backtest

Backtest Skill

This skill exists because picking the backtest window well is more important than picking the strategy. A strategy that looks great over the last 30 days of a grinding uptrend tells you nothing about how it behaves when the market turns. A window that's too short produces <10 trades and makes every ratio meaningless. A window that's too long bleeds across regime changes and flatters overfit strategies.

The /create-strategy skill hands off here after Phase B-bulk (or earlier, for single-candidate builds). Users can also invoke /backtest directly on any existing strategy_id.

Trigger

Activate when the user:

• Explicitly asks to backtest ("run a backtest on ref-004 with ETH 1h", "how would this have done over the last year", "test this")
• Has just promoted a draft from /create-strategy and needs a verdict
• Asks to re-evaluate a paper or live strategy on a new window
• Asks "is this strategy any good"

Do NOT activate for:

• Promotion decisions → /promote-strategy
• Live monitoring / evaluations → /monitor-trades
• Authoring new strategies → /create-strategy

Phase A — Collect

Required inputs:

1. strategy_id — UUID from list_strategies or a create_strategy_manual response. If the user says "the one I just made" and there's only one draft, infer it. Otherwise call list_strategies(status="draft") and present choices.

Optional overrides (skill proposes defaults in Phase B — don't ask the user for these cold):

• lookback window — days, hours, or explicit start/end dates. If user volunteers one, respect it.
• slippage_pct / fee_pct — override trading_defaults.json. Only touch if the user mentions they want realistic frictions tuned.

Pull the strategy's asset + timeframe via get_strategy(strategy_id) so the window sizing can reason about bar counts.

Phase B — Size the window

Target bar count: 2000–5000. That's the range where ratio metrics stabilize (enough for 20–100 trades) without the window spanning so much history that regime drift dominates. Translate to wall-clock via the timeframe: