Outcome Model

Train an immutable causal candidate from the bar feature store. Artifacts always start unpromoted; keep the terminal holdout frozen while developing candidates.

Export training matrix

Fixed v1 label contract: XAUUSD H1, 24 bars, 1.5 ATR target, 1 ATR stop, with one row for each long/short side:


python3 scripts/export_training_matrix.py --output data/exports/outcome_v1.parquet

Only the causal allow-list is exported. Incomplete or unreliable rows are excluded. The adjacent .manifest.json records source hashes, schema and label versions, purged chronological folds, and the untouched final holdout.

Train a candidate

Train a new immutable candidate only after reviewing the HistData report:


.venv/bin/python scripts/build_bar_features.py --symbol XAUUSD --timeframe H1 --rebuild
.venv/bin/python scripts/export_training_matrix.py --output data/exports/outcome_v1.parquet
.venv/bin/python scripts/train_outcome_model.py \
  --dataset data/exports/outcome_v1.parquet \
  --version xauusd-h1-outcome-v1-candidate-YYYYMMDD

The artifact records the promotion gates and always starts with promoted=false. Keep the terminal holdout frozen while developing candidates. Five continuous years of HistData are required before any deployability claim.

Inference endpoints

Method	Path	Role
GET	`/outcome-model/shadow`	Unpromoted causal candidate inference
GET	`/outcome-model/shadow/history`	Reveal-gated forward-shadow ledger

Operational health (candle, Capital, feature, model, parity, worker) is on GET /health.

Pipeline sketch


candles ──▶ build_bar_features ──▶ export_training_matrix
                                          │
                                          ▼
                                 train_outcome_model
                                          │
                                          ▼
                         artifact (promoted=false) + holdout frozen