提示詞詳情

模型

- （GPT 5.2）

Tokens

806

提示詞介紹

Example input [TICKERS]: NVDA, AMD, AVGO, ASML, TSM, META, AMZN, GOOGL [HORIZON]: 3–6 months [OBJECTIVE]: Long-only; maximize risk-adjusted returns using momentum + trend confirmation [CONSTRAINTS]: Max position 10%; avoid extreme single-name volatility; monthly rebalance [DATA]: I have 5y daily close+volume and 5y daily SPY; no fundamentals or options data Example output 1) Clarifying Questions (0–5) Are corporate actions already adjusted in your price series (splits/dividends)? Are you allowed to hold all names equally, or do you prefer mega-cap bias caps? Do you want a trend filter on SPY to reduce crash risk (yes/no)? Do you have borrow constraints (not relevant if long-only, confirm)? 2) Data & Assumptions Data available: 5y daily close + volume for tickers; 5y daily SPY close. Missing: intraday data, fundamentals, earnings dates/calendar, sector labels, options/implied vol. Assumptions: Prices are adjusted for splits/dividends (if not, must adjust). Liquidity is “acceptable” given large-cap nature, but still check ADV. 3) Signal Library (3–6 signals) S1: 12–1 Momentum (cross-sectional) Recipe: compute total return from t-252 to t-21; rank across universe. Intuition: intermediate momentum persistence. Failure modes: sharp reversals, crowding unwind. S2: 3–1 Momentum (tactical overlay) Recipe: total return from t-63 to t-21; rank. Intuition: captures shorter trend continuation. Failure: whipsaw in choppy tape. S3: Trend confirmation (time-series) Recipe: price above 200D MA and 50D MA slope positive. Intuition: avoid buying “falling knives.” Failure: late to turns; gap risk. S4: Relative strength vs SPY Recipe: regression-free proxy: ratio (stock / SPY) trend up over last 3–6 months. Intuition: prefer leaders vs market. Failure: market rotation. S5: Liquidity screen Recipe: rank by median dollar volume; exclude bottom quartile. Intuition: reduce execution slippage. Failure: may exclude smaller high-momentum winners. 4) Cross-Sectional Scoring Model Preprocess: Winsorize each signal at 5th/95th percentiles. Z-score signals cross-sectionally each rebalance date. Composite alpha (weights sum to 100): S1 12–1 momentum: 35 S2 3–1 momentum: 25 S3 trend confirmation: 20 S4 relative strength vs SPY: 15 S5 liquidity: 5 Missing values: if any signal missing, replace with cross-sectional median and penalize score by <X> points. 5) Risk Overlay & Constraints Mapping Risk model approach: Estimate rolling volatility (e.g., 60D) per stock. Estimate market beta to SPY (rolling regression 1y) for risk monitoring (not strict constraint if long-only). Constraints enforcement: Hard: max position 10%. Soft: penalize high-vol names via vol-scaling weights. 6) Ranked Shortlist (top 5) LOW–MED confidence (no event calendar, no fundamentals; purely price/volume) NVDA Drivers: likely strong on S1/S2/S4 if leadership persists. Risks: momentum reversal, crowded positioning; big drawdowns possible. Falsification: breaks long-term trend + relative strength down for <N> weeks. Catalysts: trend continuation / sector leadership (no specific events assumed). Confidence: MED (data adequate for momentum). META Drivers: tends to score well on relative strength when in leadership regime. Risks: sharp rotation out of megacap growth; gap risk. Falsification: ratio vs SPY rolls over + MA breaks. Catalysts: continued risk-on tape. Confidence: MED. AVGO Drivers: semiconductor leadership + momentum persistence. Risks: drawdown spikes; sector rotation. Falsification: 200D trend break + underperformance vs SPY. Catalysts: continued semis strength. Confidence: MED. AMZN Drivers: cross-sectional momentum if trend intact. Risks: whipsaw; rotation into defensives. Falsification: trend filter fails + weak relative strength. Catalysts: risk-on continuation. Confidence: MED. TSM Drivers: momentum + relative strength if strong. Risks: idiosyncratic geopolitical headline risk; gap moves. Falsification: sudden regime shift; trend break. Catalysts: sector-wide momentum. Confidence: LOW–MED (headline risk not modeled). 7) Portfolio Construction Plan Approach: long-only, score-weighted with volatility scaling. Weighting: Start with score-proportional weights on top 5–6. Apply vol-scaling: target equal risk contribution using 60D vol. Cap each at 10%. Rebalance: monthly (or every 4 weeks) with mid-month risk check. Turnover control: Replace names only if score rank drops below top <K> or trend filter fails. Transaction cost guardrails: Assume <X> bps per trade; penalize high turnover in objective. 8) Robustness Checklist (quant hygiene) Avoid lookahead: use close at t to compute signals for rebalance at t+1 open/close. Survivorship: ensure tickers exist historically; use point-in-time constituents if expanding universe later. Sensitivity: Try windows: 12–1 vs 9–1; 3–1 vs 6–1. Weights ±10–20% and confirm stable ranking. Stability: Rolling IC (information coefficient) of composite score vs future returns. Turnover and drawdown constraints. Regimes: Split backtest into high-vol vs low-vol SPY regimes. 9) Minimal Data Request (to upgrade confidence) Adjusted prices (close + corporate actions) + dividends. Daily high/low (for trend robustness) and dollar volume. Sector/industry classification. Event calendar (earnings dates) to manage gap risk. 10) Backtest Spec (walk-forward) Walk-forward: rolling 3y train (for parameter selection), 6m test; slide monthly. Benchmark: SPY. Metrics: CAGR, vol, Sharpe, max drawdown, turnover, hit rate, exposure by name.