Quant Strategy

quant-strategy — a strategy that survives an honest backtest

What this sub-skill is for

Standing up a single, standalone trading strategy: define the signal, backtest it honestly (walk-forward, no lookahead, no leakage, realistic costs + slippage), and only then judge whether it has an edge. Loaded by new-project for any quant/strategy/backtest project. This is the lighter cousin of a full, disciplined honest-eval trading project and inherits its sacred rule: honest backtests above all — a flattering equity curve is a failure, an honest losing one is a success.

Mandatory grill-questions (fold into the Definition of Ready)

• Edge thesis: what is the mechanism of the edge (why should this make money)? What would falsify it? Is it momentum / mean-reversion / carry / event / stat-arb?
• Universe & data: which instruments (crypto / equities / FX)? Data source, frequency, history length, survivorship-bias-free? Point-in-time correct (no restated data)?
• Costs reality: commissions, slippage model, spread, borrow/funding, market impact. What fill assumption (close, next-open, mid, VWAP)? These are decided up front, not tuned.
• Validation design: walk-forward windows (train/test roll), out-of-sample fraction, and a locked final holdout opened once. How many parameters — and the overfit budget?
• Lookahead audit: does any signal at bar t use bar-t (or future) information for a fill at bar t? Any indicator computed on the full series instead of expanding window?
• Risk & sizing: position sizing, max drawdown tolerance, leverage, stop logic.
• Mode: paper/research/backtest-only (default) — live is a separate, manual decision.

Project sub-agents to generate (.claude/agents/) — honest-eval roster

• quant-engineer — implements the strategy + backtester modules in Python with pytest tests, green before reporting; fail-closed on money/risk logic; never flips to live (delegate-by-default for strategy code).
• backtest-analyst — runs the backtest and evaluates it honestly: total/annualized return, Sharpe, Sortino, max drawdown, trade count, win-rate, profit-factor, cost share of gross, exposure. Flags Sharpe > 3 / win-rate > 70% as leakage suspects, fragile results (< 5 trades or one window), and IS-vs-OOS gaps > 2× as overfit. Never changes strategy code. Delegate-by-default before any result is believed.
• quant-strategy-scout — researches the strategy/signal against public evidence (papers, practitioner blogs, repos): edge thesis, evidence quality, known pitfalls, feasibility. Read-only.
• trading-reflexion — reads the ledger/reports/logs and proposes 1–2 concrete experiments (hypothesis → test → criterion) into the task queue; honors a "don't re-propose discarded ideas" rule. Proposes only — never deploys, never edits strategy or risk code.

Tools / CLIs / MCP / skills needed

• Python + .venv (honest-eval pattern): pandas/polars, numpy, a backtest engine (vectorbt / backtrader / custom event-driven), pyarrow for parquet, pytest. Install at environment-readiness (surface + offer; never auto-install). Use a dedicated venv python.
• Data: a free market-data source (free-data-only by default); cache to parquet.
• Supabase MCP (chain) only if storing runs/trades in Postgres; n8n MCP (chain) for scheduled research/backtest runs.
• Global skills/agents to chain: deep-research and the quant-strategy-scout agent for the edge thesis with evidence; the backtest-analyst agent for the honest eval; market-researcher (agent) for broker/fee/data-source facts; code-review and verify on the backtester before trusting any number.

File / asset nudges (on top of the base set)

• BACKTESTS.md — the honest results doc: per-run return/Sharpe/Sortino/max drawdown/trades/win-rate/profit-factor, the walk-forward setup, the cost & slippage assumptions, the lookahead/leakage audit done, OOS result, and known failure regimes. Bad-but-honest results recorded straight.
• STRATEGY.md — the edge thesis, entry/exit rules, parameters, and falsification criteria.
• RISK.md — sizing, leverage, drawdown limits, kill-switches.
• data/ (parquet cache, git-ignored if large; manifest tracked), backtests/ (one logged run per config: params + metrics + seed + git SHA), research/reflexion-journal.md.
• .env.template (keys only, no values; real .env git-ignored).

Stack defaults & done-bar

Default stack: Python + .venv, pandas + a vectorized/event-driven backtester, point-in-time parquet data, walk-forward validation, results logged to backtests/ and summarized in BACKTESTS.md. Paper/research mode only by default. Done-bar (all must hold):

1. The strategy passes an honest out-of-sample / walk-forward test with the locked holdout opened exactly once.
2. Backtest includes realistic costs + slippage; the edge survives a cost +50% stress (or BACKTESTS.md honestly states it does not).
3. No lookahead/leakage — audited by backtest-analyst; bar-t fills use only ≤ t−1 information; indicators use expanding/rolling windows, not the full series.
4. All assumptions (data, costs, fills, sizing) are documented in BACKTESTS.md.
5. The reflexion loop proposes experiments into the queue and never auto-deploys.

Guardrails

• Honest backtests are sacred. An overfit/leaky positive result is worthless and dangerous; an honest negative result is a real finding. Never tune the backtester to make a strategy look good.
• No lookahead, ever. A signal must not use information unavailable at decision time; fills must use realistic, executable prices. The backtest-analyst flags any violation.
• Walk-forward, not single-split fitting. One in-sample fit proves nothing; roll the windows and keep a final untouched holdout.
• Costs and slippage are not optional. A strategy that only works at zero cost has no edge. Model spread, commission, slippage, and (where relevant) funding/borrow.
• Beware overfitting the parameter search. Track the parameter/overfit budget; treat Sharpe > 3 or win-rate > 70% at high frequency as a leakage suspect until proven otherwise.
• Paper/research only by default. Never enable live trading, set a live mode, or place real orders; going live is a separate, explicit, human decision with a stop-and-confirm gate (e.g. a TRADING_MODE=live invariant that defaults off).
• Reflexion proposes, never deploys — and never edits strategy/risk code.
• Mark unverified edge claims as unverified; commits under the user's own name only — no Claude co-author.