Keel meets Kelly: what ten money-management writeups taught an engine that refuses to bet

[eaitbrahim]

Based on Will McGinnis's ten-post money-management series (April 2025 - March 2026, accompanying the open-source keeks betting library) as analyzed in keel's source review. This is the sizing half of that review, written up for readers deciding how a trading engine should treat bet sizing when it cannot yet prove an edge.

The strange headline

We read the series looking for sizing ideas to steal. The most valuable thing it gave us was a confirmation that doing nothing is the mathematically correct bet.

keel's own measurement — restated honestly under a production-faithful engine — is that no shipped rule family is net-positive at the taker fee actually paid on the venue. Run any formula in this literature on those numbers and the output is the same: the Kelly fraction is zero or negative, so the optimal bet is no bet. Fortune's Formula states the degenerate case plainly — on a fair coin, the Kelly bet is zero. A trading engine whose live path refuses to promote an unproven rule is not being timid. It is being Kelly-consistent under estimation uncertainty.

That framing organize everything else the series taught us.

The series in ten lines

1. Kelly criterion — f* = (p·b − q)/b: maximize log-wealth growth; bet more with bigger edge, nothing without one. Needs probability estimates you can trust.
2. keeks 0.3.0 / Merton share — f* = μ/(γσ²): Kelly generalized by a risk-aversion parameter γ; γ = 2 keeps ~84% of returns at ~61% less volatility.
3. Fractional Kelly — bet f × f*; half Kelly ≈ 75% of growth at half the variance; the standard concession to estimation error.
4. Drawdown-adjusted Kelly — (1 − d/D) × Kelly: de-risk linearly toward your maximum acceptable drawdown D.
5. Optimal-f — Vince's TWR = Π(1 + f·Rᵢ) maximized over actual trade history; the post itself warns it's unreliable under ~30-50 trades and needs a 50-70% safety factor.
6. Fixed fraction — bet = f × bankroll: simple, ruin-proof, edge-blind; optimal fixed fractions sit below Kelly; start at 1-2%.
7. CPPI — bet = m × (bankroll − floor): protect a floor, scale exposure to the cushion; watch gap risk.
8. Dynamic bankroll management — scale size by win/loss streaks within bounds; the post's own cons: overreaction to variance, overconfidence on lucky streaks.
9. Naive strategy — flat betting as the baseline everything must beat.
10. Strategy comparison — the honest mapping of risk tolerance to strategy, with the caveat that its own simulations are single-seed and cost-free.

One translation warning before comparing

The series sizes a stake — a fraction of bankroll wagered on a binary outcome. keel sizes a risk — a fraction of equity lost if the stop holds: qty = equity × risk_pct / |entry − stop|, with notional bounded separately by hard caps. A keel position's notional can be a large multiple of its risk when the stop is tight. Kelly comparisons against keel must be made on risk fractions, never notional. (The same trap runs the other way: a tighter stop means a larger position, not a safer one.)

What keel already has, in the series' vocabulary

Reading the ten posts with keel's source open, most of the machinery is already present — usually in a stronger form:

• Fixed fraction: keel's default posture (risk_pct = 1%), risk-defined rather than stake-defined. The series' "start at 1-2%, optimal sits below Kelly" matches keel's setting exactly.
• Drawdown control: keel's rail 11 halts trading at a 20% total / 8% weekly drawdown ceiling — the binary, guaranteed version of drawdown-adjusted sizing (a (1 − d/D) throttle is queued as a paper experiment, below).
• CPPI: the 20% ceiling is a floor at 80% of the high-water mark — halt-at-floor semantics, deliberately not the continuous m > 0 form that would re-risk into a drawdown. keel's gap risks are separately blunted: no leverage, no shorting, and a routing-time spread gate refuses live entries into thin books.
• Loss-streak protection: keel's consecutive-loss breaker halts after a losing run — the half of "dynamic bankroll management" worth keeping.
• The naive baseline: keel's simulator already benchmarks every rule against DCA — a fixed-budget naive strategy — so no rule can look good merely by losing slowly.

Three things the series argues for that keel rejected (with reasons)

• Optimal-f. Backward-looking growth maximization over a small trade history is overfitting bait. keel's promotion gate already runs a probability-of-backtest-overfitting check (PBO/CSCV) on exactly that question, with sequence-aware methodology and a 100-trade floor. Adopting Optimal-f would import the problem the gate exists to catch.
• Streak-driven sizing (raise on wins, cut on losses). keel's own hourly-evidence caveat applies: ~250 sequential trades inside one market regime are not 250 independent draws, so a streak is precisely the non-evidence a sizing modifier would celebrate. The protective half already exists as the loss-streak breaker; the aggressive half optimizes for psychology, not evidence.
• A Merton γ knob. A continuous risk-aversion parameter with no data to identify it is a guess with a dial on it. keel's fixed 1% risk under hard caps is already more conservative than any plausible γ at current edges.

What keel is actually taking from the series

Three adoptions, all instrumentation around the do-not-bet posture rather than departures from it:

1. A Kelly diagnostic (planned): at promotion time, compute f* from the rule's own fills — win rate, payoff ratio — and print it beside the gates: kelly_f_star = −0.003; risk_pct 1% vs f* ≤ 0: no edge to size on. Report-only. It restates the honest result in the sizing vocabulary at the exact moment a human decides whether to promote.
2. A quarter-Kelly ceiling on the live path (spec'd, dormant): if a rule ever clears the promotion gate, its live risk fraction caps at min(risk_pct, 0.25 × f*_forward) — the series' individual-investor recommendation — computed on forward fills, fail-closed. Until a rule promotes, there is nothing for it to bind, so it stays a specification in the review document, not code.
3. A drawdown-throttle experiment (paper only): scale effective risk (1 − d/D) between a soft floor and the existing hard ceiling, measured through keel's hash-chained experiment ledger over ≥ n=100 pooled forward signals, judged on drawdown distribution, recovery time, and forgone winners — with the series' own "slower recovery" cost stated up front.

The closing honesty

The series ends by choosing strategies by risk tolerance and psychology. keel's constraint comes earlier in the pipeline: the edge itself is unproven, and every formula in the literature agrees on what that implies. When a rule someday clears keel's gates, the fractional-Kelly ceiling is waiting to make sure the first live dollar doesn't celebrate too hard. Until then, the engine's most sophisticated sizing feature is the one it had all along: it declines.

Full findings, formulas, and per-source verdicts: the source review. keel's evidence pipeline context: the ~940 signals/year announcement.