Replay a recorded bet log

examples/replay_bet_log.py takes a chronological CSV of probability estimates and recorded outcomes, replays the same bets through Kelly, Half Kelly and Fixed Fraction (5%), then reduces each history with summarize_history(). Each strategy gets a fresh $1,000 bankroll, with all funds bettable and no per-transaction loss cap.

Warning

The committed log is synthetic, generated once with Python’s random.Random(140). It is one illustrative path, not evidence of an edge, a forecast, or a strategy recommendation. Keeks is educational only and does not provide financial advice. Replaying your own estimates does not establish that those estimates are calibrated or predict future results.

Input format

The committed examples/data/synthetic_bet_log.csv starts with:

probability,outcome
0.52,1
0.52,0
0.52,0
0.55,1
0.55,0

probability is the estimate made before the outcome was known, in [0, 1]. outcome is an integer: 1 for a win, 0 for a loss. Keep rows in their original order: reversing them changes drawdown and can change sizing when bankroll safeguards bind. The data README records the fixed-seed generation recipe; no generator runs during replay.

This example assumes the same even-money odds for every row: net profit of 1 unit per unit staked on a win, loss of 1 on a loss. Both the strategies’ transaction_cost_rate (per unit) and the simulator’s fee_per_bet (flat currency fee per settlement) are zero. Adapt these controls together when using your own log; this script does not model changing odds per row.

Reading the comparison

All three strategies replay the same recorded outcomes without drawing random numbers. Kelly’s larger stakes produce deeper losses on this path.

Summary of this synthetic replay

strategy	input_rows	bets	start	end	total_return	geometric_growth_per_period	max_drawdown	ruined
Kelly	60	60	1000.0	610.96	-0.38903999999999994	-0.008178436261167654	0.7805768701804972	False
Half Kelly	60	60	1000.0	930.98	-0.06901999999999997	-0.0011912479703595213	0.48380001132866346	False
Fixed 5%	60	60	1000.0	927.65	-0.07235000000000003	-0.001250896517715594	0.42332380952380955	False

The CSV reports input rows separately from bets: history records transactions, so skipped bets do not add entries. The plot’s horizontal axis also counts settled bets, rather than input rows. All 60 rows settle for each strategy in this fixture. Kelly and Half Kelly vary stakes with the estimated edge; Fixed 5% stakes a constant fraction when probability meets its default 0.5 threshold.

max_drawdown is the largest fractional loss from a running peak (0.78 means about 78%), not a single-bet loss limit. total_return is end / start - 1; geometric_growth_per_period compounds over recorded settlements, not days or years. Undefined metrics are blank in the CSV, rather than invented zeros. These are summaries of one path, not averages over repeated simulations.

Reproducing it and using your own log

From a repository checkout with development dependencies installed:

uv run python examples/replay_bet_log.py
# Or run the examples group:
make examples

The script writes examples/output/replay_bet_log.csv and examples/output/replay_bet_log.png using a headless matplotlib backend. Rerunning with the same input produces byte-identical CSV. To compare your own chronological CSV with the same odds assumptions:

uv run python examples/replay_bet_log.py /path/to/my_bets.csv

This overwrites the two example outputs; keep your personal log and generated results out of version control. The complete loading, replay and summary code is in replay_bet_log.py.