Strategy Generation
A strategy that exists only as prose is a mood. This skill converts ideas into
specs that machines can refuse, code that gates can check, and claims that
survive the validation pipeline — or die there honestly.
The pipeline (six stages, three gates)
IDEA → SPEC ──gate 1──→ CODE → RETURNS ──gate 2──→ RISK ──gate 3──→ PAPER
│ │ │
│ validate_strategy_spec │ backtest- │ risk-assessment
│ (this skill's script) │ validation │ profile
- DATA — establish market, universe, frequency, history budget, and
whether the data pipeline is PIT-reviewed (
point-in-time-research).
- SPEC — fill the contract in
references/spec-schema.md: named
archetype, testable entry/exit logic, explicit parameters, at least one
hard risk control, a cost assumption in bps, and variants_tried.
- Gate 1 —
python scripts/validate_strategy_spec.py SPEC.json.
FAIL means fix the spec; do not write code around a broken spec.
- CODE — implement exactly the spec. Parameters come from
signal.parameters, not from numbers invented mid-implementation. If
implementation forces a change, update the spec and re-run gate 1.
- Gate 2 — produce the returns series and run it through
backtest-validation with --trials = the spec's variants_tried
(updated if more variants were tried during coding — honesty compounds).
A FAIL verdict is a result: report it, don't tune until it passes.
- Gate 3 — run
risk-assessment on the surviving equity curve and
holdings. High-risk verdicts require explicit user confirmation.
- PAPER — the strongest deployment this skill may recommend is paper
trading. Live execution is a user decision outside this skill's authority.
Archetype discipline
Name the archetype (momentum, mean_reversion, multi_factor,
pairs_trading, technical, ml_enhanced, event_driven, value) and stay
inside its logic. "Momentum with 12 confirming indicators" is not momentum —
it is 12 extra parameters begging the overfit preflight to fail.
Counting variants (the honesty ledger)
Every parameter combination tried, universe swapped, or date range adjusted
increments variants_tried. The number follows the strategy to
backtest-validation's Deflated Sharpe Ratio. Understating it does not improve
the strategy; it just moves the lie downstream where the DSR will price it.
Guardrails
- No code before the spec passes gate 1.
- No performance claims before gate 2; no "expected returns" from thin air.
- Forbidden claims ("guaranteed", "risk-free", "稳赚", "保证收益") fail the
spec mechanically — the validator scans every string field.
- Position sizing "all-in" or absent hard risk controls fail the spec.
- Deploy advice caps at paper trading, always.
1---2name: strategy-generation3description: Turn a natural-language trading idea into a disciplined, testable strategy spec and implementation. Trigger for "帮我生成一个动量策略", "写一个BTC均值 回归策略", "设计多因子选股", "做一个ETF轮动", "generate a momentum strategy", "build me a trading strategy", "turn this idea into code", or whenever the user (1) describes a trading idea and wants it made concrete, (2) asks for strategy code, (3) wants strategy variants compared, or (4) asks to improve an existing strategy. Fire even for vague asks ("有什么好策略推荐吗" — the answer is a disciplined spec, not a tip). Do NOT trigger for validating an already-backtested result (backtest-validation) or for pure factor evaluation (factor-research).4---56# Strategy Generation78A strategy that exists only as prose is a mood. This skill converts ideas into9specs that machines can refuse, code that gates can check, and claims that10survive the validation pipeline — or die there honestly.1112## The pipeline (six stages, three gates)1314```15IDEA → SPEC ──gate 1──→ CODE → RETURNS ──gate 2──→ RISK ──gate 3──→ PAPER16 │ │ │17 │ validate_strategy_spec │ backtest- │ risk-assessment18 │ (this skill's script) │ validation │ profile19```20211. **DATA** — establish market, universe, frequency, history budget, and22 whether the data pipeline is PIT-reviewed (`point-in-time-research`).232. **SPEC** — fill the contract in `references/spec-schema.md`: named24 archetype, testable entry/exit logic, explicit parameters, at least one25 hard risk control, a cost assumption in bps, and `variants_tried`.263. **Gate 1** — `python scripts/validate_strategy_spec.py SPEC.json`.27 FAIL means fix the spec; do not write code around a broken spec.284. **CODE** — implement exactly the spec. Parameters come from29 `signal.parameters`, not from numbers invented mid-implementation. If30 implementation forces a change, update the spec and re-run gate 1.315. **Gate 2** — produce the returns series and run it through32 `backtest-validation` with `--trials` = the spec's `variants_tried`33 (updated if more variants were tried during coding — honesty compounds).34 A FAIL verdict is a result: report it, don't tune until it passes.356. **Gate 3** — run `risk-assessment` on the surviving equity curve and36 holdings. High-risk verdicts require explicit user confirmation.377. **PAPER** — the strongest deployment this skill may recommend is paper38 trading. Live execution is a user decision outside this skill's authority.3940## Archetype discipline4142Name the archetype (`momentum`, `mean_reversion`, `multi_factor`,43`pairs_trading`, `technical`, `ml_enhanced`, `event_driven`, `value`) and stay44inside its logic. "Momentum with 12 confirming indicators" is not momentum —45it is 12 extra parameters begging the overfit preflight to fail.4647## Counting variants (the honesty ledger)4849Every parameter combination tried, universe swapped, or date range adjusted50increments `variants_tried`. The number follows the strategy to51backtest-validation's Deflated Sharpe Ratio. Understating it does not improve52the strategy; it just moves the lie downstream where the DSR will price it.5354## Guardrails5556- No code before the spec passes gate 1.57- No performance claims before gate 2; no "expected returns" from thin air.58- Forbidden claims ("guaranteed", "risk-free", "稳赚", "保证收益") fail the59 spec mechanically — the validator scans every string field.60- Position sizing "all-in" or absent hard risk controls fail the spec.61- Deploy advice caps at paper trading, always.
artherahq