Loading…

How to read this: the Holdout column now runs the strategy through the same point-in-time, capital-constrained backtest as the live account — finite cash (no margin), filing-date entry (no look-ahead), oldest-first rotation, and the −8% stop-loss — so it's a like-for-like forecast of live, not an idealized one. Live is your actual realized results since the anchor date; it fills in as positions close.

Equity of the capital-constrained holdout backtest. Toggle indicators to see the path, the underwater (drawdown) curve, day-to-day P&L, and how the rolling risk-adjusted return evolved.

The honest expected-performance estimate. The data is split into sequential folds; for each fold the entire pipeline — model training and strategy-parameter selection — is re-run using only data available (with a purge/embargo) before that fold, then traded through the same capital-constrained simulator as the holdout above. Nothing here was chosen with hindsight. Because it covers far more time than the single holdout window, its Sharpe carries tighter error bars — shown as a 95% block-bootstrap confidence interval, alongside PSR, the probability the true Sharpe exceeds zero after adjusting for the sample's skew and fat tails.

How today's champion policy would have traded every year we have data for. Models are still trained point-in-time per fold, so the predictions are out-of-sample — but the strategy policy is the current one, held fixed throughout. That policy was selected knowing this history, so treat this as a characterization of the policy (which regimes suit it, how deep its drawdowns get), not as an achievable track record. For that, use the walk-forward curve above.

Live trading knobs from champion_strategy_params.json (robustness-gated). Admin-only. If active saved_strategy_params.pkl drifts from the champion, a warning appears below.