Backtesting

AlgoTradingAI provides backtesting capabilities to evaluate trading strategies against historical data. This guide explains how backtesting works and how to interpret results.

Two Different Things Named Backtesting

Backtesting (this page) replays historical candles through the strategy engine with 0.05% slippage and Zerodha brokerage modelled, and reports out-of-sample metrics.

Live Backtesting is a separate feature that places simulated orders against live prices. Those fills are modelled at the last-traded price with no slippage, spread, brokerage, or liquidity constraint applied, so its P&L is more optimistic than both a historical backtest and live execution. Neither is a prediction of future results.

How Backtesting Works

1. Data Loading: Historical OHLCV candles are fetched from Zerodha KiteConnect (or Redis cache). Candles are aligned to IST and restricted to market hours (09:15-15:30).
2. Strategy Replay: The strategy layer processes candles sequentially, generating CandidateTrade objects using pattern detection (breakout, reversal) and trend filters (SMA crossovers).
3. ML Filter: Each candidate passes through the proprietary ML quality gate. Only candidates above the confidence threshold are accepted as signals.
4. Trade Simulation: Accepted signals are evaluated against their stop-loss and target levels using subsequent candle data. No hindsight bias — decisions use only data available at the signal timestamp.
5. Metrics Calculation: Win rate, profit factor, Sharpe ratio, max drawdown, and other metrics are computed from the set of closed simulated trades.

Backtest Assumptions

Every backtest uses a consistent set of assumptions that are always disclosed:

• Slippage: 0.05% per leg (configurable in strategy params).
• Transaction Costs: Zerodha brokerage schedule (Rs 20 per intraday order, 0 for delivery).
• Data Source: Zerodha historical candles with Redis cache fallback (7-day TTL).
• Walk-Forward: 70% training / 30% validation split. Results reported on out-of-sample period only.
• No Future Data: All indicator values and signals are computed using only data available at the decision timestamp.

AI Advisory Backtest

The AI advisory pipeline supports a separate backtest mode:

• Saved advisory decisions (from /api/advisor/decisions) are replayed through the backtest engine.
• Each decision's recommended action, size, and stop-loss are simulated against actual price movements.
• Results are compared to a Buy & Hold baseline for the same symbol and period.
• Available from the Advisor Backtest page in the sidebar.

Interpreting Results

Key Metrics

• Win Rate: Percentage of trades that hit profit target before stop-loss. Above 50% is generally positive, but depends on risk:reward ratio.
• Profit Factor: Gross profit / gross loss. Above 1.0 means net profitable. Above 1.5 is considered strong.
• Sharpe Ratio: Risk-adjusted return. Above 1.0 is acceptable, above 2.0 is strong.
• Max Drawdown: Largest peak-to-trough decline. Lower is better. Compare to your personal risk tolerance.
• Exposure: Percentage of time capital was deployed. Lower exposure with similar returns indicates more selective trading.

Common Pitfalls

• Overfitting: A strategy that performs exceptionally on historical data but fails on new data. Walk-forward validation helps detect this.
• Survivorship Bias: Testing only on stocks that exist today. Delisted stocks are not included in backtests.
• Slippage Underestimation: Real slippage in illiquid options can be significantly higher than the default 0.05%.

Next Steps

• Indicators & Models — Understand the signals feeding into strategies.
• Performance Methodology — How we report results publicly.
• Risk Disclosure — Important limitations and disclaimers.