qj-strategy-reviewer

QuantJourney Strategy Reviewer

Use this skill to review a QuantJourney strategy for correctness before trusting its results. Check each item.

Timing / look-ahead

• Signals must use data available at decision time. The engine lags weights by one bar (shift(1)); do not also peek forward inside _compute_signals (e.g. .shift(-1), centered rolling windows, using iloc[i+1]).
• Rolling indicators must not include the current bar's future. Warm-up bars should be NaN/0, not filled with later data.

Data handling

• Missing bars stay unavailable, not 0% returns. Don't fillna(0) on a returns series you then treat as real observations.
• _compute_weights returning NaN → the name should drop out, not distort the normalization.

Weights and exposure

• Long/cash: weights in [0, cap], sum ≤ 1. Confirm max_position_size matches the intended cap.
• Long/short: check gross (Σ|w|) and net (Σw). Market-neutral should be net ≈ 0; a stray leg breaks neutrality.
• Renormalization should not silently move positions you meant to freeze.

Costs and realism

• A strategy that only works at zero cost does not work. Confirm slippage and commissions are set for anything turnover-heavy (intraday, reversal, daily rebalancing).
• Short strategies: borrow/financing is not modeled — is the edge just the omitted carry? Flag it.

Mode fit

• If the edge depends on stop-losses, brackets, or limit fills, it belongs in order mode, not weight mode.
• Order mode: check gap handling expectations (a gap through a stop fills at the open, not the stop price) and that per-order commission minimums aren't assumed per fill.

Output

Report findings as: severity, where (method/line), what's wrong, and the fix. Separate real correctness bugs from style. Verify claims against the code, not assumptions.