alpaca-trading-backtest

Provide a start date, end date, and strategy concept; the agent formalizes the rules, confirms the interpretation, fetches bars and quotes via the Alpaca CLI, and runs a single-file simulation script. Results include trades.csv, equity.csv, summary.json, and a report.md with benchmark comparison.

It supports next_open (signal on bar close, fill at next bar open), time_based fills at a confirmed time of day, and same_bar fills only when explicitly requested with a documented look-ahead warning. Quote-based fills use ask for buys and bid for sells plus slippage.

V1 supports one asset family per run: stocks or crypto, with multiple symbols allowed within the same family. Options require explicit contract selection and fill logic, and mixed-family portfolios are out of scope unless the generated code explicitly models them.

Execution friction combines spread and slippage in basis points, while trading-activity fees are modeled from Alpaca's Brokerage Fee Schedule PDF covering SEC, FINRA TAF, CAT, ORF, and related categories. Modeled and excluded fee categories are recorded in fee_source.json.

A no-trade result is reported directly with the cause: indicator warmup, no signal triggered, insufficient cash, missing data, or calendar filtering. The agent explains which condition applied rather than silently returning empty results.

The agent runs alpaca doctor before any backtest and stops the run if authentication fails, showing the alpaca profile login options. API keys can also be supplied via ALPACA_API_KEY and ALPACA_SECRET_KEY environment variables without writing secrets into generated code.