AgentSkillsCN

backtest

运行ML交易策略的滚动回测，支持参数自定义。

SKILL.md
--- frontmatter
name: backtest
description: Run a walk-forward backtest of the ML trading strategy with configurable parameters
disable-model-invocation: true
argument-hint: "[options: walk-forward, ensemble, stops, regime, quick]"
allowed-tools: Bash(python *), Bash(.venv/bin/python *), Read, Grep, Glob

Walk-Forward Backtest

Run a backtest using scripts/run_backtest.py. Parse the arguments from $ARGUMENTS and map them to CLI flags.

Argument Mapping

User saysCLI flags
walk-forward or wf--walk-forward
ensemble--ensemble
stops--stop-loss --trailing-stop
regime--use-regime
quick--period 6mo --test-period 42
optimize--optimize
kelly--kelly
circuit-breaker or cb--circuit-breaker
conf=X.XX--confidence X.XX
symbols=AAPL,MSFT--symbols AAPL,MSFT
capital=50000--capital 50000
no-chart--no-chart

If no arguments given, use the recommended defaults: --walk-forward --ensemble --confidence 0.55

Steps

  1. •Build the CLI command from the parsed arguments
  2. •Show the user the exact command being run
  3. •Run: .venv/bin/python scripts/run_backtest.py <flags>
    • •Use a timeout of 600 seconds (backtests can be slow)
    • •The script outputs results to stdout
  4. •After completion, read data/backtest_results.png if it exists and show the equity curve
  5. •Summarize key metrics: total return, Sharpe ratio, win rate, number of trades, max drawdown, alpha vs SPY

Important Notes

  • •Always use .venv/bin/python (system Python causes TF deadlocks)
  • •Walk-forward backtests retrain models in each window and take 5-15 minutes
  • •The default confidence threshold is 0.55 (optimal from previous analysis)
  • •Current best result: +76% return, 0.81 Sharpe, 63% win rate (walk-forward, no stops, 0.55 conf)