Backtest engine

Replay historical cross-platform tick data against a strategy. Computes ROI, Sharpe, drawdown, hit rate, CLV, p-value, and an out-of-sample split. No other comparator does this for PM × Kalshi × sportsbooks.

Sample strategy

Below is the strategy contract the engine will accept when live. You can already define strategies in this shape via the API for forward validation.

{
  "name": "PM-fade-SB-NBA-H2H",
  "description": "Buy the SB favourite when PM is at least 3pp above the SB no-vig consensus between 30min and 2h before tip-off.",
  "trigger": {
    "sport": "basketball",
    "market": "h2h",
    "condition": {
      "operator": "gte",
      "lhs": {
        "source": "pm",
        "metric": "implied"
      },
      "rhs": {
        "source": "sb",
        "metric": "no_vig"
      },
      "delta_pp": 3
    },
    "timing": {
      "min_minutes_before_start": 30,
      "max_minutes_before_start": 120
    }
  },
  "action": {
    "side": "buy_lower_implied",
    "stake_usd": 100,
    "require_verdict": "executable"
  }
}

Methodology

• No metrics below n=300. Statistical noise dominates. Engine refuses to display ROI / Sharpe if the strategy matched fewer than 300 trades in the window.
• Out-of-sample split. Last 20% of the time window is held out. In-sample vs out-of-sample ROI displayed side-by-side. If they diverge by >50%, the strategy is overfitted.
• P-value reported. Tests the null hypothesis "ROI = 0" via a one-sample t-test on per-trade returns. Below 0.05 is meaningful.
• CLV vs closing line. Each trade scored against the closing implied probability on the same venue. Long-term sharp behaviour shows up as positive CLV even when short-term ROI is noise.
• Slippage applied. Entry price uses depth at the stake size (fill1k/fill5k/fill25k from Moat 3). Strategies that look great at $100 but disappear at $5k are flagged.