研究主題：Best SPY entry/exit signals with profit target and stop loss using SPY daily and hourly bars for all available history. Use ML, technical indicators, or any method to explore alpha. Continue from Run #23 findings: RSI(2) mean reversion with 5% PT and 1% SL achieved 1.68 Sharpe. Explore NEW approaches not yet tested. · 產生時間：2026-03-12 21:26:55 UTC
本次研究迴圈共進行 10 輪， 產生 41 個假說， 其中 0 個通過驗證、 0 個被拒絕、 41 個待驗證。
總耗時 0 秒， 消耗 0 tokens。
| 輪次 | Agent 數 | Agents | 假說數 | 耗時 | Tokens |
|---|---|---|---|---|---|
| R1 | 3 | optimizer, researcher, risk-auditor | 24 | — | — |
| R2 | 3 | optimizer, researcher, risk-auditor | 9 | — | — |
| R3 | 3 | optimizer, researcher, risk-auditor | 0 | — | — |
| R4 | 3 | optimizer, researcher, risk-auditor | 0 | — | — |
| R5 | 3 | optimizer, researcher, risk-auditor | 0 | — | — |
| R6 | 3 | optimizer, researcher, risk-auditor | 0 | — | — |
| R7 | 3 | optimizer, researcher, risk-auditor | 0 | — | — |
| R8 | 3 | optimizer, researcher, risk-auditor | 0 | — | — |
| R9 | 3 | optimizer, researcher, risk-auditor | 8 | — | — |
| R10 | 3 | optimizer, researcher, risk-auditor | 0 | — | — |
| 輪次 | 綜合分數 | Sharpe | Drawdown | 樣本量 | 決策 | 改善% |
|---|---|---|---|---|---|---|
| R1 | 0.5952 | 0.0000 | 1.0000 | 1.0000 | KEEP | — |
| R2 | 0.7016 | 0.3030 | 1.0000 | 1.0000 | KEEP | 17.88% |
| R3 | 0.1625 | 0.0000 | 0.5000 | 0.0000 | DISCARD | -76.84% |
| R4 | 0.2321 | 0.0000 | 0.5000 | 0.0000 | DISCARD | -66.92% |
| R5 | 0.2750 | 0.0000 | 0.5000 | 0.0000 | DISCARD | -60.80% |
| R6 | 0.6500 | 1.0000 | 0.0000 | 1.0000 | KEEP | -7.35% |
| R7 | 0.2750 | 0.0000 | 0.5000 | 0.0000 | DISCARD | -60.80% |
| R8 | 0.1693 | 0.0000 | 0.5000 | 0.0000 | DISCARD | -75.87% |
| R9 | 0.2750 | 0.0000 | 0.5000 | 0.0000 | DISCARD | -60.80% |
| R9 | 0.2750 | 0.0000 | 0.5000 | 0.0000 | DISCARD | -60.80% |
| R10 | 0.1400 | 0.0000 | 0.5000 | 0.0000 | DISCARD | -80.05% |
| 狀態 | 維度 | 信心 | 假說 |
|---|---|---|---|
| pending | validation_framework | 0.9700 | LOOK-AHEAD BIAS CHECKLIST: All entry signals must execute at NEXT bar open (not current bar close). Specifically: (1) If signal fires on day T close, entry price = day T+1 open. (2) If exit signal fires on day T close (e.g., RSI>50), exit price = day T+1 open. (3) For intraday signals on hourly bars... |
| pending | validation_framework | 0.9600 | SHARPE CALCULATION STANDARD: All strategies must report Sharpe using the DAILY EQUITY CURVE method, not per-trade annualization. Specifically: (1) Build daily equity curve including cash days (return=0 or risk-free rate). (2) Sharpe = mean(daily_excess_returns) / std(daily_excess_returns) * sqrt(252... |
| pending | validation_framework | 0.9500 | MULTIPLE TESTING CORRECTION: When testing N parameter combinations (PT/SL grids, indicator thresholds, etc.), apply Bonferroni-Holm correction to all reported p-values and compute Deflated Sharpe Ratio (DSR). Rules: (1) For grid searches with N<=20 combinations, use Bonferroni: adjusted_alpha = 0.05... |
| pending | validation_framework | 0.9500 | MINIMUM SAMPLE SIZE REQUIREMENTS: (1) Minimum 100 trades for ANY Sharpe estimate to be reported (below this, report only win rate and avg return with explicit 'INSUFFICIENT SAMPLE' warning). (2) Minimum 30 trades per regime sub-period (bull/bear/sideways) for regime analysis. (3) Minimum 50 trades p... |
| pending | validation_framework | 0.9400 | WALK-FORWARD VALIDATION PROTOCOL: (1) Use EXPANDING window (not rolling) with minimum 5 non-overlapping OOS windows. (2) Each OOS window must be AT LEAST 3 years for daily strategies, 6 months for intraday. (3) Windows MUST NOT overlap in OOS periods — Run #23 R5 found 20-60% OOS overlap artificiall... |
| pending | validation_framework | 0.9300 | REGIME ROBUSTNESS REQUIREMENTS: Every strategy must be evaluated across at least 3 distinct market regimes with performance breakdown. Define regimes as: (1) BULL: trailing 12-month SPY return > 10% AND VIX < 20, (2) BEAR: trailing 12-month SPY return < -10% OR max drawdown in last 6 months > 15%, (... |
| pending | validation_framework | 0.9000 | TRANSACTION COST AND SLIPPAGE MODEL: (1) Minimum 10 bps round-trip for daily SPY trades (5 bps each way). (2) For intraday/hourly strategies: 5 bps each way PLUS 1 tick slippage (0.01/share). (3) For options strategies: use bid-ask midpoint minus 30% of spread width as realistic fill. (4) Report res... |
| pending | validation_framework | 0.9300 | FAT-TAIL DEPENDENCY TEST (MANDATORY): Compute strategy Sharpe after: (1) Removing top 5% of trades by P&L. (2) Removing top 10% of trades by P&L. (3) Removing all trades that hit profit target within 2 days ('lucky trades'). If Sharpe drops below 0 after removing top 10%, the strategy is a 'lottery ... |
| pending | validation_framework | 0.9100 | TRADE OVERLAP AND POSITION MANAGEMENT: (1) No new position may be entered while an existing position is still open (single-position constraint), unless the strategy is explicitly designed as a portfolio strategy with defined position limits. (2) If overlapping trades are allowed, report BOTH single-... |
| pending | validation_framework | 0.9200 | SURVIVORSHIP BIAS ACKNOWLEDGMENT: SPY is the most successful equity index ETF in history. Any buy-the-dip strategy will look spectacular in hindsight on an asset that ALWAYS recovered. Run #24 MUST: (1) Test any SPY signal on at least 2 non-US equity indices (e.g., EFA, EEM) as out-of-sample assets.... |
| pending | validation_framework | 0.9400 | BENCHMARK COMPARISON STANDARD: Every strategy must be compared against THREE benchmarks: (1) SPY buy-and-hold over the identical period with identical Sharpe methodology. (2) 60/40 SPY/AGG portfolio (for risk-adjusted comparison). (3) Random entry with same trade frequency and holding period (permut... |
| pending | validation_framework | 0.9100 | PRE-REGISTRATION AND HYPOTHESIS TRACKING: Before any backtest is run, the researcher/optimizer must declare: (1) The exact entry/exit rules with all parameters fixed. (2) The primary metric (daily equity Sharpe). (3) The success threshold (e.g., Sharpe > 0.5 net of costs, statistically significant v... |
| pending | volatility_structure | 1.0000 | VIX Term Structure Carry Trade: Go long SPY when VIX futures are in contango (VIX/VIX3M ratio < 0.9, front month cheaper than back month) and short/cash when in backwardation (ratio > 1.05). The VIX term structure is one of the most robust risk premia in finance - contango reflects the insurance pre... |
| pending | volatility_structure | 1.0000 | Realized-Implied Volatility Spread (Variance Risk Premium): Buy SPY when the variance risk premium (VRP = implied vol - realized vol) is high (VRP > 75th percentile of trailing 252-day window), indicating options are expensive relative to actual moves, and the market is pricing in fear that is likel... |
| pending | flow_sentiment | 1.0000 | Put-Call Ratio Extreme Contrarian Signal from SPY Options Data: Using the actual SPY options data in /root/data/thetadata/equity_options/eod/, compute daily put-call volume ratio and put-call open interest ratio. Entry: when put/call volume ratio > 2.0 (extreme fear, measured at EOD), go long SPY at... |
| pending | ml_volatility | 1.0000 | Gradient Boosting on VIX + Options Features (NOT price features): Train a LightGBM/sklearn GradientBoosting model to predict next-5-day SPY direction, but EXCLUSIVELY using volatility and options-derived features - NOT price-derived features (no RSI, no MA, no momentum). Features: (1) VIX level, (2)... |
| pending | cross_asset_momentum | 1.0000 | Multi-Asset Momentum Rotation (SPY + TLT + GLD): Instead of timing SPY entry/exit (proven futile in Run #23), allocate monthly across SPY, TLT (long treasury bonds), and GLD (gold) based on 12-1 month momentum (Antonacci dual momentum). Each month: rank the 3 assets by trailing 12-month return exclu... |
| pending | portfolio_construction | 1.0000 | Risk Parity with Volatility Scaling (SPY + TLT Equal Risk): Allocate between SPY and TLT such that each contributes equal portfolio volatility (inverse volatility weighting). Rebalance monthly. When combined portfolio realized vol > 15% annualized, scale down total exposure proportionally. This is N... |
| pending | intraday_options | 1.0000 | Intraday Gap Fill Strategy Using SPY Options Data: When SPY opens with a gap down > 0.5% from previous close (measured at 9:31 AM), buy SPY via a short-dated ATM call option (0-1 DTE) and exit at the gap fill level or EOD, whichever comes first. Gap fills are one of the most persistent intraday phen... |
| pending | calendar_anomalies | 1.0000 | Monthly Seasonal + VIX Regime Composite: Combine two of the most robust calendar anomalies with VIX regime filtering. Rules: (1) 'Sell in May' - invest in SPY Nov-Apr, cash May-Oct. (2) Within invested months, apply Turn-of-Month effect: increase to 125% weight on last 2 and first 3 trading days of ... |
| pending | options_selling | 1.0000 | SPX 0DTE Options Selling Strategy (Iron Condor): Rather than trying to TIME SPY (proven futile), SELL short-dated SPX options to harvest the volatility risk premium directly. Strategy: sell daily 0DTE iron condors on SPX, with wings at 0.10 delta (approx 1.5-2% OTM). Collect premium and let expire. ... |
| pending | trend_following | 1.0000 | Adaptive Trend Following with Breakout Confirmation (Turtle-style): Use a 20-day Donchian Channel breakout (buy on new 20-day high close) with ATR-based position sizing and trailing stop. Enter long when SPY closes at a 20-day high AND the 50-day SMA is rising (slope > 0). Position size = 1% risk / ... |
| pending | cross_sectional_momentum | 1.0000 | Cross-Sectional Equity Momentum Applied to Sector ETFs (Relative Strength Rotation): Instead of timing SPY, rotate monthly among the 11 SPDR sector ETFs (XLK, XLF, XLV, XLE, XLI, XLP, XLU, XLB, XLY, XLRE, XLC) based on 6-month relative strength. Go long top 3 sectors, equal-weighted. Apply absolute ... |
| pending | microstructure | 1.0000 | Overnight vs Intraday Return Decomposition Strategy: Exploit the well-documented overnight return premium. SPY has historically earned most of its returns overnight (close-to-open) rather than intraday (open-to-close). Strategy: buy SPY at 3:55 PM close, sell at next day 9:35 AM open. Repeat daily. ... |
| 狀態 | 維度 | 信心 | 假說 |
|---|---|---|---|
| pending | technical_indicators | 1.0000 | Connors RSI (CRSI) for SPY Mean Reversion: CRSI combines three components — RSI(3) of price, RSI(3) of the up/down streak length, and percentile rank of today's % change over trailing 100 days. Buy SPY at T+1 open when CRSI < 10 (extreme oversold on all three dimensions simultaneously). Exit at T+1 ... |
| pending | technical_indicators | 1.0000 | Williams %R with Divergence Detection: Williams %R is mathematically identical to an inverted Fast Stochastic %K — it measures where today's close falls within the N-day high-low range. The novel angle is DIVERGENCE detection: when price makes a new 20-day low but Williams %R(14) makes a HIGHER low ... |
| pending | technical_indicators | 1.0000 | Heikin-Ashi Trend Detection for SPY Regime Identification: Heikin-Ashi (HA) candles smooth price action by averaging open/close with prior bar values: HA_Close = (O+H+L+C)/4, HA_Open = (prev_HA_Open + prev_HA_Close)/2. A TREND is identified when HA candles have no lower wick (bullish) or no upper wi... |
| pending | technical_indicators | 1.0000 | Ichimoku Cloud System for SPY Trend + Support/Resistance: The Ichimoku Kinko Hyo system uses 5 lines: Tenkan-sen (9-period midpoint), Kijun-sen (26-period midpoint), Senkou Span A (midpoint of Tenkan/Kijun shifted forward 26 periods), Senkou Span B (52-period midpoint shifted forward 26 periods), an... |
| pending | technical_indicators | 1.0000 | Stochastic RSI (StochRSI) Extreme Readings for SPY: StochRSI applies the Stochastic oscillator formula to RSI values instead of price: StochRSI = (RSI - min(RSI,N)) / (max(RSI,N) - min(RSI,N)) over N periods. This creates a more sensitive oscillator that reaches extreme values (0 or 1) more frequent... |
| pending | technical_indicators | 1.0000 | DeMark TD Sequential (TD9/TD13) Exhaustion Counts for SPY: Tom DeMark's Sequential indicator counts consecutive bars where the close is higher/lower than the close 4 bars earlier. A completed Setup of 9 consecutive qualifying bars signals potential exhaustion. A subsequent Countdown of 13 bars (clos... |
| pending | technical_indicators | 1.0000 | Commodity Channel Index (CCI) Divergence for SPY Turning Points: CCI measures how far price has deviated from its statistical mean using: CCI = (Typical Price - SMA(TP,N)) / (0.015 * Mean Deviation). CCI divergence occurs when price makes a new high/low but CCI does NOT confirm. Strategy: buy SPY at... |
| pending | technical_indicators | 1.0000 | ADX + DI Crossover with Trend Strength Filter for SPY: The Average Directional Index (ADX) measures trend STRENGTH (not direction), while +DI and -DI measure directional movement. Strategy: (1) Only trade when ADX(14) > 25 (confirming a trend exists). (2) Buy SPY at T+1 open when +DI crosses above -... |
| pending | technical_indicators | 1.0000 | META-FINDING: Divergence Indicators as a Class Outperform Level-Based Indicators on SPY. Across the 8 indicators evaluated, a clear pattern emerges: DIVERGENCE-based signals (MACD divergence Sharpe 0.806 from Round 1, Williams %R divergence, CCI divergence) are structurally superior to LEVEL-based s... |
| 狀態 | 維度 | 信心 | 假說 |
|---|---|---|---|
| pending | risk_audit_monte_carlo | 0.8800 | Monte Carlo (10k paths, 10yr, fat-tail empirical resampling): Combined strategy has 99.1% probability of positive returns over 10yr (vs B&H 94.0%). P(beat B&H) = 100.0%. P(lose >20%) = 0.1% (vs B&H 3.1%). Median CAGR: combined 17.61% vs B&H 11.43%. The combined strategy trades ~1-2% median CAGR for ... |
| pending | risk_audit_worst_case | 0.8500 | Worst-case analysis: (A) Double crisis (2008+2020 B2B): Combined DD = -79.3% vs B&H -93.8%. (B) VIX>30 for 6 months: Lost put income ~1.44% — manageable. (C) VRP disappears: Strategy degrades gracefully to SPY+DualMomentum (still valid). KEY INSIGHT: The worst-case for the combined strategy (~-35% t... |
| pending | risk_audit_historical | 0.9200 | Historical backtest (2003-2025): Combined CAGR=13.25% Sharpe=0.85 MaxDD=-35.89% Calmar=0.369. vs B&H: CAGR=11.06% Sharpe=0.657 MaxDD=-55.19% Calmar=0.2. Bootstrap Sharpe diff: 0.1929, 95% CI [0.0755, 0.3289], P(Combined>B&H)=100.0%. Sharpe improvement is statistically significant but drawdown improv... |
| pending | risk_audit_final_verdict | 0.9500 | FINAL SCORECARD — Alpha: 4/10 (risk premium, not alpha). Robustness: 8/10 (components well-researched, but regime-dependent). Complexity: 6/10 (lower=simpler) (monthly put management, options knowledge needed). Risk: 6/10 (higher=safer) (-25% to -35% DD still realistic). HONEST VERDICT: CONDITIONAL ... |
| pending | risk_audit_monte_carlo | 0.8800 | Monte Carlo (10k paths, 10yr, fat-tail empirical resampling): Combined strategy has 99.1% probability of positive returns over 10yr (vs B&H 94.0%). P(beat B&H) = 100.0%. P(lose >20%) = 0.1% (vs B&H 3.1%). Median CAGR: combined 17.61% vs B&H 11.43%. The combined strategy trades ~1-2% median CAGR for ... |
| pending | risk_audit_worst_case | 0.8500 | Worst-case analysis: (A) Double crisis (2008+2020 B2B): Combined DD = -79.3% vs B&H -93.8%. (B) VIX>30 for 6 months: Lost put income ~1.44% — manageable. (C) VRP disappears: Strategy degrades gracefully to SPY+DualMomentum (still valid). KEY INSIGHT: The worst-case for the combined strategy (~-35% t... |
| pending | risk_audit_historical | 0.9200 | Historical backtest (2003-2025): Combined CAGR=13.25% Sharpe=0.85 MaxDD=-35.89% Calmar=0.369. vs B&H: CAGR=11.06% Sharpe=0.657 MaxDD=-55.19% Calmar=0.2. Bootstrap Sharpe diff: 0.1929, 95% CI [0.0755, 0.3289], P(Combined>B&H)=100.0%. Sharpe improvement is statistically significant but drawdown improv... |
| pending | risk_audit_final_verdict | 0.9500 | FINAL SCORECARD — Alpha: 4/10 (risk premium, not alpha). Robustness: 8/10 (components well-researched, but regime-dependent). Complexity: 6/10 (lower=simpler) (monthly put management, options knowledge needed). Risk: 6/10 (higher=safer) (-25% to -35% DD still realistic). HONEST VERDICT: CONDITIONAL ... |
| 輪次 | 維度 | 方法 | 失敗原因 | 分數 |
|---|---|---|---|---|
| R3 | expiration_choice | Round 3 research on expiration_choice | Composite score 0.1625 below baseline | 0.1625 |
| R4 | position_sizing | Round 4 research on position_sizing | Composite score 0.2321 below baseline | 0.2321 |
| R5 | exit_rules | Round 5 research on exit_rules | Composite score 0.2750 below baseline | 0.2750 |
| R7 | volatility_surface | Round 7 research on volatility_surface | Composite score 0.2750 below baseline | 0.2750 |
| R8 | correlation | Round 8 research on correlation | Composite score 0.1693 below baseline | 0.1693 |
| R9 | entry_timing | Monte Carlo (10k paths, 10yr, fat-tail empirical resampling): Combined strategy has 99.1% probability of positive returns over 10yr (vs B&H 94.0%). P(beat B&H) = 100.0%. P(lose >20%) = 0.1% (vs B&H 3.... | Composite score 0.2750 below baseline | 0.2750 |
| R9 | entry_timing | Monte Carlo (10k paths, 10yr, fat-tail empirical resampling): Combined strategy has 99.1% probability of positive returns over 10yr (vs B&H 94.0%). P(beat B&H) = 100.0%. P(lose >20%) = 0.1% (vs B&H 3.... | Composite score 0.2750 below baseline | 0.2750 |
| R10 | strike_selection | Round 10 research on strike_selection | Composite score 0.1400 below baseline | 0.1400 |
| Agent | 方法 | 成功率 | 提出數 | 確認數 | 樣本數 |
|---|---|---|---|---|---|
| risk-auditor | validation_framework | 0.00% | 12 | 0 | 12 |
| researcher | volatility_structure | 0.00% | 2 | 0 | 2 |
| researcher | flow_sentiment | 0.00% | 1 | 0 | 1 |
| researcher | ml_volatility | 0.00% | 1 | 0 | 1 |
| researcher | cross_asset_momentum | 0.00% | 1 | 0 | 1 |
| researcher | portfolio_construction | 0.00% | 1 | 0 | 1 |
| researcher | intraday_options | 0.00% | 1 | 0 | 1 |
| researcher | calendar_anomalies | 0.00% | 1 | 0 | 1 |
| researcher | options_selling | 0.00% | 1 | 0 | 1 |
| researcher | trend_following | 0.00% | 1 | 0 | 1 |
| researcher | cross_sectional_momentum | 0.00% | 1 | 0 | 1 |
| researcher | microstructure | 0.00% | 1 | 0 | 1 |
| researcher | technical_indicators | 0.00% | 9 | 0 | 9 |
| risk-auditor | risk_audit_monte_carlo | 0.00% | 1 | 0 | 1 |
| risk-auditor | risk_audit_worst_case | 0.00% | 1 | 0 | 1 |
| risk-auditor | risk_audit_historical | 0.00% | 1 | 0 | 1 |
| risk-auditor | risk_audit_final_verdict | 0.00% | 1 | 0 | 1 |
| risk-auditor | risk_audit_monte_carlo | 0.00% | 2 | 0 | 2 |
| risk-auditor | risk_audit_worst_case | 0.00% | 2 | 0 | 2 |
| risk-auditor | risk_audit_historical | 0.00% | 2 | 0 | 2 |
| risk-auditor | risk_audit_final_verdict | 0.00% | 2 | 0 | 2 |