Backtesting

Determinism

Backtests (research/backtest.py) are deterministic: fixed seeds, frozen

datasets, no wall-clock dependence. Same inputs ⇒ same results.

Friction model

Spread, slippage and latency are modeled, not ignored:

• fills on historical bid/ask (direction-aware),
• tick-level SL/TP first-touch resolution,
• logical latency applied to the decision path,
• robustness stress re-runs under degraded friction (spread/slippage/latency

shocks) — a strategy that only works under ideal fills does not pass.

Boundary honesty

• Backtests produce evidence, not promises. Results feed the validation

pipeline; they never bypass the OOS gate.

• Fees/commissions follow the configured friction profile; metrics without the

underlying evidence render n/a — never fabricated zeros.

• Streaming replay (research/streaming_replay.py) is the high-fidelity

sibling: it runs the shared engine over historical events with a logical

clock (zero sleeps) and simulated fills — and is test-enforced to never call

order_send.

Reading a backtest report

Reports include funnel (signals → orders → fills → exits), per-regime and

per-confidence stratification, MFE/MAE, and exit-mechanism attribution. See

Validation for how these numbers are judged.