Experiments
Each phase below corresponds to a distinct round of attempts. They are presented in chronological order; the structural pattern across all of them is summarised on Methodology and the dead-ends are catalogued on What didn't work.
| Phase | Title | Outcome |
|---|---|---|
| Phase 5 | Eight strategies on a 5-ticker × 3y universe | Three "no edge" verdicts, harness regime trade-off discovered |
| Phase 6 | Harness recalibration (6.A-E) | Cheat-test transform_fn, mean-PnL ratio co-condition, window-stability test, position-aware losses, orchestrator launch |
| Phase 7 | Nifty 50 universe expansion (46 stocks) | Position-floor works at scale; shuffled_target now fails consistently — cross-stock memorisation surfaced |
| Phase 8 | Per-stock rolling z-score + static_features_test | L43 hypothesis partial — z-score helps but isn't sufficient |
| Phase 9 | Single-model-per-stock baseline + L47 bug | Harness now trustworthy for per-stock evaluation; per-stock has no time-edge either |
| Phase 11 | Five parallel agents (linear / perm-test / neg-tone / sectors / dropout) | Linear baseline first leakage-clean at scale (L49); dropout candidate's apparent edge later collapsed |
| Phase 12 | Window-stability verdict | L53 bound recalibration; L54 dropout-as-window-artifact confirmation. v0.6 release-ready |
Note: Phase 10 was a planned write-up phase that was folded into Phase 11. The numbering is preserved for traceability with the orchestrator state DB.