 0%. |\n| [D] **Ignoring correlation** across bets | Kelly assumes independent bets. Correlated bets (same sector, same vintage) require a lower portfolio-level Kelly. |\n| [D] Using Kelly on **negative-EV** bets | Kelly returns zero or negative fraction when EV ≤ 0. No sizing rescues a -EV bet. Don't bet. |\n| [D] Treating EV as the **only number that matters** | Bernoulli 1738: utility of money is non-linear. A +$1000 EV bet that risks your rent ≠ one that risks your rounding error. |\n| [D] Forgetting Kelly's **brutal variance** | Even correct full-Kelly expects 50%+ drawdowns. Most professionals use half/quarter-Kelly to survive psychologically. |\n| [D] Applying Kelly to **non-financial \"bets\"** | Kelly assumes a compoundable bankroll. Relationships, careers, time, attention don't compound across independent trials. |\n| [D] Computing Kelly **once**, ignoring updates | A static fraction on a now-stale edge is the textbook path to ruin. Pre-commit to a re-estimation trigger. |\n| *→ Add [O] entries here after each real use — paste the actual failure pattern* | *What went wrong and why* |\n\n## Red Flags\n\n- Kelly applied to a decision that happens only once\n- Full Kelly used with subjective probabilities and no calibration check\n- EV is negative but a positive fraction is being computed\n- No stop-and-re-estimate trigger is named\n- Bets are obviously correlated but Kelly is computed per-bet without portfolio adjustment\n- The \"bankroll\" is not liquid or fungible (career time, relationship capital)\n- A historical drawdown wiped out the strategy; same fraction still applied without re-examining the edge\n\n## Verification\n\n- [ ] Decision confirmed *repeated*, not one-shot\n- [ ] Outcomes, probabilities, and payoffs stated explicitly in correct units\n- [ ] EV computed per unit staked; edge gate (positive/zero/negative) named\n- [ ] Estimation uncertainty named (measured/estimated/mixed); 80% CI on edge given\n- [ ] Kelly fraction computed with correct formula (binary vs. continuous)\n- [ ] Fractional-Kelly adjustment applied with explicit justification\n- [ ] Stop-and-re-estimate trigger named with concrete conditions\n- [ ] Correlation across bets considered; per-bet Kelly adjusted if needed\n\n---\n*Part of **deciqAI Knowledge Skills** — 163 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. **See it run → https://www.deciqai.com/skills/expected-value-and-kelly?utm_source=clawhub&utm_medium=marketplace&utm_campaign=knowledge-skills&utm_content=expected-value-and-kelly** · ⭐ Star the repo → https://github.com/deciqAI/knowledge-skills · Contributions welcome.*\n\nFile v1.0.2:_meta.json\n\n{\n \"ownerId\": \"kn754b8sk22s8c6gjxt02bftbn88q7ye\",\n \"slug\": \"expected-value-and-kelly\",\n \"version\": \"1.0.2\",\n \"publishedAt\": 1783471648602\n}\n\nFile v1.0.2:references/sources.md\n\n# Sources — expected-value-and-kelly\n\n> *Primary sources for the [expected-value-and-kelly](../SKILL.md) skill.*\n\n- **Kelly, J. L., Jr.** \"A New Interpretation of Information Rate.\" *Bell System Technical Journal*, Vol. 35, No. 4 (July 1956), pp. 917–926. **The primary source for the Kelly formula and its proof.** https://www.princeton.edu/~wbialek/rome/refs/kelly_56.pdf\n- **Thorp, Edward O.** *Beat the Dealer: A Winning Strategy for the Game of Twenty-One*. Random House, 1962 (2nd ed. Vintage, 1966). **Primary source** for the first practical application of Kelly to a positive-edge real-world bet (blackjack card-counting).\n- **Thorp, Edward O.** \"The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market.\" Paper for the 10th International Conference on Gambling and Risk Taking, Montreal, 1997. The clearest concise summary by Thorp himself; covers the fractional-Kelly trade-off. https://www.eecs.harvard.edu/cs286r/courses/fall12/papers/Thorpe_KellyCriterion2007.pdf\n- **Thorp, Edward O.** *A Man for All Markets: From Las Vegas to Wall Street, How I Beat the Dealer and the Market*. Random House, 2017. Autobiography; primary source on the Princeton-Newport years and Thorp's evolution from blackjack to markets.\n- **Bernoulli, Daniel.** \"Exposition of a New Theory on the Measurement of Risk.\" Commentarii Academiae Scientiarum Imperialis Petropolitanae, Tomus V (St. Petersburg, 1738). English translation by Louise Sommer in *Econometrica*, Vol. 22, No. 1 (January 1954), pp. 23–36. **Primary source** for the utility-versus-EV distinction.\n- **Benter, William.** \"Computer Based Horse Race Handicapping and Wagering Systems: A Report.\" In *Efficiency of Racetrack Betting Markets*, edited by Donald B. Hausch, Victor S. Y. Lo, and William T. Ziemba, pp. 183–198. Academic Press, 1994 (reprinted World Scientific, 2008). **Primary source** for combining a fitted probability model with public odds and sizing pari-mutuel bets by fractional Kelly under estimated inputs and market impact.\n- **Poundstone, William.** *Fortune's Formula: The Untold Story of the Scientific Betting System That Beat the Casinos and Wall Street*. Hill & Wang, 2005. The popular history; useful as a secondary reference and for the Bell Labs / Shannon / Kelly intellectual lineage.\n- The popular framing \"the house always wins\" is **not** cited here as a source — by this skill's own rule, an aphorism is not evidence. The Kelly criterion is precisely the formula for *the conditions under which the house does not always win* — when the player has a measurable positive edge.\n\nFile v1.0.2:examples/bill-benter-hong-kong-horse-racing-1985-2001.md\n\n# Method in Action: Bill Benter and Computer-Model Horse Betting in Hong Kong (1985 → 2001)\n\n> *Example for the [expected-value-and-kelly](../SKILL.md) skill.*\n\nA worked example in a domain far from a card table: **pari-mutuel horse racing**, where the odds are set by the crowd's own money and your bet moves the payout against you. Primary-source documented in Benter's own 1994 technical paper.\n\nIn the mid-1980s **William \"Bill\" Benter** — who had learned card-counting from Thorp's *Beat the Dealer* and been barred from casinos — turned to the Hong Kong Jockey Club, then the largest pari-mutuel racing pool in the world. His insight was statistical, not tipster's intuition: build a **multinomial logit model** that estimates each horse's true win probability from fundamental factors (past times, weight, jockey, barrier, going), then combine that estimate with the **public's implied probability** encoded in the tote odds. The crowd is a strong predictor; a model that ignores it loses, and a model that only mimics it has no edge. Benter's edge came from the *residual* — the systematic gap between his model's probabilities and the crowd's.\n\nThat residual defined the bet. On any race, a horse is a positive-EV wager only when Benter's estimated win probability *p* exceeds the break-even probability implied by its tote odds after the Jockey Club's ~17–19% takeout. Most horses in most races failed that gate. The discipline was to bet only the small fraction that cleared it — and to size each qualifying bet by Kelly.\n\nPari-mutuel racing punishes overbetting twice. First, Kelly's ordinary lesson: a positive-EV bet staked too large courts ruin through variance. Second, a mechanical trap absent from casino play — **your own money enters the pool and lowers the odds you receive.** A full-Kelly stake computed against the pre-bet odds is, after your bet lands, an *overbet* against the true post-bet payout. Benter's paper is explicit that they used **fractional Kelly**, both to absorb estimation error in the probability model and to blunt this self-impact on the tote. Run over hundreds of thousands of bets across many seasons, the syndicate is widely reported to have won on the order of a billion dollars — the integral effect of a small per-race edge, correctly sized and relentlessly repeated.\n\nWalk the EV-Kelly Sizing through Benter's case:\n\n- **Repeatability (Step 1):** Hundreds of races per season, many seasons, thousands of qualifying bets. Emphatically repeated. ✓\n- **Bet map (Step 2):** For a candidate horse, model win probability *p*; payoff *W* set by the tote (dividend minus stake) after ~17–19% takeout; loss *L* = the stake.\n- **EV (Step 3):** EV = p·W − q·L per unit staked. Computed per horse, per race.\n- **Edge gate (Step 4):** Bet only when *p* exceeds the takeout-adjusted break-even implied by the odds. The overwhelming majority of horses fail this gate and are not bet.\n- **Estimation uncertainty (Step 5):** High and irreducible — *p* comes from a fitted statistical model, not from a countable deck. This is estimated, not measured, edge.\n- **Kelly fraction (Step 6):** f\\* = (bp − q) / b per qualifying bet, with *b* from the tote odds.\n- **Fractional Kelly (Step 7):** A fraction of full Kelly — for model estimation error *and* to limit the bet's own depressive effect on the pari-mutuel payout.\n- **Stop trigger (Step 8):** Continuous model re-estimation each season; degrade or halt when out-of-sample calibration drifts or the residual edge shrinks as the pool grows more efficient.\n\nThe lesson Benter's case adds to Thorp's: when *p* is a model output and *W* is a payout your own bet degrades, full Kelly is doubly wrong. Fractional Kelly is not caution — it is the correct sizing under estimated inputs and market impact.\n\nPrimary source: Benter, William. \"Computer Based Horse Race Handicapping and Wagering Systems: A Report.\" In *Efficiency of Racetrack Betting Markets*, edited by Donald B. Hausch, Victor S. Y. Lo, and William T. Ziemba, pp. 183–198. Academic Press, 1994 (reprinted World Scientific, 2008).\n\nFile v1.0.2:examples/ed-thorp-blackjack-and-princeton-newport-1961-1988.md\n\n# Method in Action: Ed Thorp, Blackjack, and Princeton-Newport (1961 → 1988)\n\n> *Example for the [expected-value-and-kelly](../SKILL.md) skill.*\n\nA worked example. Not a casino fable — primary-source documented, including the Bell Labs origin.\n\nIn 1960, **Ed Thorp**, a young mathematics professor at MIT, was reading Baldwin et al.'s 1956 paper on basic-strategy blackjack and realized something the casino-gambling literature had missed: because cards in blackjack are drawn *without replacement* from a finite deck, the composition of the remaining deck shifts the player's edge from hand to hand. A player who **tracks** the cards already played (a \"counter\") sees a small but real positive edge on certain deck compositions — typically when the remaining deck is rich in tens and aces.\n\nThorp's contribution had two parts. First, he proved the math: a counting system (his \"Ten-Count,\" later refined to the \"High-Low\" by Harvey Dubner) yields an average edge of roughly **+1% over the house** when applied disciplinedly. Second — and this is what most blackjack popularizers miss — he applied **Kelly sizing** to the bet: vary the wager up and down with the count, betting more when the count favored the player and the minimum when it didn't.\n\nTo prove it in the field, Thorp went to Las Vegas and Reno casinos in spring 1961, funded by **Manny Kimmel**, a New York businessman with $10,000 in seed capital. Over a series of trips, they confirmed the math at the tables. The result was published in **Thorp's 1962 book *Beat the Dealer*** (Random House) — a primary source for everything in this Method-in-Action section. The casinos' second-order response was rapid: they changed rules, introduced multiple decks and shoe shuffles, and trained dealers to spot counters. Within a few years, Thorp's specific edge in single-deck games had been engineered out of most casinos.\n\nBut the *framework* — combine a measurable edge with disciplined Kelly sizing across a repeated series of independent bets — translated. In 1969, Thorp co-founded **Princeton-Newport Partners**, one of the first quantitative hedge funds. The fund's strategy was convertible-bond arbitrage, statistical arbitrage, and other measurable-edge market trades, sized using Kelly principles (Thorp typically used a fractional Kelly given the higher estimation uncertainty in markets vs. casinos). **From 1969 to 1988, the fund returned 19.1% annually before fees with no down year** (Thorp, *A Man for All Markets*, 2017, ch. 14). The Princeton-Newport partnership ended for unrelated legal reasons in 1988.\n\nWalk through the EV-Kelly Sizing on Thorp's blackjack case:\n\n- **Repeatability (Step 1):** Tens of thousands of hands per casino trip. Strongly repeated. ✓\n- **Bet map (Step 2):** Per hand at a favorable count, win probability *p ≈ 0.505*, loss *q ≈ 0.495* (the +1% edge); payoff approximately 1:1 (with some variance from doubles, splits, blackjack 3:2).\n- **EV (Step 3):** ≈ +0.01 per unit staked. *Positive but small.*\n- **Edge gate (Step 4):** Positive. Proceed.\n- **Estimation uncertainty (Step 5):** In 1961, with a measurable card-counting system and verifiable math, this was as close to **casino-grade certainty** as bets get. Estimation uncertainty was unusually low (one of the few real-world cases where near-full Kelly is defensible).\n- **Kelly fraction (Step 6):** With *p = 0.505, q = 0.495, b = 1*: f\\* = (1 × 0.505 − 0.495) / 1 = **0.01 = 1% of bankroll per hand at a favorable count**. (Thorp varied the bet *up* to roughly this level at the highest-positive counts and dropped to the table minimum at negative counts.)\n- **Fractional Kelly (Step 7):** Because the edge math was unusually well-measured, Thorp used closer to full Kelly than is typical for market trades. In later markets work at Princeton-Newport, he scaled to fractional Kelly given larger estimation uncertainty.\n- **Stop trigger (Step 8):** Casino rule changes (multiple decks, frequent shuffles) — *the casinos' second-order response* — eventually made the edge structure obsolete in most venues. Thorp re-estimated repeatedly and pivoted out.\n\nThe inversion-style lesson Thorp himself emphasized later: **the most expensive mistake in EV-Kelly is using full Kelly with an estimated edge.** Half-Kelly costs roughly 25% of growth rate but cuts variance and ruin risk by far more (Thorp 1997 conference paper, §4).\n\n**Sources:** Kelly, J. L. **\"A New Interpretation of Information Rate.\"** *Bell System Technical Journal*, Vol. 35, No. 4 (July 1956), pp. 917–926 — the primary source for the formula and its derivation. https://www.princeton.edu/~wbialek/rome/refs/kelly_56.pdf . Thorp, Edward O. ***Beat the Dealer: A Winning Strategy for the Game of Twenty-One*** (Random House, 1962; 2nd ed. Vintage 1966) — primary source for the blackjack application. Thorp, Edward O. **\"The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market.\"** Paper for the 10th International Conference on Gambling and Risk Taking, Montreal, 1997 — the clearest concise summary by Thorp himself. https://www.eecs.harvard.edu/cs286r/courses/fall12/papers/Thorpe_KellyCriterion2007.pdf . Thorp, Edward O. ***A Man for All Markets*** (Random House, 2017) — autobiography; primary source for the Princeton-Newport years. Bernoulli, Daniel. **\"Exposition of a New Theory on the Measurement of Risk\"** (St. Petersburg, 1738; English tr. Louise Sommer, *Econometrica*, Vol. 22, No. 1, January 1954, pp. 23–36) — primary source for the utility-vs-EV distinction. Poundstone, William. ***Fortune's Formula: The Untold Story of the Scientific Betting System That Beat the Casinos and Wall Street*** (Hill & Wang, 2005) — the popular history of Kelly + Thorp, useful as a secondary reference.\n\nFile v1.0.2:skill-card.md\n\n## Description:
\nGuides agents through expected value and Kelly criterion sizing for repeated capital allocation, betting, ad spend, portfolio, or test-ramp decisions where probabilities and payoffs can be estimated.
\n\nThis skill is ready for commercial/non-commercial use.
\n\n## Publisher:
\n[deciqai](https://clawhub.ai/user/deciqai)
\n\n### License/Terms of Use:
\nMIT-0
\n\n\n## Use Case:
\nDevelopers and external users use this skill to help an agent decide whether a repeated decision has positive expected value and how to size exposure using full or fractional Kelly. It is suited to repeated financial, betting, ad allocation, venture portfolio, and experimentation decisions with explicit probabilities, payoffs, uncertainty, and stop triggers.
\n\n### Deployment Geography for Use:
\nGlobal
\n\n## Known Risks and Mitigations:
\nRisk: Users may treat EV or Kelly output as a guarantee of profit or personalized financial advice.
\nMitigation: Present outputs as decision-support only and require users to independently verify probabilities, payoff assumptions, legal constraints, and risk tolerance before acting.
\nRisk: Kelly sizing can be misleading when probabilities, payoffs, independence, or bankroll assumptions are wrong.
\nMitigation: Use fractional Kelly by default for estimated edges, include uncertainty ranges, check correlation, and define stop-and-re-estimate triggers.
\nRisk: The method is not appropriate for one-shot life decisions or cases where probabilities and payoffs cannot be estimated.
\nMitigation: Confirm the decision is repeated and measurable before applying the process; redirect one-shot decisions to another decision framework.
\n\n\n## Reference(s):
\n- [Sources - expected-value-and-kelly](references/sources.md)
\n- [Ed Thorp, Blackjack, and Princeton-Newport](examples/ed-thorp-blackjack-and-princeton-newport-1961-1988.md)
\n- [Bill Benter and Computer-Model Horse Betting in Hong Kong](examples/bill-benter-hong-kong-horse-racing-1985-2001.md)
\n- [Kelly, J. L., Jr., A New Interpretation of Information Rate](https://www.princeton.edu/~wbialek/rome/refs/kelly_56.pdf)
\n- [Thorp, Edward O., The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market](https://www.eecs.harvard.edu/cs286r/courses/fall12/papers/Thorpe_KellyCriterion2007.pdf)
\n- [ClawHub skill page](https://clawhub.ai/deciqai/skills/expected-value-and-kelly)
\n\n\n## Skill Output:
\n**Output Type(s):** [Text, Markdown, Guidance]
\n**Output Format:** [Markdown decision-support report with structured EV-Kelly sizing fields]
\n**Output Parameters:** [1D]
\n**Other Properties Related to Output:** [May include step-by-step coaching questions, expected value calculations, Kelly or fractional-Kelly sizing, uncertainty notes, stop triggers, and correlation checks.]
\n\n## Skill Version(s):
\n1.0.2 (source: server release evidence)
\n\n## Ethical Considerations:
\nUsers should evaluate whether this skill is appropriate for their environment, review any generated or modified files before relying on them, and apply their organization's safety, security, and compliance requirements before deployment.
\n\nArchive v1.0.1: 5 files, 10423 bytes\n\nFiles: examples/ed-thorp-blackjack-and-princeton-newport-1961-1988.md (5794b), references/sources.md (2146b), skill-card.md (3095b), SKILL.md (9123b), _meta.json (143b)\n\nFile v1.0.1:SKILL.md\n\n---\nname: expected-value-and-kelly\ndescription: \"Activate when: user asks 'how much should I bet/invest on this?', 'what's the expected value here?', 'Kelly criterion', 'optimal bet size', 'fractional Kelly', 'how big a position should I take?', or is allocating capital across repeated decisions (ad spend by segment, VC portfolio construction, position sizing, A/B test ramp).\n Do NOT activate when: the decision is one-shot and non-repeating (career change, marriage) — use regret-minimization instead; or when the user cannot estimate probabilities or payoffs even roughly.\"\n---\n\n# Expected Value and the Kelly Criterion\n\n## Overview\n\nTwo questions decide most repeated bets: **is this bet good?** (EV) and **how big?** (Kelly). Most professional ruin comes from positive-EV bets sized wrong. **EV = p · W − q · L.** If EV ≤ 0, do not bet. **Kelly f\\* = (bp − q) / b** maximizes long-term geometric growth (Kelly, Bell Labs, 1956). Full Kelly requires casino-grade certainty; default to half- or quarter-Kelly for estimated edges.\n\nNeighbors: [first-principles](../first-principles/SKILL.md) · [occams-razor](../occams-razor/SKILL.md) · [second-order-thinking](../second-order-thinking/SKILL.md) · [inversion](../inversion/SKILL.md) · [regret-minimization](../regret-minimization/SKILL.md) (for non-repeating life decisions).\n\n## When to Use\n\n- Decision **repeats many times** — capital allocation, position sizing, VC portfolio, ad spend, A/B test budget\n- *How big to bet* matters as much as *whether to bet*; you have a **measurable or estimable edge**\n- Someone says: \"expected value,\" \"EV,\" \"Kelly,\" \"optimal bet size,\" \"how much should we put on this?\"\n\n**When NOT to use:** one-shot life decisions → [regret-minimization](../regret-minimization/SKILL.md); negative-EV bets (don't bet); unestimable probabilities; correlated bets without portfolio adjustment.\n\n## Coaching Novices (Adaptive Front Door)\n\n**Engine mode:** user has a concrete repeated bet → run The Process directly. **Coach mode:** user is unfamiliar → guide step by step.\n\nIn Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.\n\n1. One-line what-it-is: EV tells you **whether the bet is worth taking**; Kelly tells you **what fraction of bankroll to stake** — sized to maximize long-term growth without ruin.\n2. Check fit against When to Use / When NOT to use. If one-shot life decision, redirect to [regret-minimization](../regret-minimization/SKILL.md). If EV is negative, say \"don't bet\" and stop.\n3. Elicit their real bet. Ask for a concrete repeated decision with measurable inputs. > **[WAIT — do not advance until user responds]**\n4. Walk The Process one step per turn: outcomes → probabilities → payoffs → EV → Kelly → fractional Kelly. > **[WAIT — do not advance until user responds]**\n5. Close by naming their sizing rule: \"bet f\\* × bankroll, use half-Kelly given estimation uncertainty\" — and the trigger that would change it. > **[WAIT — do not advance until user responds]**\n\n## The Process\n\nRun the **EV-Kelly Sizing** (EV → Kelly → fractional Kelly → stop trigger):\n\n1. **Confirm the decision is repeated.** If \"once,\" stop → use [regret-minimization](../regret-minimization/SKILL.md).\n2. **Map the bet.** Win prob *p*, loss prob *q = 1−p*, payoff on win *W*, loss *L*, odds *b = W/L*.\n3. **Compute EV.** EV = Σ(pᵢ · payoffᵢ). State per unit staked.\n4. **Edge gate.** EV > 0? If no, stop — do not bet.\n5. **Estimate input uncertainty.** Are *p* and *W* measured or estimated? Write 80% CI on edge.\n6. **Compute Kelly fraction.** f\\* = (bp − q) / b. For continuous: f\\* ≈ μ/σ².\n7. **Apply fractional Kelly.** Half-Kelly under modest uncertainty; quarter-Kelly under serious uncertainty.\n8. **Set stop trigger.** \"I will re-estimate if: (a) drawdown > X%, (b) outcomes diverge N σ over Y trials, (c) regime change invalidates edge model.\"\n\n### Output: EV-Kelly Sizing\n\n```markdown\n# EV-Kelly Sizing: \n## Repeatability: — if one-shot, STOP.\n## Bet map: p=<>, q=<>, W=<>, L=<>, b=<>\n## EV: p·W − q·L = — Edge: \n## Estimation uncertainty: ; 80% CI on edge: \n## Kelly fraction: f* = → practitioner: = \n## Stop trigger: \"I will re-estimate if .\"\n## Correlation check: bets independent? \n```\n*→ Method in Action: [Ed Thorp, Blackjack, and Princeton-Newport (1961 → 1988)](examples/ed-thorp-blackjack-and-princeton-newport-1961-1988.md)*\n## Sizing Packs\n\n| Domain | Fractional-Kelly | Stop trigger |\n|---|---|---|\n| Active equity | quarter-Kelly or less | drawdown > 2× expected annual vol |\n| Venture capital | portfolio-level quarter-Kelly | hit rate diverges from model by vintage |\n| Ad spend by segment | half- to full Kelly | ROAS falls >2σ over N conversions |\n| A/B test ramp | fractional Kelly on traffic % | regression in primary metric |\n\n## Applying It Well\n\n- **Positive EV is necessary but not sufficient.** A positive-EV bet sized too large still wrecks you.\n- **Default to fractional Kelly.** Full Kelly is for measured edges. Half- or quarter-Kelly for estimated edges — growth cost is small, protection is large.\n- **Correlation eats Kelly fast.** All positions long tech, all VC bets in one vintage — adjust portfolio Kelly down.\n- **Kelly is for bankroll, not for life.** Use [regret-minimization](../regret-minimization/SKILL.md) for career, marriage, time.\n- **Update continuously.** Static Kelly on a stale edge is how winning strategies ride into ruin.\n\n*→ Primary sources: [references/sources.md](references/sources.md)*\n## Common Rationalizations\n\n**[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.**\n\n| Fake move | Reality |\n|---|---|\n| [D] Applying Kelly to **one-shot** decisions | Kelly maximizes long-run geometric growth rate across many bets. For a one-time bet, EV is the right concern; for life decisions, use [regret-minimization](../regret-minimization/SKILL.md). |\n| [D] Treating **estimated *p*** as **known *p*** | Kelly is brutal when probability estimates are wrong. Use fractional Kelly to compensate for estimated inputs. |\n| [D] Using **full Kelly with subjective probabilities** | Full Kelly is for casino-grade certainty. Half-Kelly costs ~25% of growth and cuts ruin risk by far more (Thorp 1997). |\n| [D] Confusing **positive EV** with **\"should bet\"** | EV ignores variance, bankroll, path-dependency. A +1% EV bet that ruins you 1% of the time is not equivalent to one that ruins you 0%. |\n| [D] **Ignoring correlation** across bets | Kelly assumes independent bets. Correlated bets (same sector, same vintage) require a lower portfolio-level Kelly. |\n| [D] Using Kelly on **negative-EV** bets | Kelly returns zero or negative fraction when EV ≤ 0. No sizing rescues a -EV bet. Don't bet. |\n| [D] Treating EV as the **only number that matters** | Bernoulli 1738: utility of money is non-linear. A +$1000 EV bet that risks your rent ≠ one that risks your rounding error. |\n| [D] Forgetting Kelly's **brutal variance** | Even correct full-Kelly expects 50%+ drawdowns. Most professionals use half/quarter-Kelly to survive psychologically. |\n| [D] Applying Kelly to **non-financial \"bets\"** | Kelly assumes a compoundable bankroll. Relationships, careers, time, attention don't compound across independent trials. |\n| [D] Computing Kelly **once**, ignoring updates | A static fraction on a now-stale edge is the textbook path to ruin. Pre-commit to a re-estimation trigger. |\n| *→ Add [O] entries here after each real use — paste the actual failure pattern* | *What went wrong and why* |\n\n## Red Flags\n\n- Kelly applied to a decision that happens only once\n- Full Kelly used with subjective probabilities and no calibration check\n- EV is negative but a positive fraction is being computed\n- No stop-and-re-estimate trigger is named\n- Bets are obviously correlated but Kelly is computed per-bet without portfolio adjustment\n- The \"bankroll\" is not liquid or fungible (career time, relationship capital)\n- A historical drawdown wiped out the strategy; same fraction still applied without re-examining the edge\n\n## Verification\n\n- [ ] Decision confirmed *repeated*, not one-shot\n- [ ] Outcomes, probabilities, and payoffs stated explicitly in correct units\n- [ ] EV computed per unit staked; edge gate (positive/zero/negative) named\n- [ ] Estimation uncertainty named (measured/estimated/mixed); 80% CI on edge given\n- [ ] Kelly fraction computed with correct formula (binary vs. continuous)\n- [ ] Fractional-Kelly adjustment applied with explicit justification\n- [ ] Stop-and-re-estimate trigger named with concrete conditions\n- [ ] Correlation across bets considered; per-bet Kelly adjusted if needed\n\n---\n*Part of **deciqAI Knowledge Skills** — open-source thinking skills that make rigor executable for AI agents. Built by deciqAI · https://deciqai.com · Contributions welcome — see the template at the repo root.*\n\nFile v1.0.1:_meta.json\n\n{\n \"ownerId\": \"kn754b8sk22s8c6gjxt02bftbn88q7ye\",\n \"slug\": \"expected-value-and-kelly\",\n \"version\": \"1.0.1\",\n \"publishedAt\": 1783456390053\n}\n\nFile v1.0.1:references/sources.md\n\n# Sources — expected-value-and-kelly\n\n> *Primary sources for the [expected-value-and-kelly](../SKILL.md) skill.*\n\n- **Kelly, J. L., Jr.** \"A New Interpretation of Information Rate.\" *Bell System Technical Journal*, Vol. 35, No. 4 (July 1956), pp. 917–926. **The primary source for the Kelly formula and its proof.** https://www.princeton.edu/~wbialek/rome/refs/kelly_56.pdf\n- **Thorp, Edward O.** *Beat the Dealer: A Winning Strategy for the Game of Twenty-One*. Random House, 1962 (2nd ed. Vintage, 1966). **Primary source** for the first practical application of Kelly to a positive-edge real-world bet (blackjack card-counting).\n- **Thorp, Edward O.** \"The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market.\" Paper for the 10th International Conference on Gambling and Risk Taking, Montreal, 1997. The clearest concise summary by Thorp himself; covers the fractional-Kelly trade-off. https://www.eecs.harvard.edu/cs286r/courses/fall12/papers/Thorpe_KellyCriterion2007.pdf\n- **Thorp, Edward O.** *A Man for All Markets: From Las Vegas to Wall Street, How I Beat the Dealer and the Market*. Random House, 2017. Autobiography; primary source on the Princeton-Newport years and Thorp's evolution from blackjack to markets.\n- **Bernoulli, Daniel.** \"Exposition of a New Theory on the Measurement of Risk.\" Commentarii Academiae Scientiarum Imperialis Petropolitanae, Tomus V (St. Petersburg, 1738). English translation by Louise Sommer in *Econometrica*, Vol. 22, No. 1 (January 1954), pp. 23–36. **Primary source** for the utility-versus-EV distinction.\n- **Poundstone, William.** *Fortune's Formula: The Untold Story of the Scientific Betting System That Beat the Casinos and Wall Street*. Hill & Wang, 2005. The popular history; useful as a secondary reference and for the Bell Labs / Shannon / Kelly intellectual lineage.\n- The popular framing \"the house always wins\" is **not** cited here as a source — by this skill's own rule, an aphorism is not evidence. The Kelly criterion is precisely the formula for *the conditions under which the house does not always win* — when the player has a measurable positive edge.\n\nFile v1.0.1:examples/ed-thorp-blackjack-and-princeton-newport-1961-1988.md\n\n# Method in Action: Ed Thorp, Blackjack, and Princeton-Newport (1961 → 1988)\n\n> *Example for the [expected-value-and-kelly](../SKILL.md) skill.*\n\nA worked example. Not a casino fable — primary-source documented, including the Bell Labs origin.\n\nIn 1960, **Ed Thorp**, a young mathematics professor at MIT, was reading Baldwin et al.'s 1956 paper on basic-strategy blackjack and realized something the casino-gambling literature had missed: because cards in blackjack are drawn *without replacement* from a finite deck, the composition of the remaining deck shifts the player's edge from hand to hand. A player who **tracks** the cards already played (a \"counter\") sees a small but real positive edge on certain deck compositions — typically when the remaining deck is rich in tens and aces.\n\nThorp's contribution had two parts. First, he proved the math: a counting system (his \"Ten-Count,\" later refined to the \"High-Low\" by Harvey Dubner) yields an average edge of roughly **+1% over the house** when applied disciplinedly. Second — and this is what most blackjack popularizers miss — he applied **Kelly sizing** to the bet: vary the wager up and down with the count, betting more when the count favored the player and the minimum when it didn't.\n\nTo prove it in the field, Thorp went to Las Vegas and Reno casinos in spring 1961, funded by **Manny Kimmel**, a New York businessman with $10,000 in seed capital. Over a series of trips, they confirmed the math at the tables. The result was published in **Thorp's 1962 book *Beat the Dealer*** (Random House) — a primary source for everything in this Method-in-Action section. The casinos' second-order response was rapid: they changed rules, introduced multiple decks and shoe shuffles, and trained dealers to spot counters. Within a few years, Thorp's specific edge in single-deck games had been engineered out of most casinos.\n\nBut the *framework* — combine a measurable edge with disciplined Kelly sizing across a repeated series of independent bets — translated. In 1969, Thorp co-founded **Princeton-Newport Partners**, one of the first quantitative hedge funds. The fund's strategy was convertible-bond arbitrage, statistical arbitrage, and other measurable-edge market trades, sized using Kelly principles (Thorp typically used a fractional Kelly given the higher estimation uncertainty in markets vs. casinos). **From 1969 to 1988, the fund returned 19.1% annually before fees with no down year** (Thorp, *A Man for All Markets*, 2017, ch. 14). The Princeton-Newport partnership ended for unrelated legal reasons in 1988.\n\nWalk through the EV-Kelly Sizing on Thorp's blackjack case:\n\n- **Repeatability (Step 1):** Tens of thousands of hands per casino trip. Strongly repeated. ✓\n- **Bet map (Step 2):** Per hand at a favorable count, win probability *p ≈ 0.505*, loss *q ≈ 0.495* (the +1% edge); payoff approximately 1:1 (with some variance from doubles, splits, blackjack 3:2).\n- **EV (Step 3):** ≈ +0.01 per unit staked. *Positive but small.*\n- **Edge gate (Step 4):** Positive. Proceed.\n- **Estimation uncertainty (Step 5):** In 1961, with a measurable card-counting system and verifiable math, this was as close to **casino-grade certainty** as bets get. Estimation uncertainty was unusually low (one of the few real-world cases where near-full Kelly is defensible).\n- **Kelly fraction (Step 6):** With *p = 0.505, q = 0.495, b = 1*: f\\* = (1 × 0.505 − 0.495) / 1 = **0.01 = 1% of bankroll per hand at a favorable count**. (Thorp varied the bet *up* to roughly this level at the highest-positive counts and dropped to the table minimum at negative counts.)\n- **Fractional Kelly (Step 7):** Because the edge math was unusually well-measured, Thorp used closer to full Kelly than is typical for market trades. In later markets work at Princeton-Newport, he scaled to fractional Kelly given larger estimation uncertainty.\n- **Stop trigger (Step 8):** Casino rule changes (multiple decks, frequent shuffles) — *the casinos' second-order response* — eventually made the edge structure obsolete in most venues. Thorp re-estimated repeatedly and pivoted out.\n\nThe inversion-style lesson Thorp himself emphasized later: **the most expensive mistake in EV-Kelly is using full Kelly with an estimated edge.** Half-Kelly costs roughly 25% of growth rate but cuts variance and ruin risk by far more (Thorp 1997 conference paper, §4).\n\n**Sources:** Kelly, J. L. **\"A New Interpretation of Information Rate.\"** *Bell System Technical Journal*, Vol. 35, No. 4 (July 1956), pp. 917–926 — the primary source for the formula and its derivation. https://www.princeton.edu/~wbialek/rome/refs/kelly_56.pdf . Thorp, Edward O. ***Beat the Dealer: A Winning Strategy for the Game of Twenty-One*** (Random House, 1962; 2nd ed. Vintage 1966) — primary source for the blackjack application. Thorp, Edward O. **\"The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market.\"** Paper for the 10th International Conference on Gambling and Risk Taking, Montreal, 1997 — the clearest concise summary by Thorp himself. https://www.eecs.harvard.edu/cs286r/courses/fall12/papers/Thorpe_KellyCriterion2007.pdf . Thorp, Edward O. ***A Man for All Markets*** (Random House, 2017) — autobiography; primary source for the Princeton-Newport years. Bernoulli, Daniel. **\"Exposition of a New Theory on the Measurement of Risk\"** (St. Petersburg, 1738; English tr. Louise Sommer, *Econometrica*, Vol. 22, No. 1, January 1954, pp. 23–36) — primary source for the utility-vs-EV distinction. Poundstone, William. ***Fortune's Formula: The Untold Story of the Scientific Betting System That Beat the Casinos and Wall Street*** (Hill & Wang, 2005) — the popular history of Kelly + Thorp, useful as a secondary reference.\n\nFile v1.0.1:skill-card.md\n\n## Description:
\nHelps agents evaluate repeated decisions by calculating expected value, Kelly sizing, fractional Kelly adjustments, and stop triggers for capital allocation, betting, ad spend, A/B testing, or similar repeatable choices.
\n\nThis skill is ready for commercial/non-commercial use.
\n\n## Publisher:
\n[deciqai](https://clawhub.ai/user/deciqai)
\n\n### License/Terms of Use:
\nMIT-0
\n\n\n## Use Case:
\nExternal users and agents use this skill to decide whether a repeated decision has positive expected value and to size exposure using Kelly or fractional Kelly. It is most relevant for measurable, repeatable allocation decisions where probabilities, payoffs, bankroll, uncertainty, and correlation can be stated explicitly.
\n\n### Deployment Geography for Use:
\nGlobal
\n\n## Known Risks and Mitigations:
\nRisk: The skill discusses financial and betting stake sizes and could be mistaken for personalized financial advice.
\nMitigation: Use outputs as educational calculation support only; verify inputs carefully and avoid treating the result as a personalized recommendation.
\nRisk: Incorrect probability, payoff, or bankroll assumptions can produce misleading Kelly fractions.
\nMitigation: Require explicit probabilities, payoffs, uncertainty estimates, and use conservative fractional Kelly sizing when inputs are estimated.
\nRisk: Correlated bets or stale edge estimates can make per-bet Kelly sizing too aggressive.
\nMitigation: Check correlation across positions and define stop-and-re-estimate triggers for drawdowns, outcome divergence, or regime changes.
\n\n\n## Reference(s):
\n- [ClawHub skill release page](https://clawhub.ai/deciqai/skills/expected-value-and-kelly)
\n- [deciqAI publisher profile](https://clawhub.ai/user/deciqai)
\n- [Sources - expected-value-and-kelly](references/sources.md)
\n- [Ed Thorp, Blackjack, and Princeton-Newport example](examples/ed-thorp-blackjack-and-princeton-newport-1961-1988.md)
\n- [Kelly, J. L., Jr., A New Interpretation of Information Rate](https://www.princeton.edu/~wbialek/rome/refs/kelly_56.pdf)
\n- [Thorp, The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market](https://www.eecs.harvard.edu/cs286r/courses/fall12/papers/Thorpe_KellyCriterion2007.pdf)
\n- [deciqAI](https://deciqai.com)
\n\n\n## Skill Output:
\n**Output Type(s):** [markdown, guidance, configuration]
\n**Output Format:** [Markdown EV-Kelly sizing worksheet with calculations, checks, and stop-trigger guidance]
\n**Output Parameters:** [1D]
\n**Other Properties Related to Output:** [Outputs are educational decision-support text and should not be treated as personalized financial advice.]
\n\n## Skill Version(s):
\n1.0.1 (source: server release evidence)
\n\n## Ethical Considerations:
\nUsers should evaluate whether this skill is appropriate for their environment, review any generated or modified files before relying on them, and apply their organization's safety, security, and compliance requirements before deployment.
\n\nArchive v1.0.0: 5 files, 10186 bytes\n\nFiles: examples/ed-thorp-blackjack-and-princeton-newport-1961-1988.md (5794b), references/sources.md (2146b), skill-card.md (2531b), SKILL.md (9123b), _meta.json (143b)\n\nFile v1.0.0:SKILL.md\n\n---\nname: expected-value-and-kelly\ndescription: \"Activate when: user asks 'how much should I bet/invest on this?', 'what's the expected value here?', 'Kelly criterion', 'optimal bet size', 'fractional Kelly', 'how big a position should I take?', or is allocating capital across repeated decisions (ad spend by segment, VC portfolio construction, position sizing, A/B test ramp).\n Do NOT activate when: the decision is one-shot and non-repeating (career change, marriage) — use regret-minimization instead; or when the user cannot estimate probabilities or payoffs even roughly.\"\n---\n\n# Expected Value and the Kelly Criterion\n\n## Overview\n\nTwo questions decide most repeated bets: **is this bet good?** (EV) and **how big?** (Kelly). Most professional ruin comes from positive-EV bets sized wrong. **EV = p · W − q · L.** If EV ≤ 0, do not bet. **Kelly f\\* = (bp − q) / b** maximizes long-term geometric growth (Kelly, Bell Labs, 1956). Full Kelly requires casino-grade certainty; default to half- or quarter-Kelly for estimated edges.\n\nNeighbors: [first-principles](../first-principles/SKILL.md) · [occams-razor](../occams-razor/SKILL.md) · [second-order-thinking](../second-order-thinking/SKILL.md) · [inversion](../inversion/SKILL.md) · [regret-minimization](../regret-minimization/SKILL.md) (for non-repeating life decisions).\n\n## When to Use\n\n- Decision **repeats many times** — capital allocation, position sizing, VC portfolio, ad spend, A/B test budget\n- *How big to bet* matters as much as *whether to bet*; you have a **measurable or estimable edge**\n- Someone says: \"expected value,\" \"EV,\" \"Kelly,\" \"optimal bet size,\" \"how much should we put on this?\"\n\n**When NOT to use:** one-shot life decisions → [regret-minimization](../regret-minimization/SKILL.md); negative-EV bets (don't bet); unestimable probabilities; correlated bets without portfolio adjustment.\n\n## Coaching Novices (Adaptive Front Door)\n\n**Engine mode:** user has a concrete repeated bet → run The Process directly. **Coach mode:** user is unfamiliar → guide step by step.\n\nIn Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.\n\n1. One-line what-it-is: EV tells you **whether the bet is worth taking**; Kelly tells you **what fraction of bankroll to stake** — sized to maximize long-term growth without ruin.\n2. Check fit against When to Use / When NOT to use. If one-shot life decision, redirect to [regret-minimization](../regret-minimization/SKILL.md). If EV is negative, say \"don't bet\" and stop.\n3. Elicit their real bet. Ask for a concrete repeated decision with measurable inputs. > **[WAIT — do not advance until user responds]**\n4. Walk The Process one step per turn: outcomes → probabilities → payoffs → EV → Kelly → fractional Kelly. > **[WAIT — do not advance until user responds]**\n5. Close by naming their sizing rule: \"bet f\\* × bankroll, use half-Kelly given estimation uncertainty\" — and the trigger that would change it. > **[WAIT — do not advance until user responds]**\n\n## The Process\n\nRun the **EV-Kelly Sizing** (EV → Kelly → fractional Kelly → stop trigger):\n\n1. **Confirm the decision is repeated.** If \"once,\" stop → use [regret-minimization](../regret-minimization/SKILL.md).\n2. **Map the bet.** Win prob *p*, loss prob *q = 1−p*, payoff on win *W*, loss *L*, odds *b = W/L*.\n3. **Compute EV.** EV = Σ(pᵢ · payoffᵢ). State per unit staked.\n4. **Edge gate.** EV > 0? If no, stop — do not bet.\n5. **Estimate input uncertainty.** Are *p* and *W* measured or estimated? Write 80% CI on edge.\n6. **Compute Kelly fraction.** f\\* = (bp − q) / b. For continuous: f\\* ≈ μ/σ².\n7. **Apply fractional Kelly.** Half-Kelly under modest uncertainty; quarter-Kelly under serious uncertainty.\n8. **Set stop trigger.** \"I will re-estimate if: (a) drawdown > X%, (b) outcomes diverge N σ over Y trials, (c) regime change invalidates edge model.\"\n\n### Output: EV-Kelly Sizing\n\n```markdown\n# EV-Kelly Sizing: \n## Repeatability: — if one-shot, STOP.\n## Bet map: p=<>, q=<>, W=<>, L=<>, b=<>\n## EV: p·W − q·L = — Edge: \n## Estimation uncertainty: ; 80% CI on edge: \n## Kelly fraction: f* = → practitioner: = \n## Stop trigger: \"I will re-estimate if .\"\n## Correlation check: bets independent? \n```\n*→ Method in Action: [Ed Thorp, Blackjack, and Princeton-Newport (1961 → 1988)](examples/ed-thorp-blackjack-and-princeton-newport-1961-1988.md)*\n## Sizing Packs\n\n| Domain | Fractional-Kelly | Stop trigger |\n|---|---|---|\n| Active equity | quarter-Kelly or less | drawdown > 2× expected annual vol |\n| Venture capital | portfolio-level quarter-Kelly | hit rate diverges from model by vintage |\n| Ad spend by segment | half- to full Kelly | ROAS falls >2σ over N conversions |\n| A/B test ramp | fractional Kelly on traffic % | regression in primary metric |\n\n## Applying It Well\n\n- **Positive EV is necessary but not sufficient.** A positive-EV bet sized too large still wrecks you.\n- **Default to fractional Kelly.** Full Kelly is for measured edges. Half- or quarter-Kelly for estimated edges — growth cost is small, protection is large.\n- **Correlation eats Kelly fast.** All positions long tech, all VC bets in one vintage — adjust portfolio Kelly down.\n- **Kelly is for bankroll, not for life.** Use [regret-minimization](../regret-minimization/SKILL.md) for career, marriage, time.\n- **Update continuously.** Static Kelly on a stale edge is how winning strategies ride into ruin.\n\n*→ Primary sources: [references/sources.md](references/sources.md)*\n## Common Rationalizations\n\n**[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.**\n\n| Fake move | Reality |\n|---|---|\n| [D] Applying Kelly to **one-shot** decisions | Kelly maximizes long-run geometric growth rate across many bets. For a one-time bet, EV is the right concern; for life decisions, use [regret-minimization](../regret-minimization/SKILL.md). |\n| [D] Treating **estimated *p*** as **known *p*** | Kelly is brutal when probability estimates are wrong. Use fractional Kelly to compensate for estimated inputs. |\n| [D] Using **full Kelly with subjective probabilities** | Full Kelly is for casino-grade certainty. Half-Kelly costs ~25% of growth and cuts ruin risk by far more (Thorp 1997). |\n| [D] Confusing **positive EV** with **\"should bet\"** | EV ignores variance, bankroll, path-dependency. A +1% EV bet that ruins you 1% of the time is not equivalent to one that ruins you 0%. |\n| [D] **Ignoring correlation** across bets | Kelly assumes independent bets. Correlated bets (same sector, same vintage) require a lower portfolio-level Kelly. |\n| [D] Using Kelly on **negative-EV** bets | Kelly returns zero or negative fraction when EV ≤ 0. No sizing rescues a -EV bet. Don't bet. |\n| [D] Treating EV as the **only number that matters** | Bernoulli 1738: utility of money is non-linear. A +$1000 EV bet that risks your rent ≠ one that risks your rounding error. |\n| [D] Forgetting Kelly's **brutal variance** | Even correct full-Kelly expects 50%+ drawdowns. Most professionals use half/quarter-Kelly to survive psychologically. |\n| [D] Applying Kelly to **non-financial \"bets\"** | Kelly assumes a compoundable bankroll. Relationships, careers, time, attention don't compound across independent trials. |\n| [D] Computing Kelly **once**, ignoring updates | A static fraction on a now-stale edge is the textbook path to ruin. Pre-commit to a re-estimation trigger. |\n| *→ Add [O] entries here after each real use — paste the actual failure pattern* | *What went wrong and why* |\n\n## Red Flags\n\n- Kelly applied to a decision that happens only once\n- Full Kelly used with subjective probabilities and no calibration check\n- EV is negative but a positive fraction is being computed\n- No stop-and-re-estimate trigger is named\n- Bets are obviously correlated but Kelly is computed per-bet without portfolio adjustment\n- The \"bankroll\" is not liquid or fungible (career time, relationship capital)\n- A historical drawdown wiped out the strategy; same fraction still applied without re-examining the edge\n\n## Verification\n\n- [ ] Decision confirmed *repeated*, not one-shot\n- [ ] Outcomes, probabilities, and payoffs stated explicitly in correct units\n- [ ] EV computed per unit staked; edge gate (positive/zero/negative) named\n- [ ] Estimation uncertainty named (measured/estimated/mixed); 80% CI on edge given\n- [ ] Kelly fraction computed with correct formula (binary vs. continuous)\n- [ ] Fractional-Kelly adjustment applied with explicit justification\n- [ ] Stop-and-re-estimate trigger named with concrete conditions\n- [ ] Correlation across bets considered; per-bet Kelly adjusted if needed\n\n---\n*Part of **deciqAI Knowledge Skills** — open-source thinking skills that make rigor executable for AI agents. Built by deciqAI · https://deciqai.com · Contributions welcome — see the template at the repo root.*\n\nFile v1.0.0:_meta.json\n\n{\n \"ownerId\": \"kn754b8sk22s8c6gjxt02bftbn88q7ye\",\n \"slug\": \"expected-value-and-kelly\",\n \"version\": \"1.0.0\",\n \"publishedAt\": 1782631095841\n}\n\nFile v1.0.0:references/sources.md\n\n# Sources — expected-value-and-kelly\n\n> *Primary sources for the [expected-value-and-kelly](../SKILL.md) skill.*\n\n- **Kelly, J. L., Jr.** \"A New Interpretation of Information Rate.\" *Bell System Technical Journal*, Vol. 35, No. 4 (July 1956), pp. 917–926. **The primary source for the Kelly formula and its proof.** https://www.princeton.edu/~wbialek/rome/refs/kelly_56.pdf\n- **Thorp, Edward O.** *Beat the Dealer: A Winning Strategy for the Game of Twenty-One*. Random House, 1962 (2nd ed. Vintage, 1966). **Primary source** for the first practical application of Kelly to a positive-edge real-world bet (blackjack card-counting).\n- **Thorp, Edward O.** \"The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market.\" Paper for the 10th International Conference on Gambling and Risk Taking, Montreal, 1997. The clearest concise summary by Thorp himself; covers the fractional-Kelly trade-off. https://www.eecs.harvard.edu/cs286r/courses/fall12/papers/Thorpe_KellyCriterion2007.pdf\n- **Thorp, Edward O.** *A Man for All Markets: From Las Vegas to Wall Street, How I Beat the Dealer and the Market*. Random House, 2017. Autobiography; primary source on the Princeton-Newport years and Thorp's evolution from blackjack to markets.\n- **Bernoulli, Daniel.** \"Exposition of a New Theory on the Measurement of Risk.\" Commentarii Academiae Scientiarum Imperialis Petropolitanae, Tomus V (St. Petersburg, 1738). English translation by Louise Sommer in *Econometrica*, Vol. 22, No. 1 (January 1954), pp. 23–36. **Primary source** for the utility-versus-EV distinction.\n- **Poundstone, William.** *Fortune's Formula: The Untold Story of the Scientific Betting System That Beat the Casinos and Wall Street*. Hill & Wang, 2005. The popular history; useful as a secondary reference and for the Bell Labs / Shannon / Kelly intellectual lineage.\n- The popular framing \"the house always wins\" is **not** cited here as a source — by this skill's own rule, an aphorism is not evidence. The Kelly criterion is precisely the formula for *the conditions under which the house does not always win* — when the player has a measurable positive edge.\n\nFile v1.0.0:examples/ed-thorp-blackjack-and-princeton-newport-1961-1988.md\n\n# Method in Action: Ed Thorp, Blackjack, and Princeton-Newport (1961 → 1988)\n\n> *Example for the [expected-value-and-kelly](../SKILL.md) skill.*\n\nA worked example. Not a casino fable — primary-source documented, including the Bell Labs origin.\n\nIn 1960, **Ed Thorp**, a young mathematics professor at MIT, was reading Baldwin et al.'s 1956 paper on basic-strategy blackjack and realized something the casino-gambling literature had missed: because cards in blackjack are drawn *without replacement* from a finite deck, the composition of the remaining deck shifts the player's edge from hand to hand. A player who **tracks** the cards already played (a \"counter\") sees a small but real positive edge on certain deck compositions — typically when the remaining deck is rich in tens and aces.\n\nThorp's contribution had two parts. First, he proved the math: a counting system (his \"Ten-Count,\" later refined to the \"High-Low\" by Harvey Dubner) yields an average edge of roughly **+1% over the house** when applied disciplinedly. Second — and this is what most blackjack popularizers miss — he applied **Kelly sizing** to the bet: vary the wager up and down with the count, betting more when the count favored the player and the minimum when it didn't.\n\nTo prove it in the field, Thorp went to Las Vegas and Reno casinos in spring 1961, funded by **Manny Kimmel**, a New York businessman with $10,000 in seed capital. Over a series of trips, they confirmed the math at the tables. The result was published in **Thorp's 1962 book *Beat the Dealer*** (Random House) — a primary source for everything in this Method-in-Action section. The casinos' second-order response was rapid: they changed rules, introduced multiple decks and shoe shuffles, and trained dealers to spot counters. Within a few years, Thorp's specific edge in single-deck games had been engineered out of most casinos.\n\nBut the *framework* — combine a measurable edge with disciplined Kelly sizing across a repeated series of independent bets — translated. In 1969, Thorp co-founded **Princeton-Newport Partners**, one of the first quantitative hedge funds. The fund's strategy was convertible-bond arbitrage, statistical arbitrage, and other measurable-edge market trades, sized using Kelly principles (Thorp typically used a fractional Kelly given the higher estimation uncertainty in markets vs. casinos). **From 1969 to 1988, the fund returned 19.1% annually before fees with no down year** (Thorp, *A Man for All Markets*, 2017, ch. 14). The Princeton-Newport partnership ended for unrelated legal reasons in 1988.\n\nWalk through the EV-Kelly Sizing on Thorp's blackjack case:\n\n- **Repeatability (Step 1):** Tens of thousands of hands per casino trip. Strongly repeated. ✓\n- **Bet map (Step 2):** Per hand at a favorable count, win probability *p ≈ 0.505*, loss *q ≈ 0.495* (the +1% edge); payoff approximately 1:1 (with some variance from doubles, splits, blackjack 3:2).\n- **EV (Step 3):** ≈ +0.01 per unit staked. *Positive but small.*\n- **Edge gate (Step 4):** Positive. Proceed.\n- **Estimation uncertainty (Step 5):** In 1961, with a measurable card-counting system and verifiable math, this was as close to **casino-grade certainty** as bets get. Estimation uncertainty was unusually low (one of the few real-world cases where near-full Kelly is defensible).\n- **Kelly fraction (Step 6):** With *p = 0.505, q = 0.495, b = 1*: f\\* = (1 × 0.505 − 0.495) / 1 = **0.01 = 1% of bankroll per hand at a favorable count**. (Thorp varied the bet *up* to roughly this level at the highest-positive counts and dropped to the table minimum at negative counts.)\n- **Fractional Kelly (Step 7):** Because the edge math was unusually well-measured, Thorp used closer to full Kelly than is typical for market trades. In later markets work at Princeton-Newport, he scaled to fractional Kelly given larger estimation uncertainty.\n- **Stop trigger (Step 8):** Casino rule changes (multiple decks, frequent shuffles) — *the casinos' second-order response* — eventually made the edge structure obsolete in most venues. Thorp re-estimated repeatedly and pivoted out.\n\nThe inversion-style lesson Thorp himself emphasized later: **the most expensive mistake in EV-Kelly is using full Kelly with an estimated edge.** Half-Kelly costs roughly 25% of growth rate but cuts variance and ruin risk by far more (Thorp 1997 conference paper, §4).\n\n**Sources:** Kelly, J. L. **\"A New Interpretation of Information Rate.\"** *Bell System Technical Journal*, Vol. 35, No. 4 (July 1956), pp. 917–926 — the primary source for the formula and its derivation. https://www.princeton.edu/~wbialek/rome/refs/kelly_56.pdf . Thorp, Edward O. ***Beat the Dealer: A Winning Strategy for the Game of Twenty-One*** (Random House, 1962; 2nd ed. Vintage 1966) — primary source for the blackjack application. Thorp, Edward O. **\"The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market.\"** Paper for the 10th International Conference on Gambling and Risk Taking, Montreal, 1997 — the clearest concise summary by Thorp himself. https://www.eecs.harvard.edu/cs286r/courses/fall12/papers/Thorpe_KellyCriterion2007.pdf . Thorp, Edward O. ***A Man for All Markets*** (Random House, 2017) — autobiography; primary source for the Princeton-Newport years. Bernoulli, Daniel. **\"Exposition of a New Theory on the Measurement of Risk\"** (St. Petersburg, 1738; English tr. Louise Sommer, *Econometrica*, Vol. 22, No. 1, January 1954, pp. 23–36) — primary source for the utility-vs-EV distinction. Poundstone, William. ***Fortune's Formula: The Untold Story of the Scientific Betting System That Beat the Casinos and Wall Street*** (Hill & Wang, 2005) — the popular history of Kelly + Thorp, useful as a secondary reference.\n\nFile v1.0.0:skill-card.md\n\n## Description:
\nHelps an agent evaluate repeated sizing decisions by mapping probabilities and payoffs, computing expected value, applying the Kelly criterion or fractional Kelly, and naming stop triggers.
\n\nThis skill is ready for commercial/non-commercial use.
\n\n## Publisher:
\n[deciqai](https://clawhub.ai/user/deciqai)
\n\n### License/Terms of Use:
\nMIT-0
\n\n\n## Use Case:
\nExternal users and developers use this skill to guide agents through repeated capital-allocation, position-sizing, ad-spend, A/B-test ramp, or similar decisions where probabilities and payoffs can be estimated. The skill produces an EV-Kelly sizing analysis with uncertainty, fractional-Kelly, correlation, and stop-trigger checks.
\n\n### Deployment Geography for Use:
\nGlobal
\n\n## Known Risks and Mitigations:
\nRisk: Users may treat educational EV-Kelly output as professional financial advice or apply it to one-shot decisions.
\nMitigation: Present outputs as educational analysis, require a repeatable decision with estimable probabilities and payoffs, and redirect one-shot or unestimable decisions away from Kelly sizing.
\nRisk: Incorrect probabilities, payoffs, or correlation assumptions can lead to oversized recommendations.
\nMitigation: Require explicit input uncertainty, fractional-Kelly sizing, correlation checks, and concrete stop-and-reestimate triggers.
\n\n\n## Reference(s):
\n- [ClawHub skill page](https://clawhub.ai/deciqai/skills/expected-value-and-kelly)
\n- [Sources - expected-value-and-kelly](references/sources.md)
\n- [A New Interpretation of Information Rate](https://www.princeton.edu/~wbialek/rome/refs/kelly_56.pdf)
\n- [The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market](https://www.eecs.harvard.edu/cs286r/courses/fall12/papers/Thorpe_KellyCriterion2007.pdf)
\n\n\n## Skill Output:
\n**Output Type(s):** [text, markdown, guidance]
\n**Output Format:** [Markdown EV-Kelly sizing worksheet with calculations and decision guidance]
\n**Output Parameters:** [1D]
\n**Other Properties Related to Output:** [Educational calculation guidance; no executable behavior.]
\n\n## Skill Version(s):
\n1.0.0 (source: server release evidence)
\n\n## Ethical Considerations:
\nUsers should evaluate whether this skill is appropriate for their environment, review any generated or modified files before relying on them, and apply their organization's safety, security, and compliance requirements before deployment.
","readmeExcerpt":"Skill: Expected Value and the Kelly Criterion Owner: deciqai Summary: Activate when: user asks 'how much should I bet/invest on this?', 'what's the expected value here?', 'Kelly criterion', 'optimal bet size', 'fractional Kelly... Tags: latest:1.0.4 Version history: v1.0.4 | 2026-07-16T17:59:26.940Z | user Description tail link + agents machine-readable metadata line (deciqai.com/s/expected-value-and-kelly.json) v1.0","codeSnippets":[],"executableExamples":[{"language":"markdown","snippet":"# EV-Kelly Sizing: \n## Repeatability: — if one-shot, STOP.\n## Bet map: p=<>, q=<>, W=<>, L=<>, b=<>\n## EV: p·W − q·L = — Edge: \n## Estimation uncertainty: ; 80% CI on edge: \n## Kelly fraction: f* = → practitioner: = \n## Stop trigger: \"I will re-estimate if .\"\n## Correlation check: bets independent? "},{"language":"markdown","snippet":"# EV-Kelly Sizing: \n## Repeatability: — if one-shot, STOP.\n## Bet map: p=<>, q=<>, W=<>, L=<>, b=<>\n## EV: p·W − q·L = — Edge: \n## Estimation uncertainty: ; 80% CI on edge: \n## Kelly fraction: f* = → practitioner: = \n## Stop trigger: \"I will re-estimate if .\"\n## Correlation check: bets independent? "},{"language":"markdown","snippet":"# EV-Kelly Sizing: \n## Repeatability: — if one-shot, STOP.\n## Bet map: p=<>, q=<>, W=<>, L=<>, b=<>\n## EV: p·W − q·L = — Edge: \n## Estimation uncertainty: ; 80% CI on edge: \n## Kelly fraction: f* = → practitioner: = \n## Stop trigger: \"I will re-estimate if .\"\n## Correlation check: bets independent? "},{"language":"markdown","snippet":"# EV-Kelly Sizing: \n## Repeatability: — if one-shot, STOP.\n## Bet map: p=<>, q=<>, W=<>, L=<>, b=<>\n## EV: p·W − q·L = — Edge: \n## Estimation uncertainty: ; 80% CI on edge: \n## Kelly fraction: f* = → practitioner: = \n## Stop trigger: \"I will re-estimate if .\"\n## Correlation check: bets independent? "},{"language":"markdown","snippet":"# EV-Kelly Sizing: \n## Repeatability: — if one-shot, STOP.\n## Bet map: p=<>, q=<>, W=<>, L=<>, b=<>\n## EV: p·W − q·L = — Edge: \n## Estimation uncertainty: ; 80% CI on edge: \n## Kelly fraction: f* = → practitioner: = \n## Stop trigger: \"I will re-estimate if .\"\n## Correlation check: bets independent? "}],"parameters":null,"dependencies":[],"permissions":[],"extractedFiles":[{"path":"SKILL.md","content":"---\nname: expected-value-and-kelly\ndescription: \"Activate when: user asks 'how much should I bet/invest on this?', 'what's the expected value here?', 'Kelly criterion', 'optimal bet size', 'fractional Kelly', 'how big a position should I take?', or is allocating capital across repeated decisions (ad spend by segment, VC portfolio construction, position sizing, A/B test ramp).\n Do NOT activate when: the decision is one-shot and non-repeating (career change, marriage) — use regret-minimization instead; or when the user cannot estimate probabilities or payoffs even roughly. More: deciqai.com/c/expected-value-and-kelly\"\n---\n\n# Expected Value and the Kelly Criterion\n\n## Overview\n\nTwo questions decide most repeated bets: **is this bet good?** (EV) and **how big?** (Kelly). Most professional ruin comes from positive-EV bets sized wrong. **EV = p · W − q · L.** If EV ≤ 0, do not bet. **Kelly f\\* = (bp − q) / b** maximizes long-term geometric growth (Kelly, Bell Labs, 1956). Full Kelly requires casino-grade certainty; default to half- or quarter-Kelly for estimated edges.\n\nNeighbors: first-principles · occams-razor · second-order-thinking · inversion · regret-minimization (for non-repeating life decisions).\n\n## When to Use\n\n- Decision **repeats many times** — capital allocation, position sizing, VC portfolio, ad spend, A/B test budget\n- *How big to bet* matters as much as *whether to bet*; you have a **measurable or estimable edge**\n- Someone says: \"expected value,\" \"EV,\" \"Kelly,\" \"optimal bet size,\" \"how much should we put on this?\"\n- Sizing bets in a boom with power-law payoffs and possible ruin — **how much to allocate to AI startups / GPU-compute capex / AI-exposed equities** given frothy AI valuations, uncertain AI adoption, and correlated bets\n\n**When NOT to use:** one-shot life decisions → regret-minimization; negative-EV bets (don't bet); unestimable probabilities; correlated bets without portfolio adjustment.\n\n## Coaching Novices (Adaptive Front Door)\n\n**Engine mode:** user has a concrete repeated bet → run The Process directly. **Coach mode:** user is unfamiliar → guide step by step.\n\nIn Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.\n\n1. One-line what-it-is: EV tells you **whether the bet is worth taking**; Kelly tells you **what fraction of bankroll to stake** — sized to maximize long-term growth without ruin.\n2. Check fit against When to Use / When NOT to use. If one-shot life decision, redirect to regret-minimization. If EV is negative, say \"don't bet\" and stop.\n3. Elicit their real bet. Ask for a concrete repeated decision with measurable inputs. > **[WAIT — do not advance until user responds]**\n4. Walk The Process one step per turn: outcomes → probabilities → payoffs → EV → Kelly → fractional Kelly. > **[WAIT — do not advance until user responds]**\n5. Close by naming their sizing rule: \"bet f\\* × bankroll, use half-Kelly given estimation uncertainty\" — and the trigger that "},{"path":"_meta.json","content":"{\n \"ownerId\": \"kn754b8sk22s8c6gjxt02bftbn88q7ye\",\n \"slug\": \"expected-value-and-kelly\",\n \"version\": \"1.0.4\",\n \"publishedAt\": 1784224766940\n}"},{"path":"references/sources.md","content":"# Sources — expected-value-and-kelly\n\n> *Primary sources for the [expected-value-and-kelly](../SKILL.md) skill.*\n\n- **Kelly, J. L., Jr.** \"A New Interpretation of Information Rate.\" *Bell System Technical Journal*, Vol. 35, No. 4 (July 1956), pp. 917–926. **The primary source for the Kelly formula and its proof.** https://www.princeton.edu/~wbialek/rome/refs/kelly_56.pdf\n- **Thorp, Edward O.** *Beat the Dealer: A Winning Strategy for the Game of Twenty-One*. Random House, 1962 (2nd ed. Vintage, 1966). **Primary source** for the first practical application of Kelly to a positive-edge real-world bet (blackjack card-counting).\n- **Thorp, Edward O.** \"The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market.\" Paper for the 10th International Conference on Gambling and Risk Taking, Montreal, 1997. The clearest concise summary by Thorp himself; covers the fractional-Kelly trade-off. https://www.eecs.harvard.edu/cs286r/courses/fall12/papers/Thorpe_KellyCriterion2007.pdf\n- **Thorp, Edward O.** *A Man for All Markets: From Las Vegas to Wall Street, How I Beat the Dealer and the Market*. Random House, 2017. Autobiography; primary source on the Princeton-Newport years and Thorp's evolution from blackjack to markets.\n- **Bernoulli, Daniel.** \"Exposition of a New Theory on the Measurement of Risk.\" Commentarii Academiae Scientiarum Imperialis Petropolitanae, Tomus V (St. Petersburg, 1738). English translation by Louise Sommer in *Econometrica*, Vol. 22, No. 1 (January 1954), pp. 23–36. **Primary source** for the utility-versus-EV distinction.\n- **Benter, William.** \"Computer Based Horse Race Handicapping and Wagering Systems: A Report.\" In *Efficiency of Racetrack Betting Markets*, edited by Donald B. Hausch, Victor S. Y. Lo, and William T. Ziemba, pp. 183–198. Academic Press, 1994 (reprinted World Scientific, 2008). **Primary source** for combining a fitted probability model with public odds and sizing pari-mutuel bets by fractional Kelly under estimated inputs and market impact.\n- **Poundstone, William.** *Fortune's Formula: The Untold Story of the Scientific Betting System That Beat the Casinos and Wall Street*. Hill & Wang, 2005. The popular history; useful as a secondary reference and for the Bell Labs / Shannon / Kelly intellectual lineage.\n- **U.S. Securities and Exchange Commission filings and quarterly earnings materials of Microsoft, Alphabet, Amazon, Meta, and Nvidia (2023–2025).** Durable **primary sources** for AI-driven data-center capital-expenditure guidance and Nvidia's data-center-segment revenue growth during the 2023–2026 AI investment boom — used in preference to fast-moving press figures, which should be verified against these filings. https://www.sec.gov/cgi-bin/browse-edgar\n- **Contemporaneous business-press coverage of frontier-AI-lab financing rounds and valuations (e.g., The Wall Street Journal, Financial Times, The Information, 2023–2026).** **Secondary** reference for reported startup valuations (OpenAI, Anthropic, a"},{"path":"examples/bill-benter-hong-kong-horse-racing-1985-2001.md","content":"# Method in Action: Bill Benter and Computer-Model Horse Betting in Hong Kong (1985 → 2001)\n\n> *Example for the [expected-value-and-kelly](../SKILL.md) skill.*\n\nA worked example in a domain far from a card table: **pari-mutuel horse racing**, where the odds are set by the crowd's own money and your bet moves the payout against you. Primary-source documented in Benter's own 1994 technical paper.\n\nIn the mid-1980s **William \"Bill\" Benter** — who had learned card-counting from Thorp's *Beat the Dealer* and been barred from casinos — turned to the Hong Kong Jockey Club, then the largest pari-mutuel racing pool in the world. His insight was statistical, not tipster's intuition: build a **multinomial logit model** that estimates each horse's true win probability from fundamental factors (past times, weight, jockey, barrier, going), then combine that estimate with the **public's implied probability** encoded in the tote odds. The crowd is a strong predictor; a model that ignores it loses, and a model that only mimics it has no edge. Benter's edge came from the *residual* — the systematic gap between his model's probabilities and the crowd's.\n\nThat residual defined the bet. On any race, a horse is a positive-EV wager only when Benter's estimated win probability *p* exceeds the break-even probability implied by its tote odds after the Jockey Club's ~17–19% takeout. Most horses in most races failed that gate. The discipline was to bet only the small fraction that cleared it — and to size each qualifying bet by Kelly.\n\nPari-mutuel racing punishes overbetting twice. First, Kelly's ordinary lesson: a positive-EV bet staked too large courts ruin through variance. Second, a mechanical trap absent from casino play — **your own money enters the pool and lowers the odds you receive.** A full-Kelly stake computed against the pre-bet odds is, after your bet lands, an *overbet* against the true post-bet payout. Benter's paper is explicit that they used **fractional Kelly**, both to absorb estimation error in the probability model and to blunt this self-impact on the tote. Run over hundreds of thousands of bets across many seasons, the syndicate is widely reported to have won on the order of a billion dollars — the integral effect of a small per-race edge, correctly sized and relentlessly repeated.\n\nWalk the EV-Kelly Sizing through Benter's case:\n\n- **Repeatability (Step 1):** Hundreds of races per season, many seasons, thousands of qualifying bets. Emphatically repeated. ✓\n- **Bet map (Step 2):** For a candidate horse, model win probability *p*; payoff *W* set by the tote (dividend minus stake) after ~17–19% takeout; loss *L* = the stake.\n- **EV (Step 3):** EV = p·W − q·L per unit staked. Computed per horse, per race.\n- **Edge gate (Step 4):** Bet only when *p* exceeds the takeout-adjusted break-even implied by the odds. The overwhelming majority of horses fail this gate and are not bet.\n- **Estimation uncertainty (Step 5):** High and irreducible — *p* comes from a f"},{"path":"examples/ed-thorp-blackjack-and-princeton-newport-1961-1988.md","content":"# Method in Action: Ed Thorp, Blackjack, and Princeton-Newport (1961 → 1988)\n\n> *Example for the [expected-value-and-kelly](../SKILL.md) skill.*\n\nA worked example. Not a casino fable — primary-source documented, including the Bell Labs origin.\n\nIn 1960, **Ed Thorp**, a young mathematics professor at MIT, was reading Baldwin et al.'s 1956 paper on basic-strategy blackjack and realized something the casino-gambling literature had missed: because cards in blackjack are drawn *without replacement* from a finite deck, the composition of the remaining deck shifts the player's edge from hand to hand. A player who **tracks** the cards already played (a \"counter\") sees a small but real positive edge on certain deck compositions — typically when the remaining deck is rich in tens and aces.\n\nThorp's contribution had two parts. First, he proved the math: a counting system (his \"Ten-Count,\" later refined to the \"High-Low\" by Harvey Dubner) yields an average edge of roughly **+1% over the house** when applied disciplinedly. Second — and this is what most blackjack popularizers miss — he applied **Kelly sizing** to the bet: vary the wager up and down with the count, betting more when the count favored the player and the minimum when it didn't.\n\nTo prove it in the field, Thorp went to Las Vegas and Reno casinos in spring 1961, funded by **Manny Kimmel**, a New York businessman with $10,000 in seed capital. Over a series of trips, they confirmed the math at the tables. The result was published in **Thorp's 1962 book *Beat the Dealer*** (Random House) — a primary source for everything in this Method-in-Action section. The casinos' second-order response was rapid: they changed rules, introduced multiple decks and shoe shuffles, and trained dealers to spot counters. Within a few years, Thorp's specific edge in single-deck games had been engineered out of most casinos.\n\nBut the *framework* — combine a measurable edge with disciplined Kelly sizing across a repeated series of independent bets — translated. In 1969, Thorp co-founded **Princeton-Newport Partners**, one of the first quantitative hedge funds. The fund's strategy was convertible-bond arbitrage, statistical arbitrage, and other measurable-edge market trades, sized using Kelly principles (Thorp typically used a fractional Kelly given the higher estimation uncertainty in markets vs. casinos). **From 1969 to 1988, the fund returned 19.1% annually before fees with no down year** (Thorp, *A Man for All Markets*, 2017, ch. 14). The Princeton-Newport partnership ended for unrelated legal reasons in 1988.\n\nWalk through the EV-Kelly Sizing on Thorp's blackjack case:\n\n- **Repeatability (Step 1):** Tens of thousands of hands per casino trip. Strongly repeated. ✓\n- **Bet map (Step 2):** Per hand at a favorable count, win probability *p ≈ 0.505*, loss *q ≈ 0.495* (the +1% edge); payoff approximately 1:1 (with some variance from doubles, splits, blackjack 3:2).\n- **EV (Step 3):** ≈ +0.01 per unit staked. *Positive but sma"}],"languages":[],"docsSourceLabel":"CLAWHUB","editorialOverview":"Activate when: user asks 'how much should I bet/invest on this?', 'what's the expected value here?', 'Kelly criterion', 'optimal bet size', 'fractional Kelly... Skill: Expected Value and the Kelly Criterion Owner: deciqai Summary: Activate when: user asks 'how much should I bet/invest on this?', 'what's the expected value here?', 'Kelly criterion', 'optimal bet size', 'fractional Kelly... Tags: latest:1.0.4 Version history: v1.0.4 | 2026-07-16T17:59:26.940Z | user Description tail link + agents machine-readable metadata line (deciqai.com/s/expected-value-and-kelly.json) v1.0","editorialQuality":{"score":100,"threshold":65,"status":"ready","wordCount":2247,"uniquenessScore":48,"reasons":[]}},"media":{"evidence":{"source":"no-media","verified":false,"confidence":"low","updatedAt":"2026-10-12T01:50:01.517Z","emptyReason":"No screenshots, media assets, or demo links are available."},"primaryImageUrl":null,"mediaAssetCount":0,"assets":[],"demoUrl":null},"ownerResources":{"evidence":{"source":"unclaimed","verified":false,"confidence":"low","updatedAt":"2026-10-12T01:50:01.517Z","emptyReason":"This page has not been claimed by the agent owner."},"hasCustomPage":false,"customPageUpdatedAt":null,"customLinks":[],"structuredLinks":{"docsUrl":null,"demoUrl":null,"supportUrl":null,"pricingUrl":null,"statusUrl":null},"customPage":null},"relatedAgents":{"evidence":{"source":"protocol-neighbors","verified":false,"confidence":"medium","updatedAt":"2026-10-12T13:19:52.492Z","emptyReason":null},"items":[{"id":"8ebccd8e-3863-4187-8355-c3f14e1f9edf","entityType":"agent","canonicalPath":"/agent/iofficeai-aionui","slug":"iofficeai-aionui","name":"AionUi","description":"Free, local, open-source 24/7 Cowork app and OpenClaw for Gemini CLI, Claude Code, Codex, OpenCode, Qwen Code, Goose CLI, Auggie, and more | 🌟 Star if you like it!","url":"https://github.com/iOfficeAI/AionUi","homepage":"https://www.aionui.com","source":"GITHUB_REPOS","protocols":["MCP","OPENCLAW"],"capabilities":[],"safetyScore":100,"overallRank":70,"updatedAt":"2026-10-09T19:11:12.944Z","createdAt":"2026-02-25T03:38:16.584Z","downloads":null},{"id":"b917f68a-ebff-438e-84f8-3f4b2494c0bc","entityType":"agent","canonicalPath":"/agent/activepieces-activepieces","slug":"activepieces-activepieces","name":"activepieces","description":"AI Agents & MCPs & AI Workflow Automation • (~400 MCP servers for AI agents) • AI Automation / AI Agent with MCPs • AI Workflows & AI Agents • MCPs for AI Agents","url":"https://github.com/activepieces/activepieces","homepage":"https://www.activepieces.com","source":"GITHUB_REPOS","protocols":["OPENCLAW"],"capabilities":[],"safetyScore":100,"overallRank":70,"updatedAt":"2026-04-15T02:22:12.426Z","createdAt":"2026-02-25T03:38:12.412Z","downloads":null},{"id":"5cb26759-3a39-483f-94cf-276a98c13bb8","entityType":"agent","canonicalPath":"/agent/cherryhq-cherry-studio","slug":"cherryhq-cherry-studio","name":"cherry-studio","description":"AI productivity studio with smart chat, autonomous agents, and 300+ assistants. Unified access to frontier LLMs","url":"https://github.com/CherryHQ/cherry-studio","homepage":"https://cherry-ai.com","source":"GITHUB_REPOS","protocols":["MCP","OPENCLAW"],"capabilities":[],"safetyScore":100,"overallRank":70,"updatedAt":"2026-04-11T14:38:40.986Z","createdAt":"2026-02-25T03:38:19.379Z","downloads":null},{"id":"6f6582d0-5d76-4f0f-b81d-86520247950b","entityType":"agent","canonicalPath":"/agent/copilotkit-copilotkit","slug":"copilotkit-copilotkit","name":"CopilotKit","description":"The Frontend for Agents & Generative UI. React + Angular","url":"https://github.com/CopilotKit/CopilotKit","homepage":"https://docs.copilotkit.ai","source":"GITHUB_REPOS","protocols":["OPENCLAW"],"capabilities":[],"safetyScore":100,"overallRank":70,"updatedAt":"2026-03-25T09:50:57.846Z","createdAt":"2026-02-25T03:39:14.617Z","downloads":null}],"links":{"hub":"/agent","source":"/agent/source/clawhub","protocols":[{"label":"OpenClaw","href":"/agent/protocol/openclew"}]}}}