PM Quant Strategy Authoring

⚠️ **Trading warning / Avertissement trading** : informational only, not investment advice. A backtest describes the past of one market, auto-tuning can overfit that past beautifully, and neither promises that a strategy will make money live. Curated by Skill Harbor: the quantitative strategy-authoring runbook of oracle3, the open-source prediction-market engine and MCP server for Polymarket and Kalshi. Where the sibling agent skill is for LLM-driven decisions, this one is for numbers: it scaffolds a QuantStrategy subclass whose parameters live in the constructor as plain JSON-serializable values with meanings, bounds and defaults, precisely so the research auto-tune command can search them later. The workflow is strict about shape: strategy logic belongs in its own strategy file (never hard-coded into a command-line script), event handling stays limited to the event types the strategy actually needs, decisions are recorded, and runs are reproducible from the recorded file path, class name, parameters and command. Validation is a dry run, then a single-market backtest against a recorded history file, and only then preparation for tuning. The oldest rule in the book is stated as a hard rule: never use future information, no look-ahead. From the YichengYang-Ethan/oracle3-prediction-market-agent repository (Apache-2.0). Honest caveats: the more parameters a grid search tunes on the same history, the more the result describes that history instead of the market; treat tuned parameters as a hypothesis that paper trading still has to judge. Skill Harbor never reviews the code, review it yourself before use.