PRD §3 C1. Jesse ships this; keel has nothing equivalent.
The question it answers
Right now keel's live open question is: a reconstructed 14.9% win rate against a 14.9% break-even — is that real, or one lucky path? PBO/CSCV cannot answer it (wrong failure mode), and Deflated Sharpe only discounts for the number of trials tried.
Resampling attacks it directly.
What Jesse does
monte_carlo_trades — reshuffles the sequence of completed trades. Same trades, different order; if the equity curve's shape depended on lucky ordering, the distribution reveals it.monte_carlo_candles — perturbs the price series itself and re-runs.- Candle perturbation pipelines (
jesse/candle_pipelines/) — gaussian_noise, gaussian_resampler, moving_block_bootstrap, pluggable onto a strategy.
Moving-block bootstrap is the most valuable of these for keel, because it preserves local autocorrelation while resampling — the right null for a trend-following rule, where naive IID resampling would destroy the very structure turtle_breakout trades.
Fit with keel
keel/sim/portfolio_sim.py already walks a shared timeline through the same strategy.engine.evaluate() the live path uses — the resampled series can feed the same entry point.- Metrics are hand-rolled pure
Decimal (keel/sim/metrics.py), so a distribution over N runs is straightforward.
- Results should land in the hash-chained trials ledger (
keel/research/ledger.py) with provenance, like every other experiment.
Complementarity worth noting
This is frequentist resampling; the Quant Lab note (#432) proposes a Bayesian π_edge = P(p > p_be | D). They answer the same question from different directions and their agreement or disagreement is itself informative. Both should be available before the pooled review.
Scope
keel trials monte-carlo --rule <id> --mode {trades,candles} with a seed for reproducibility (determinism is a keel constraint).- Report the observed equity curve's percentile within the resampled distribution.
- Ledger every run.
PRD §3 C1. Jesse ships this; keel has nothing equivalent.
The question it answers
Right now keel's live open question is: a reconstructed 14.9% win rate against a 14.9% break-even — is that real, or one lucky path? PBO/CSCV cannot answer it (wrong failure mode), and Deflated Sharpe only discounts for the number of trials tried.
Resampling attacks it directly.
What Jesse does
monte_carlo_trades— reshuffles the sequence of completed trades. Same trades, different order; if the equity curve's shape depended on lucky ordering, the distribution reveals it.monte_carlo_candles— perturbs the price series itself and re-runs.jesse/candle_pipelines/) —gaussian_noise,gaussian_resampler,moving_block_bootstrap, pluggable onto a strategy.Moving-block bootstrap is the most valuable of these for keel, because it preserves local autocorrelation while resampling — the right null for a trend-following rule, where naive IID resampling would destroy the very structure
turtle_breakouttrades.Fit with keel
keel/sim/portfolio_sim.pyalready walks a shared timeline through the samestrategy.engine.evaluate()the live path uses — the resampled series can feed the same entry point.Decimal(keel/sim/metrics.py), so a distribution over N runs is straightforward.keel/research/ledger.py) with provenance, like every other experiment.Complementarity worth noting
This is frequentist resampling; the Quant Lab note (#432) proposes a Bayesian
π_edge = P(p > p_be | D). They answer the same question from different directions and their agreement or disagreement is itself informative. Both should be available before the pooled review.Scope
keel trials monte-carlo --rule <id> --mode {trades,candles}with a seed for reproducibility (determinism is a keel constraint).