Thirteen strategies on a marginal edge¶

examples/strategy_comparison.py simulates the scenario most profitable bettors actually live in: a marginal edge, not a big one. Every strategy bets 5,000 rounds at a 52% win probability, a 0.95 payoff (typical -110 to -105 odds after line shopping), a 1% transaction cost, and a full-stake loss — an expected value of about +0.4% per bet. At an edge that thin, the Kelly fraction is a fraction of a percent of the bankroll, variance dominates over any realistic sample, and ruin is a live possibility. The script runs 500 seeded simulations per strategy from a $1,000 bankroll and records how many end in ruin.

Thirteen configurations are compared: full, half and quarter Kelly; drawdown-adjusted Kelly; Optimal F; fixed fractions at 5% and 10%; the naive expected-value rule; CPPI; dynamic bankroll management; and the Merton share at risk aversion 1, 2 and 5.

Warning

These are simulated results from a model, not a forecast and not investment advice. The simulator assumes a known, constant win probability and an independent binary outcome per bet — neither of which holds in a real book — and the edge is an assumption, not an estimate. Keeks is an educational library; treat every number below as a property of the model, not a prediction about money.

What the spread says¶

Two-panel figure over 500 seeded simulations of 5,000 bets for thirteen strategies. The top panel is a boxplot of final bankrolls on a linear dollar axis: every box is squashed near zero with a handful of extreme outliers, the largest near 9,000,000 dollars for Fixed 5% and 3,000,000 for Dynamic. The bottom panel plots mean final bankroll with standard-deviation error bars: Fixed 5% shows a small positive mean with error bars several hundred thousand dollars tall, Dynamic shows a mean near zero with error bars about half that size, and every other strategy sits near zero with almost invisible error bars. — Final-bankroll distributions (top) and means with standard deviations (bottom) across 500 seeded simulations per strategy.¶

The two panels make the same point from opposite directions. The boxplot shows medians pinned near zero — ruin is common at this edge — with a few enormous survivors, one Fixed 5% path ending near $9 million. The mean panel shows why the mean is the wrong summary here: it is pulled upward by exactly those survivors, which is why Fixed 5%’s mean carries error bars taller than every other strategy’s entire distribution. The script prints its results table sorted by median, not mean, for this reason.

For a controlled study of the same question — how growth, drawdown and early stops trade against each other as the edge, the cost input, the estimate quality and the loss cap move — see the nine-strategy benchmark, which works under even-money assumptions in benchmarks/. This example is the messier, more realistic cousin: thinner edge, real odds shape, and a cost model that actually bites.

Reproducing it¶

One command, from a checkout of the repository:

uv run python examples/strategy_comparison.py

It runs 6,500 seeded simulations (13 strategies × 500 paths), prints the full results table — mean, median, min, max, standard deviation and ruin rate per strategy, sorted by median — and writes the chart plus examples/output/strategy_comparison.csv.