Skip to main content

ML Pipeline and Configuration

Main Flow​

  1. Backtests can write ML chunk files (--ml).
  2. npx @tradejs/cli ml-export merges chunk files into one JSONL dataset.
  3. Train script builds derived windows (holdout, prod, walk-forward).
  4. Python train container generates model artifacts and reports.
  5. Runtime inference uses gRPC (ML_GRPC_ADDRESS).

Real Commands​

npx @tradejs/cli backtest --config trendline --ml
npx @tradejs/cli ml-export
npx @tradejs/cli ml-inspect
npx @tradejs/cli ml-train:trendline:xgboost

Useful .env Example​

ML_GRPC_ADDRESS=127.0.0.1:50051
ML_TRAIN_RECENT_DAYS=60
ML_TRAIN_TEST_DAYS=30
ML_TRAIN_WALK_FORWARD_FOLDS=2
ML_TRAIN_FEATURE_PROFILE=all
ML_TRAIN_FEATURE_SET=enriched
ML_TRAIN_ENSEMBLE=1
ML_TRAIN_ENSEMBLE_MEMBERS=3

Dataset File Patterns​

Source chunk files:

ml-dataset-<strategy>-<chunkId>.jsonl

Derived split files (generated automatically):

*.holdout-train.<key>.jsonl
*.holdout-test.<key>.jsonl
*.prod.<key>.jsonl
*.walk-forward-fold-<N>.train.<key>.jsonl
*.walk-forward-fold-<N>.test.<key>.jsonl

Local Artifact Chain (Export -> Train -> Signals)​

  1. npx @tradejs/cli backtest --ml writes chunk files to local project folder: data/ml/export/ml-dataset-<strategy>-chunk-<chunkId>.jsonl
  2. npx @tradejs/cli ml-export merges chunks into: data/ml/export/ml-dataset-<strategy>-merged-<timestamp>.jsonl
  3. npx @tradejs/cli ml-train:latest reads export files from data/ml/export and writes model aliases to: data/ml/models/<Strategy>.joblib or ensemble aliases data/ml/models/<Strategy>.modelN.joblib
  4. ML infer service must read the same model directory (MODEL_DIR).
  5. npx @tradejs/cli signals (with strategy config ML_ENABLED=true) sends signal.strategy to gRPC Predict. Inference service loads <Strategy>.joblib / <Strategy>.modelN.joblib for that strategy.

If strategy name and model alias prefix match, runtime automatically uses the trained local model.

Quality and Causality​

  • Train validates lookahead leakage on timestamp-like fields.
  • Guard can be disabled only for debugging:
ML_TRAIN_DISABLE_CAUSALITY_GUARD=1
  • Runtime inference uses the same feature trimming policy as train.

Reports​

Each train run saves markdown and HTML reports with:

  • holdout metrics,
  • walk-forward metrics,
  • threshold tables,
  • holdout TOP feature table.