Skip to content

CHANGELOG.md

Latest commit

 

History

History
36 lines (29 loc) · 1.85 KB

CHANGELOG.md

File metadata and controls

36 lines (29 loc) · 1.85 KB
Raw

Changelog

v1.1.0

Added a fourth guardrail and closed three silent-failure bugs found in a self-audit.

Added

  • deflated_sharpe module — Probabilistic Sharpe Ratio (PSR), Deflated Sharpe Ratio (DSR), and minimum track-record length (Bailey & López de Prado). The multiple-testing correction the first three guardrails don't do: it asks whether a Sharpe survives after accounting for how many variants were tried. Stdlib-only (no new dependency).
  • examples/deflation_demo.py — shows the demo strategy's Sharpe (significant on 1 trial) failing once a realistic search size is admitted.
  • py.typed marker; .gitattributes (LF normalization); this changelog.
  • Tests: test_deflated_sharpe.py, test_silent_failures.py, test_report_and_data.py (73 → 92 tests; report.py/data.py now covered).

Fixed

  • Engine now raises on a non-finite (NaN/inf) strategy weight instead of silently producing a fake flat 0%-return curve.
  • block_bootstrap_delta returns an honest "UNDEFINED" verdict when the metric is undefined on the point estimate or most resamples (e.g. Calmar on a zero-drawdown series), instead of fabricating a confident-looking p_delta_gt_0 = 0.0.
  • audit_no_lookahead reports INCONCLUSIVE (and passed=False) when the strategy never made a real decision on any sampled date, instead of a vacuous {} == {} pass; adds n_nonempty_checked.
  • data.load_prices handles yfinance's single-ticker Series return shape.

Docs

  • Clarified that the cash symbol is a real tradeable instrument (its rebalancing is costed), that {} means "go flat" not "hold", and the block-length caveat for path-dependent metrics.

v1.0.0

Initial public release: look-ahead-safe engine, three guardrails (look-ahead audit, matched-exposure control, block-bootstrap significance), demo strategy, property fuzzer + mutation test, three case studies.