permutation_test

Defined in fynance.research

permutation_test(strategy, data, *, metric='sharpe', n_permutations=200, seed=0)

Permutation test for a spurious edge.

Runs the strategy on the real series, then on n_permutations shuffles of the asset’s returns (which destroys any temporal structure). If the strategy scores as well on shuffled data as on the real data, its edge is not real.

The price series must be strictly positive: the null is built on log-returns (np.diff(np.log(prices))), so a non-positive price would yield nan / -inf returns.

Each run is seeded with a distinct seed derived from seed (the observed run and every permutation), so a stochastic strategy does not see the same model RNG draws on every path — which would bias the null variance. The whole test stays reproducible for a fixed seed.

Parameters:

strategyfynance.strategy.Strategy

The strategy to evaluate.

dataPriceSeries or array-like

Strictly positive price series (np.log is taken).

metricstr

Metric key from the run summary (default "sharpe").

n_permutationsint

Number of shuffles forming the null distribution.

seedint

Master seed for the shuffles and the runs (per-run seeds are derived from it, so the test is fully reproducible).

Returns:

dict

observed, p_value, null_mean, null_std. The p-value is the (smoothed) fraction of shuffles scoring at least the observed metric.