cyb50-quant/research/chinext50_regime_project/memory/2026-04-11.md

2026-04-11

• User highlighted unresolved profit problem after phase-1 scaffold: improve trend follow-back after false cuts, focused on execution timing + entry-specific exit experiments.
• Created and completed OpenSpec change experiment-timing-and-entry-exit; archived as openspec/changes/archive/2026-04-10-experiment-timing-and-entry-exit/.
• Added new experiment pipeline: pipelines/regime_lite_experiments.py.
  • timing modes: lag1, fast_entry
  • entry-specific overlay: trend-entry hold floor + stop guard
  • outputs: CSV/JSON/HTML + baseline/best ledgers
• Added tests: tests/test_regime_lite_experiments.py.
• README updated with experiment command + artifacts.
• Targeted tests passed: py -m pytest -q tests/test_regime_lite_experiments.py tests/test_regime_lite_pipeline.py -> 5 passed.
• Real PIT experiment run:
  • command: py pipelines/regime_lite_experiments.py --pit-csv outputs/ingestion_derived_full50_20260409_v2/pit/chinext50_pit.csv --output-dir outputs/regime_lite_experiments_20260411
  • report: outputs/regime_lite_experiments_20260411/regime_lite_experiment_report.html
  • recommended variant: combo_fast_hold3
  • baseline -> best deltas:
  • annual_return: 0.04814 -> 0.09358 (+0.04544)
  • max_drawdown: 0.31242 -> 0.25812 (-0.05430)
  • annual_turnover: 8.48588 -> 6.10884 (-2.37704)
  • avg_days_to_reach_exposure_after_trend_entry: 11.16 -> 0.66
• Spec validation after archive: openspec validate --specs --strict -> 32 passed.
• 2026-04-11: Added split-sample robustness runs for experiment board:
  • 2020-2023: recommended combo_fast_hold3, annual_return_delta_vs_baseline +0.0542, max_drawdown_delta -0.0469, annual_turnover_delta -2.9647.
  • 2024-2026: recommended combo_fast_hold5, annual_return_delta_vs_baseline +0.0307, max_drawdown_delta -0.0340, annual_turnover_delta -1.5692.
  • Conclusion: timing+entry-exit improvements are directionally consistent across both splits, with shorter hold favored in earlier sample and slightly longer hold favored in recent sample.