Skip to main content

Grid Strategy Backtesting

The grid backtest simulates the AI-driven grid market-making strategy against historical data. It tests the grid placement, fill simulation, and exit order logic without placing real orders.

Quick Start​

uv run python backtest.py \
  --symbol ETH \
  --strategy grid \
  --start-date 2024-01-01 \
  --end-date 2024-12-01

How the Grid Backtest Works​

  1. Data download: Historical OHLCV data is fetched and cached.

  2. AI grid decisions: Every scheduler.interval_minutes, the GridAgent calls the LLM to determine direction and grid width — same as live trading.

  3. Grid simulation: The math engine computes grid levels. Fill simulation checks whether each historical candle's high/low would have triggered fills at each grid level.

  4. Fill handling: When a simulated fill occurs, exit orders (TP/SL) are also simulated. The grid is re-evaluated at each scheduler interval.

  5. P&L tracking: Realized P&L from fills, unrealized P&L from open grid positions, and fees are all tracked.

Using a Grid-Specific Config​

Grid backtesting uses config.grid.yaml by default. Override with --config:

uv run python backtest.py --symbol ETH --strategy grid \
  --start-date 2024-01-01 --end-date 2024-12-01 \
  --config config.grid.yaml

Sample config.grid.yaml for backtesting:

trading:
  symbols: [ETH]
  max_total_investment: 500
  max_leverage: 3               # conservative for backtesting
  grid_limit_order_take_profit_enabled: true
  grid_limit_order_stop_loss_enabled: true
  grid_reduce_only_exit_orders_enabled: true

agent:
  grid_width:
    min_pct: 0.02
    max_pct: 0.10
    fallback_pct: 0.04
    ai_blend_weight: 0.35

scheduler:
  interval_minutes: 15          # longer interval = fewer LLM calls = cheaper backtest

Output​

backtest_results/
└── ETH_grid_20241201_143022/
    ├── live_report.json         # grid state snapshots + fills
    ├── summary.json             # aggregate metrics
    ├── grid_levels.png          # visualization of grid placement over time
    └── fill_log.csv             # every simulated fill

Key Metrics for Grid Strategy​

{
  "total_return_pct": 18.7,
  "grid_fills": 1847,
  "avg_profit_per_fill_usd": 0.43,
  "max_concurrent_levels": 8,
  "max_unrealized_loss_usd": -234.50,
  "direction_accuracy": 0.61,
  "regime_distribution": {
    "trending": 0.35,
    "ranging": 0.48,
    "volatile": 0.17
  }
}

Direction Accuracy​

Shows what percentage of AI direction calls (LONG/SHORT/NEUTRAL) aligned with the subsequent price movement. Higher is better; 50% is chance level.

Regime Distribution​

Shows how much time was spent in each market regime during the backtest period. Grid strategies perform best in ranging markets (high fill rate) and worst in strong trends (accumulates directional exposure).

Grid Backtest vs. Single Agent Backtest​

AspectSingle AgentGrid
Decision frequencyEvery intervalEvery interval (grid re-eval) + continuous fills
LLM callsOne per intervalOne per interval (grid direction)
P&L driversDirectional tradesFill frequency × spread
RiskStop-out from adverse directionTrend riding against grid direction
Best market conditionTrendingRanging

Analyzing Grid Performance by Regime​

To understand when your grid config performs well, look at the regime_distribution in the backtest results. If most backtest time was ranging and the strategy was profitable, verify it also holds up during the trending/volatile periods.

tip

Run the grid backtest over multiple date ranges to capture different market regimes. A grid config tuned only on 2024's ranging periods may underperform in 2025's trending markets.

Resuming Interrupted Grid Backtests​

uv run python backtest.py --symbol ETH --strategy grid \
  --start-date 2024-01-01 --end-date 2024-12-01 \
  --resume-from backtest_results/ETH_grid_20241201_143022/live_report.json

Comparing Grid Configurations​

Use A/B Comparison Backtesting to test different ai_blend_weight or grid_width settings side by side.

Next Steps​