Experimental models — research results
Backtesting
Comparison of model configurations across a training window and a held-out test window, with confidence intervals where the API reports them.
Experimental model — unvalidated
In testing. No result here is confirmed out-of-sample.
Period selector
Applies to server-side backtest runs
Selections are submitted to the backtest endpoint. No client-side backtest is simulated.
Audited out-of-sample summary
performanceSummary from /api/dashboard — Clopper-Pearson 95% intervals as reported by the backend
Experiment counts
Each research family uses its own experiment unit. Counts are reported separately and never summed.
Chronological Model Backtest
Walk-forward evaluation from /api/backtest/results — training data strictly prior to each test draw
V8 Weighted Ensemble — Holdout Evaluation
GET /api/v8/results — train / validation / held-out test split, 9 optimised signal weights
Distinction labels are mandatory: a historical observation is measured on data the model was fitted to; an out-of-sample result is measured only on pre-draw records.