backtest-audit

Static analysis for Python backtesting code.
Catches look-ahead bias, data leakage, and statistical errors — before they invalidate your results.

pip install backtest-audit
backtest-audit check ./strategies/

Why

A backtest that looks profitable is often built on data the strategy could never have had. Look-ahead bias, data leakage, and missing transaction costs are the three most common ways a backtest lies to you. backtest-audit catches these at the code level, without running your strategy.

Output

── strategies/mean_reversion.py ─────────────────────── 2 issues ──

  [LAB001] Negative shift — future data pulled into current timestep  line 34
  df['signal'] = df['close'].shift(-5)
  ↳ Replace shift(-n) with shift(n). Negative periods reference future rows, invalidating the backtest.

  [LEAK002] fit_transform() detected — ensure this is called only on training data  line 61
  X_scaled = scaler.fit_transform(X)
  ↳ Replace fit_transform(X) with fit(X_train).transform(X_train), then transform(X_test) separately.

──────────────────────────────────────────────────────────────────────
1 error, 1 warning

Rules

Full documentation with explanations and code examples: docs/rules/

Usage

# Check a single file
backtest-audit check my_backtest.py

# Check a directory
backtest-audit check ./strategies/

# Output as JSON (for CI pipelines)
backtest-audit check ./strategies/ --format json

# Ignore specific rules
backtest-audit check ./strategies/ --ignore LAB002,STAT002

# List all rules
backtest-audit rules

Installation

pip install backtest-audit

Requires Python 3.8+. The only dependency is click.

How It Works

backtest-audit parses your Python source files into an abstract syntax tree (AST) and walks the tree looking for known problematic patterns. It does not execute your code. Each rule is a pattern matcher — it looks for a specific construct that is known to produce invalid backtest results.

This means it is fast, safe to run on any codebase, and produces no false positives from runtime behavior.

Limitations

• Does not run your backtest or evaluate strategy profitability
• Cannot detect all forms of look-ahead bias — only statically visible patterns
• LEAK001 checks for train_test_split presence in the file, not correct application
• Manual chronological splits (e.g., df[:split_date]) are not yet recognized
• Does not support non-Python backtesting frameworks

License

MIT