A working copy of the tuning dashboard from my options research stack. It shows a strategy's 1,260-configuration parameter grid over 160 sessions, the split-half test that shows whether a surface's shape is real, tuners refitted every night and walked forward against a frozen baseline, and config freezing. Every number here comes from a synthetic panel generated in your browser (seed ); no market data is involved.
Surface
Drag to rotate. The ◆ is the frozen baseline.
Flat
Click a cell to select that config.
Selected config against the baseline
Cumulative $ over the sessions in view.
Split-half test
The referee: refit every night, scored on the next session
The config each tuner trades on session u is fitted only on sessions before u. Haircut is the tuner's in-sample $/session minus its walk-forward $/session on the same sessions: what choosing cost. SDC is the worst drawdown any ordering of the sessions could build; it breaks ties and never vetoes.
Frozen candidates
Freezing fixes a config as of a session. Every later session becomes its forward record, scored against the baseline (B0, always frozen). In the production tool, each frozen config also trades live as a shadow book.
Synthetic data: a weak, smooth signal under heavy-tailed noise that is correlated across neighbouring configs, so an in-sample peak usually looks like a hill and usually isn't one. Generator and arithmetic: synth.js, referee.js.