feat: add backtester with parameter sweep and walk-forward validation

[HaoweiChan]

Summary

Adds a full-featured backtesting extension to OpenAlice with 6 MCP tools accessible by Alice (or any MCP client).

Phase 1 — Core Engine

• Strategy DSL: safe recursive-descent parser for entry/exit logic — supports price variables, technical indicators (RSI, EMA, SMA, BBANDS, MACD, ATR with arbitrary periods), arithmetic, comparisons, boolean logic, and strategy parameter variables
• Indicator Series: wraps existing analysis-kit indicators to compute full series over historical candle data
• Historical Data: fetches OHLCV candles via OpenBB SDK (equity, crypto, currency) with auto asset-class detection, symbol normalization (BTC/USD → BTCUSD), and on-disk candle caching
• Simulation Engine: candle-by-candle execution with configurable slippage (bps), commissions, stop-loss/take-profit, leverage, and position tracking
• Metrics: total return, Sharpe ratio, max drawdown, win rate, profit factor, equity curve
• Persistence: trade logs (JSONL) and summary reports (JSON) saved to data/backtests/

Phase 2 — Advanced Features

• Parameter Grid Search (runParameterSweep): cartesian product over user-defined parameter ranges, runs all combinations concurrently, ranks by Sharpe ratio, persists results
• Walk-Forward Validation (runWalkForward): splits data into N rolling train/test windows, optimizes on training, validates on test, reports overfitting score
• Volume Realism: flags trades where position size exceeds a configurable fraction of candle volume (warnings, not rejections)
• CCXT Data Source: paginated fetchOHLCV via connected exchange (e.g. Binance), automatic fallback to OpenBB when CCXT unavailable
• DSL Parameter Variables: strategy parameters are injected into the DSL context, enabling expressions like RSI_14 < rsi_oversold

MCP Tools Exposed

Tool	Description
runBacktest	Single strategy backtest with volume warnings
runParameterSweep	Grid search over parameter ranges, ranked by Sharpe
runWalkForward	Train/test validation to detect overfitting
fetchHistoricalOhlcv	Raw candle data (OpenBB or CCXT)
listBacktests	Browse saved backtest results
readBacktestSummary	Read a specific saved backtest summary

Files Changed

• New: src/extension/backtester/ (11 modules — adapter, data, dsl, engine, index, indicators, io, realism, sweep, types, validation)
• Modified: src/main.ts (tool registration with lazy CCXT account getter), src/extension/trading/providers/ccxt/CcxtAccount.ts (added public fetchOHLCV method)
• Specs: OpenSpec artifacts for both phases (proposals, designs, delta specs, task breakdowns, main specs)

Test plan

• runBacktest — EMA crossover on BTC/USD daily, 20 trades, +1.12%, Sharpe 1.21
• fetchHistoricalOhlcv (OpenBB) — 60 candles fetched
• fetchHistoricalOhlcv (CCXT) — 60 candles fetched from Binance
• listBacktests — lists 9 saved results
• readBacktestSummary — reads saved JSON correctly
• runParameterSweep — 12 combinations (4×3), ranked output, 108ms
• runWalkForward — 3 windows, correctly flags overfitting (IS Sharpe 5-12 vs OOS negative)

Made with Cursor