Get Backtest Summary

MDD-01 (2026-10-09): per-step coverage floor and what it removed.

A step is measured only when both of its monthly snapshots hold at least min_snapshot_universe distinct scored symbols AND at least max(min_basket_names, min_priced_fraction x basket) of the basket is priced at both ends. Any other step abstains: it is absent from the curves and from every figure (never published as a flat 0.00% month) and is listed in abstained. measured_steps_contiguous is False when an abstained step sits BETWEEN measured ones, i.e. the curve joins non-adjacent periods.

Per-series equity curve segment — normalized to 100.0 at index 0.

P0-3 (2026-08-03): universe is the equal-weight mean of every scored symbol over the same rebalance dates — the like-for-like comparator. qqq is retired (always empty, kept for response-shape compatibility). An empty spy with benchmark_degraded=True on the response means the external feed failed — never rendered as a flat 0% line.

Headline metrics — None when sufficient_data=False.

P0-3: cagr / sharpe / alpha_vs_spy are additionally None until n_rebalances >= 24 (annualizing a ~4-month window is statistically meaningless — df=3). MDD-01 (2026-10-09): max_drawdown is gated the same way (a sample maximum is most biased on a short panel); the response's max_drawdown_suppressed says so. alpha_vs_spy is a CAGR DIFFERENCE, not risk-adjusted alpha; surfaces must label it "CAGR gap vs SPY".

Per-benchmark figures over the SAME measured steps as the strategy.

MDD-01e (2026-10-09): max_drawdown is published exactly when the strategy's is (n_rebalances >= 24 measured steps) so the strategy drawdown is never shown without its comparator; None otherwise, and None for SPY when the feed degraded.

Per-benchmark figures over the SAME measured steps as the strategy.

MDD-01e (2026-10-09): max_drawdown is published exactly when the strategy's is (n_rebalances >= 24 measured steps) so the strategy drawdown is never shown without its comparator; None otherwise, and None for SPY when the feed degraded.

Per-benchmark figures over the SAME measured steps as the strategy.

MDD-01e (2026-10-09): max_drawdown is published exactly when the strategy's is (n_rebalances >= 24 measured steps) so the strategy drawdown is never shown without its comparator; None otherwise, and None for SPY when the feed degraded.

D3 / VALGATE hard gates (item 37, .planning/PRICE-PREDICTION-TIMING-PLAN-2026-09-03.md).

Deflated Sharpe Ratio + Probability of Backtest Overfitting (Bailey & López de Prado 2014 — computed once in app/services/factor_ic_significance.py and never re-derived here) applied to the persisted strategy result. dsr/pbo/passes_gates are all None and validation_status stays 'not_evaluated' until the annualized Sharpe itself is populated (fewer than MIN_REBALANCES_FOR_ANNUALIZED periods — nothing to deflate yet). passes_gates is true only when dsr >= 0.95 AND, where PBO is applicable, pbo < 0.35; a False or None value here means this result must NOT be read as validated, regardless of how the headline metrics look.

pbo_applicable (2026-10-08): PBO measures selection among >= 2 candidate configurations. This strategy is one fixed configuration, so pbo is null and pbo_applicable is false: no overfitting-of-selection test was run, and the gate rests on the DSR alone. It is never published as 0.0. pbo_n_candidates is the number of configurations compared.

dsr IS A PROBABILITY IN [0, 1] AS OF 2026-10-07, NOT A RATIO. It estimates P(true Sharpe > the Sharpe the best of n_trials null configurations would be expected to show by luck), so 0.95 is a one-sided 95% confidence level. The field previously carried a ratio-valued statistic whose documented 1.5 threshold was unreachable at any trial count of 4 or more (DSR-01); it was never non-null on this surface, because the live result has had fewer than MIN_REBALANCES_FOR_ANNUALIZED rebalances throughout, so no published value changed. Both fields stay None rather than 0.0 so "not measurable" can never read as "measured".

Compliance copy surfaced to the frontend (BACK-05).

All language locks to "historical model performance" framing and explicitly negates "forecast" / "recommendation" / "guarantee".