Rolling Ensemble — Live Dashboard
Walk-forward validation of the rolling model-selection methodology, and the live paper-trade backfill it's driving.
Walk-Forward Validation
Validating the rolling model-selection methodology now live on develop/main/paper_trade, across four historical test months (2026-04 through 2026-07).
↻
Covers all 12 trained datasets. Originally scoped to the 7 datasets backing the currently-deployed production ensemble (re-picking lgbm vs. xgboost within that lineup). Now also includes alpha_v1_original, alpha_v2_with_cond, alpha_v3_1_cond_filtered_eff, alpha_v5_rollingCondMapped, and alpha_v5_1_rollingCondMapped_eff — trained variants that exist in the repo but were never part of the production selection — since any of the 12 may end up used for inference.
Train — 1 year
12 datasets × {lgbm, xgboost} = 24 candidate models, 3-fold CV each.
Validate — 1 month
Better of lgbm/xgboost kept per dataset, by Kendall Tau. Never sees the test month.
Test — 1 month
12 selected models rank-averaged into one ensemble, scored on the held-out month.
Sharpe by test month
Long, short, and blended — one bar group per held-out month
Combined PnL curve — 2026-04-01 through 2026-07-29
The four held-out test months (2026-04 through 2026-07), extended past 2026-07-23 with a live tail scored by the current deployed model -- 81-day continuous series
Full results
Kendall Tau and Sharpe/cumulative return for the held-out test month, per window
Which estimator won, per dataset per window
Picked by validation-month Kendall Tau — 27 lgbm vs. 21 xgboost selections overall, switches most months
Long and short Sharpe are computed on each side's own daily return series; Total blends both books rather than averaging the two ratios. See experiments/13_walkforward_ensemble/ on claude/exp for the full methodology and scripts.
Live Paper-Trade Backfill
↻
Walk-forward-consistent. 2026-04-02 through 2026-07-23 is scored by the trained window snapshots above (each month scored by a model that never saw it); 2026-07-24 onward uses the current live-deploy model, which is genuinely unseen for those recent days too. This replaces an earlier version that replayed the whole history through one fixed model, in-sample for most of it, and showed impossible +52%/+65% monthly returns.
Daily capital deployed (NT$)
Monthly report
P&L and return, scored against each book's average deployed capital that month
Long and short are each scored against their own deployed capital (e.g. long_pnl / long_capital); Total blends both books' capital and PnL together, so it isn't the simple average of the two sides.