Skip to content

CHANGELOG.md

Latest commit

 

History

History
64 lines (55 loc) · 4.36 KB

CHANGELOG.md

File metadata and controls

64 lines (55 loc) · 4.36 KB
Raw

Changelog

Unreleased — Data contracts become executable

  • An AlphaBench-inspired research protocol compares bounded CoE, ToT and evolutionary candidate search under one DSL, one budget and the existing deterministic falsification pipeline. Search diagnostics are reported per arm; no proposal engine can inspect held-out feedback, judge factors absolutely or bypass validation.
  • xalpha-doctor audits price and fundamental inputs before discovery. Duplicate bars, impossible OHLC, missing notice dates, insufficient chronological capacity, absent neutralization controls and missing tradability fields fail closed in machine-readable JSON.
  • xalpha-demo runs the installed package end to end without a checkout or private dataset. It emits deterministic synthetic inputs, the doctor report, frozen config and full research result in one directory.
  • The daily forward-record summary is rendered from language templates, keeping generated figures out of source code while preserving the English and Chinese reader views.
  • The user-supplied literature registry machine-checks paper metadata, claimed public-code mappings, implementation boundaries and next falsifiable steps. Black--Litterman and Hoeffding now have executable contribution cards; unrelated papers remain visibly excluded.
  • The README, project site and successful synthetic demo expose one consistent, voluntary GitHub Star path without opening a browser, requesting credentials or automating engagement.

0.5.0 — The record publishes itself

The daily record now runs on GitHub Actions from public data, and publishes the name before the session it applies to.

  • tools/daily_record.py — selects the name, scores every published pick whose holding window has elapsed, and redraws. Only picks that were actually published are scored; recomputing what the rule would say today would quietly turn the forward record back into a backtest.
  • tools/audit_returns.py — audit somebody else's backtest in their own fork: CSCV probability of backtest overfitting, deflated Sharpe against the declared trial count, and White's Reality Check across the family. The example that ships with it is 24 variants of pure noise whose best has an annualised Sharpe of 1.11 — against 1.18 expected from noise at that trial count.
  • Specification v4 (dea0e608) pins the data provider, adjustment convention, board scope, window and listing filter. v3 named a universe rule without saying where the universe came from, and two faithful implementations disagreed on 614 of ~4,700 eligible names; v3 keeps the one pick it published and is retired rather than edited.
  • Coverage gates. Under parallel load the data source answers a throttled query with an empty result set and error_code "0", indistinguishable from a symbol with no history — two runs over the same 5,202 codes differed by 143,028 bars. Empty answers are retried on a rebuilt session and swept serially; below 98% symbol coverage or 95% quoting on the ranked session the run aborts rather than selecting a name from whichever symbols replied.
  • Chinese README, anchor navigation, and the first tagged releases.

0.4.0 — Forward records

  • forward.py — frozen specifications with four refusals: no overwrite, digest verified on load, one entry per session, and scoring only fully elapsed holding windows. Factors travel inside the specification as portable DSL so a published record is reproducible by its reader.
  • universe.py — sealed_bar_limits infers limit-locked sessions from the bars. build_panel previously defaulted the flags to False, so a panel built from plain OHLCV backtested fills on locked boards; those legs carry +6.05% against +0.38% for legs that could be traded. point_in_time_eligibility decides membership only from prior sessions.
  • book.py — long-only top-N construction sharing the tranche and turnover accounting with the neutral book, so a difference between them is construction and never charging.
  • session_index in the panel, making trend-quality factors (ts_corr(close, t, w)²) expressible in the DSL at all.

0.3.0 and earlier

Point-in-time alignment, the audited factor DSL, bounded synthesis with purged walk-forward, counterfactual and placebo controls, PBO and deflated Sharpe, the evidence lab (stationary bootstrap, Reality Check, Romano–Wolf step-down), and the decision toolkit.