OPEN-SOURCE RESEARCH SYSTEM

RESEARCH · PAPER INFERENCE

ARES Engine

Quantitative machine-learning research infrastructure for chronological, cost-aware evaluation—built to make false confidence harder, not to market a trading return.

View SourceAI Engineering

THE RESEARCH PROBLEM

Quantitative ML is easy to fool yourself with.

Future information can leak backward. Search can overfit a narrow history. Friction can erase an apparent edge. A strong result can be unstable across folds or regimes. ARES makes those assumptions visible and testable.

CHRONOLOGICAL PIPELINE

Time only moves forward.

  1. 01Data GenerationsImmutable, verified input generations.
  2. 02Quality + FeaturesCausal features and declared horizons.
  3. 03Purged Walk-ForwardExpanding training history; purge before validation.
  4. 04Model SearchCandidates evaluated across chronological folds.
  5. 05Cost-Aware ScoreFees, slippage, turnover, drawdown, and stability.
  6. 06Champion / ChallengerHard gates plus a required improvement margin.
  7. 07Paper InferenceVerified bundles emit read-only long / flat / short signals.

MAKE TIME VISIBLE

Separate research history from the future sequence.

Purge belongs to ordinary walk-forward folds. Embargo belongs to the separate final-holdout workflow. The holdout is not a mandatory promotion or paper-inference gate.

ROBUST SCORE

Five terms. Separate hard gates.

The scalar score ranks candidates; separate gates still enforce minimum Sharpe, return, drawdown, turnover, trade count, solvency, and doubled-cost stress under the default configuration.

01

Median Sharpe

A central risk-adjusted fold result.

02

Median Return × 2

A weighted return contribution across folds.

03

Mean AUC Above 0.5

Signal discrimination contributes only above chance.

04

Worst Drawdown Penalty

The worst fold drawdown reduces the score.

05

Sharpe Variability Penalty

Unstable fold performance reduces the score.

CHAMPION / CHALLENGER

New is not automatically better.

CHALLENGERverified bundle
→
HARD GATESrisk · cost · solvency
→
+0.05 MARGINdefault improvement
→
CHAMPIONatomic manifest pointer

OPTIONAL LOCKED HOLDOUT

One local commitment. Historical evidence only.

protocol boundary · current main

commandares locked-holdout … --holdout-bars 720

one_time_evaluationtrue

historical_result_onlytrue

profitability_establishedfalse

order_execution_enabledfalse

The public repository does not contain a committed real-market holdout report or verified benchmark win.

PUBLIC ARCHITECTURE

Research infrastructure you can inspect.

19documented components in the public system map
31public test modules on current main
7latest CI jobs passed
0live-order paths

Version 0.1.0 remains alpha. Current main includes newer optional holdout work that is not part of that tagged release.

LIMITATIONS

A research engine is not evidence of profitability.

Historical tests do not predict future returnsCosts, slippage, latency, and liquidity remain assumptionsOHLC bars cannot establish intrabar ordering or fillsThe locked holdout is optional and locally enforcedPaper inference does not place live ordersNo investment advice or independent certification

SOURCE + TESTS + ARCHITECTURE

Inspect ARES on GitHub.

Read the pipeline, validation rules, promotion logic, holdout protocol, tests, and documented boundaries.

Open RepositoryAI Engineering