backtesting-py-oracle_skill

Overview

This skill provides a hardened backtesting.py configuration and checklist for validating ClickHouse SQL oracle results and range-bar pattern backtests. It codifies required constructor flags, common anti-patterns, and deterministic fixes so Python trade output matches SQL sweep results. Use it to run reproducible, bit-atomic comparisons between SQL-generated signals and backtesting.py trades.

How this skill works

The skill inspects backtesting.py runtime configuration, trade sorting, rolling-quantile signal generation, and strategy sizing to detect divergences from SQL oracle outputs. It enforces hedging and non-exclusive orders, prescribes EntryTime sorting of trades, and supplies NaN-safe rolling-quantile logic and data-range defaults. It also outlines a multi-position strategy template that mirrors SQL behavior for overlapping signals.

When to use it

• When validating ClickHouse SQL oracle vs backtesting.py trade outputs
• When running Backtest() with overlapping signals or multi-position modes
• When entry price, timestamp, or signal counts differ between SQL and Python
• When rolling quantile features produce NaNs that poison signals
• When configuring sizing to avoid margin exhaustion with concurrent positions

Best practices

• Always set hedging=True and exclusive_orders=False to allow overlapping trades
• Sort stats._trades by EntryTime before mapping trades to signals
• Skip NaN values when computing rolling quantiles to match SQL NULL handling
• Use an expanded data start (e.g., 2017-01-01) to match SQL data range
• Prefer fixed fractional sizing (e.g., size=0.01) for overlapping positions to avoid margin exhaustion

Example use cases

• Oracle validation run comparing SQL gen600_strategy outputs with backtesting.py trades
• Backtest configuration for multi-position range-bar pattern experiments
• Debugging entry-price mismatches by re-sorting trades on EntryTime
• Fixing cross-asset failures caused by rolling-quantile NaN propagation
• Reproducing SQL signal timestamps by recording signal bar timestamps in next()