--- run_id: 23 date: "2026-03-12 19:56:57 UTC" question: "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." total_hypotheses: 156 confirmed: 0 rejected: 0 pending: 156 rounds: 10 --- # 研究迴圈報告 — Run #23 > 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. > 產生時間：2026-03-12 19:56:57 UTC ## 執行摘要 本次研究迴圈共進行 **10** 輪，產生 **156** 個假說。 最高信心假說（confidence=0.9600）： > After 6 rounds of rigorous investigation, the honest verdict on RSI(2) for SPY is: it is a STATISTICALLY MARGINAL timing signal with ZERO alpha over buy-and-hold on Sharpe, but with a REAL drawdown-re... ## 各輪摘要 | 輪次 | Agent 數 | 假說數 | 持續時間 | Token 用量 | |------|----------|--------|----------|------------| | R1 | 3 | 22 | — | — | | R2 | 3 | 42 | — | — | | R3 | 3 | 18 | — | — | | R4 | 3 | 15 | — | — | | R5 | 3 | 24 | — | — | | R6 | 3 | 11 | — | — | | R7 | 3 | 9 | — | — | | R8 | 3 | 5 | — | — | | R9 | 3 | 5 | — | — | | R10 | 3 | 5 | — | — | ## 輪次評分（Keep/Discard Gate） | 輪次 | 綜合分數 | Sharpe | Drawdown | 樣本量 | 決策 | 改善% | |------|----------|--------|----------|--------|------|-------| | R1 | 0.2545 | 0.0000 | 0.5000 | 0.0000 | KEEP | — | | R2 | 0.2750 | 0.0000 | 0.5000 | 0.0000 | KEEP | 8.06% | | R2 | 0.2750 | 0.0000 | 0.5000 | 0.0000 | KEEP | 8.06% | | R3 | 0.2417 | 0.0000 | 0.5000 | 0.0000 | DISCARD | -12.11% | | R4 | 0.2300 | 0.0000 | 0.5000 | 0.0000 | DISCARD | -16.36% | | R5 | 0.2750 | 0.0000 | 0.5000 | 0.0000 | KEEP | 0.00% | | R5 | 0.2750 | 0.0000 | 0.5000 | 0.0000 | KEEP | 0.00% | | R6 | 0.2221 | 0.0000 | 0.5000 | 0.0000 | DISCARD | -19.24% | | R7 | 0.2750 | 0.0000 | 0.5000 | 0.0000 | KEEP | 0.00% | | R8 | 0.2750 | 0.0000 | 0.5000 | 0.0000 | KEEP | 0.00% | | R9 | 0.1932 | 0.0000 | 0.5000 | 0.0000 | DISCARD | -29.75% | | R10 | 0.2750 | 0.0000 | 0.5000 | 0.0000 | KEEP | 0.00% | ## 假說清單 ### Round 1 - ⏳ **[entry_exit_signals]** (confidence=0.8000) RSI(2) mean reversion (buy RSI<10, sell RSI>90) with tight stop loss (1%) and wide profit target (5%) is the best SPY entry/exit signal strategy - ⏳ **[exit_rules]** (confidence=0.7500) Tight stop losses (1-2%) paired with wide profit targets (3-5%) produce optimal risk-adjusted returns for mean reversion strategies on SPY - ⏳ **[entry_exit_signals]** (confidence=0.6500) Bollinger Band bounce strategy (buy below lower band, sell above upper) is a viable secondary signal for SPY with high win rate - ⏳ **[entry_exit_signals]** (confidence=0.7000) Trend-following strategies (MA crossover, MACD) underperform mean reversion strategies for SPY entry/exit timing - ⏳ **[validation]** (confidence=0.7000) RSI2 mean reversion strategies show positive Sharpe improvement from train to test period, suggesting genuine alpha rather than overfitting - ⏳ **[risk_framework]** (confidence=0.9500) Multiple testing inflation: testing N profit-target/stop-loss combinations inflates Type I error rate exponentially - ⏳ **[risk_framework]** (confidence=0.9300) Look-ahead bias in technical indicator construction and ML feature engineering - ⏳ **[risk_framework]** (confidence=0.8800) Survivorship and selection bias: SPY is the most successful equity index ETF in history - ⏳ **[risk_framework]** (confidence=0.9200) Regime dependence: 2010-2024 dominated by QE, low rates, and tech mega-cap concentration - ⏳ **[risk_framework]** (confidence=0.9000) Transaction cost sensitivity destroys edge for short-term SPY signals - ⏳ **[risk_framework]** (confidence=0.9100) Insufficient sample size for per-regime and conditional analysis - ⏳ **[risk_framework]** (confidence=0.8700) Sharpe ratio inflation from autocorrelated returns and annualization method - ⏳ **[risk_framework]** (confidence=0.8900) Stop-loss whipsaw in volatile markets erodes strategy edge - ⏳ **[risk_framework]** (confidence=0.9400) ML model risks: feature leakage, non-stationarity, and overfitting to noise - ⏳ **[risk_framework]** (confidence=0.9100) Profit target optimization is the most dangerous form of curve-fitting in short-term strategies - ⏳ **[risk_framework]** (confidence=0.8500) Hourly bar aggregation introduces hidden assumptions about trading session and data alignment - ⏳ **[risk_framework]** (confidence=0.9500) Risk framework summary: minimum acceptable thresholds for strategy validation - ⏳ **[mean_reversion_rsi2]** (confidence=0.8200) RSI(2) Mean Reversion with ATR-based PT/SL: Buy SPY when RSI(2) < 10 (extreme oversold), exit at 2x ATR profit target or 1.5x ATR stop loss, with max 10-day holding period. Connors RSI(2) is a well-documented short-term mean reversion signal. Our empirical analysis shows RSI(2)<10 produces 0.74% avg... - ⏳ **[multi_timeframe_trend_reversion]** (confidence=0.7200) Multi-Timeframe Trend + Mean Reversion: Use daily SMA(50)>SMA(200) as trend filter (uptrend regime), then enter on hourly RSI(14)<30 oversold dips within the uptrend. Exit at 1% PT or 0.5% SL intraday, or hold overnight if unrealized gain > 0.3%. This combines trend-following regime detection with s... - ⏳ **[ml_random_forest_signals]** (confidence=0.6500) Random Forest Classifier on Technical Feature Set: Train a Random Forest model using 15+ technical indicators as features to predict next-day SPY direction (up/down). Features include RSI(2), RSI(14), MACD histogram, Bollinger %B, ADX, volume ratio, 5-day return Z-score, VIX level (if available), da... - ⏳ **[volume_confirmed_mean_reversion]** (confidence=0.7000) Volume-Weighted Mean Reversion: Buy SPY when (1) price closes below lower Bollinger Band(20,2), (2) volume is > 1.5x its 20-day average (panic selling confirmation), and (3) RSI(14) < 35. This triple-filter identifies high-conviction mean reversion opportunities where institutional capitulation crea... - ⏳ **[trend_following_adaptive_exit]** (confidence=0.5800) Adaptive Trend Following with MACD + ADX Filter and Volatility-Adjusted Exits: Enter long when MACD crosses above signal line AND ADX > 20 (confirming trend strength). Use ATR-based trailing stop (3x ATR trail) instead of fixed PT. This captures extended trending moves in SPY that fixed PT strategie... ### Round 2 - ⏳ **[technical_indicators]** (confidence=0.8500) Plain RSI(2)<10 remains the strongest mean-reversion signal for SPY; advanced composite indicators (Connors RSI, Capitulation Score) do NOT outperform on a risk-adjusted basis when sample size is adequate - ⏳ **[technical_indicators]** (confidence=0.7000) VIX danger-zone filter (avoid VIX 20-35 and rising) modestly improves RSI(2) risk profile by cutting max drawdown nearly in half, at the cost of ~45% fewer trades and lower total return - ⏳ **[technical_indicators]** (confidence=0.7200) SMA(200) trend filter with RSI(2) provides a good risk-return tradeoff, outperforming KAMA filter due to more signals - ⏳ **[technical_indicators]** (confidence=0.7500) StochRSI + Keltner Channel combination generates adequate trade volume but underperforms plain RSI(2) on risk-adjusted returns - ⏳ **[technical_indicators]** (confidence=0.7800) Wider stop-loss (2%) combined with tighter profit target (3%) yields highest Sharpe for RSI(2) strategy, contrary to Round 1 finding that PT5%/SL1% was optimal - ⏳ **[technical_indicators]** (confidence=0.8000) Connors RSI is a poor replacement for simple RSI(2) on SPY daily timeframe — the composite smoothing destroys the short-term oversold signal - ⏳ **[risk_audit]** (confidence=0.9200) Deflated Sharpe Ratio: After correcting for ~204 configurations tested across R1+R2, the best Sharpe of 1.27-1.68 deflates to approximately 0.55-0.85, barely above the SPY buy-and-hold baseline of 0.5-0.7 - ⏳ **[risk_audit]** (confidence=0.9000) Sample size of ~100 trades over 25 years (4/year) produces unacceptably wide confidence intervals on Sharpe ratio, making any Sharpe estimate unreliable - ⏳ **[risk_audit]** (confidence=0.8800) Regime analysis reveals critical vulnerability: RSI(2) mean reversion likely suffered catastrophic drawdowns during 2008 GFC and early 2020 COVID crash when mean reversion failed for weeks - ⏳ **[risk_audit]** (confidence=0.8200) Transaction costs of 1-5 bps roundtrip have minimal impact on this strategy due to low trade frequency and high average return per trade, but the analysis should be verified - ⏳ **[risk_audit]** (confidence=0.8700) PT/SL parameter instability between R1 (PT5%/SL1% best) and R2 (PT3%/SL2% best) is a red flag for overfitting — the optimal exit parameters shifted materially between rounds - ⏳ **[risk_audit]** (confidence=0.8500) Look-ahead bias risk is LOW for this strategy but walk-forward implementation has a critical flaw: single 70/30 split is not true walk-forward validation - ⏳ **[risk_audit]** (confidence=0.9300) Survivorship bias is the MOST SERIOUS unaddressed risk: SPY is the ultimate survivor and any buy-the-dip strategy will look spectacular in hindsight on the asset that always recovered - ⏳ **[risk_audit]** (confidence=0.8600) The test-period Sharpe exceeding train-period Sharpe (+21% improvement) is anomalous and likely driven by the COVID crash creating perfect RSI(2) mean-reversion setups rather than genuine out-of-sample alpha - ⏳ **[risk_audit]** (confidence=0.9500) The CapScore indicator with Sharpe=69.78 on 2 trades exemplifies a data snooping failure pattern that must be systematically prevented in future rounds - ⏳ **[risk_audit]** (confidence=0.8500) Overall risk assessment: RSI(2) mean reversion on SPY shows genuine but modest alpha after adjustments — estimated true Sharpe 0.8-1.0, with survivorship bias as the primary unresolved concern - ⏳ **[advanced_indicators]** (confidence=0.8000) Connors RSI (CRSI) composite signal outperforms raw RSI(2) for SPY mean reversion: Buy when CRSI < 5, sell when CRSI > 95, with PT5%/SL1%. CRSI combines RSI(3) of price, RSI(3) of up/down streak length, and percentile rank of magnitude — capturing 3 orthogonal dimensions of oversold vs RSI(2)'s sing... - ⏳ **[advanced_indicators]** (confidence=0.7500) Stochastic RSI with Keltner Channel confirmation beats RSI(2) by adding volatility-envelope context: Buy when StochRSI(14,14,3,3) K-line < 0.05 AND price is below lower Keltner Channel (EMA20 - 2*ATR14). Exit at Keltner midline (EMA20) or PT5%/SL1%. This dual-filter selects oversold conditions that ... - ⏳ **[intermarket_filters]** (confidence=0.7800) VIX regime filter applied to RSI(2) signals substantially improves Sharpe by avoiding mean-reversion entries during volatility regime transitions: Only take RSI(2)<10 buy signals when VIX is either (a) below 20 (calm market, high mean-reversion probability) or (b) above 35 AND declining (panic peak,... - ⏳ **[adaptive_indicators]** (confidence=0.7200) Kaufman Adaptive Moving Average (KAMA) trend filter combined with RSI(2) entries outperforms fixed SMA(200) filter: Use KAMA(10,2,30) as the trend filter — only take RSI(2)<10 buys when price > KAMA. KAMA adapts its smoothing to market efficiency ratio, becoming fast in trends and slow in chop, prov... - ⏳ **[composite_signal]** (confidence=0.7600) RSI(2) enhanced with consecutive-down-day streak and volume surge creates a 'capitulation score' that isolates the highest-return mean reversion entries: Score = (100 - RSI2)/100 * streak_factor * volume_factor. Only enter when score > 0.8. This should produce fewer but much higher quality trades wi... - ⏳ **[technical_indicators]** (confidence=0.8500) Plain RSI(2)<10 remains the strongest mean-reversion signal for SPY; advanced composite indicators (Connors RSI, Capitulation Score) do NOT outperform on a risk-adjusted basis when sample size is adequate - ⏳ **[technical_indicators]** (confidence=0.7000) VIX danger-zone filter (avoid VIX 20-35 and rising) modestly improves RSI(2) risk profile by cutting max drawdown nearly in half, at the cost of ~45% fewer trades and lower total return - ⏳ **[technical_indicators]** (confidence=0.7200) SMA(200) trend filter with RSI(2) provides a good risk-return tradeoff, outperforming KAMA filter due to more signals - ⏳ **[technical_indicators]** (confidence=0.7500) StochRSI + Keltner Channel combination generates adequate trade volume but underperforms plain RSI(2) on risk-adjusted returns - ⏳ **[technical_indicators]** (confidence=0.7800) Wider stop-loss (2%) combined with tighter profit target (3%) yields highest Sharpe for RSI(2) strategy, contrary to Round 1 finding that PT5%/SL1% was optimal - ⏳ **[technical_indicators]** (confidence=0.8000) Connors RSI is a poor replacement for simple RSI(2) on SPY daily timeframe — the composite smoothing destroys the short-term oversold signal - ⏳ **[risk_audit]** (confidence=0.9200) Deflated Sharpe Ratio: After correcting for ~204 configurations tested across R1+R2, the best Sharpe of 1.27-1.68 deflates to approximately 0.55-0.85, barely above the SPY buy-and-hold baseline of 0.5-0.7 - ⏳ **[risk_audit]** (confidence=0.9000) Sample size of ~100 trades over 25 years (4/year) produces unacceptably wide confidence intervals on Sharpe ratio, making any Sharpe estimate unreliable - ⏳ **[risk_audit]** (confidence=0.8800) Regime analysis reveals critical vulnerability: RSI(2) mean reversion likely suffered catastrophic drawdowns during 2008 GFC and early 2020 COVID crash when mean reversion failed for weeks - ⏳ **[risk_audit]** (confidence=0.8200) Transaction costs of 1-5 bps roundtrip have minimal impact on this strategy due to low trade frequency and high average return per trade, but the analysis should be verified - ⏳ **[risk_audit]** (confidence=0.8700) PT/SL parameter instability between R1 (PT5%/SL1% best) and R2 (PT3%/SL2% best) is a red flag for overfitting — the optimal exit parameters shifted materially between rounds - ⏳ **[risk_audit]** (confidence=0.8500) Look-ahead bias risk is LOW for this strategy but walk-forward implementation has a critical flaw: single 70/30 split is not true walk-forward validation - ⏳ **[risk_audit]** (confidence=0.9300) Survivorship bias is the MOST SERIOUS unaddressed risk: SPY is the ultimate survivor and any buy-the-dip strategy will look spectacular in hindsight on the asset that always recovered - ⏳ **[risk_audit]** (confidence=0.8600) The test-period Sharpe exceeding train-period Sharpe (+21% improvement) is anomalous and likely driven by the COVID crash creating perfect RSI(2) mean-reversion setups rather than genuine out-of-sample alpha - ⏳ **[risk_audit]** (confidence=0.9500) The CapScore indicator with Sharpe=69.78 on 2 trades exemplifies a data snooping failure pattern that must be systematically prevented in future rounds - ⏳ **[risk_audit]** (confidence=0.8500) Overall risk assessment: RSI(2) mean reversion on SPY shows genuine but modest alpha after adjustments — estimated true Sharpe 0.8-1.0, with survivorship bias as the primary unresolved concern - ⏳ **[advanced_indicators]** (confidence=0.8000) Connors RSI (CRSI) composite signal outperforms raw RSI(2) for SPY mean reversion: Buy when CRSI < 5, sell when CRSI > 95, with PT5%/SL1%. CRSI combines RSI(3) of price, RSI(3) of up/down streak length, and percentile rank of magnitude — capturing 3 orthogonal dimensions of oversold vs RSI(2)'s sing... - ⏳ **[advanced_indicators]** (confidence=0.7500) Stochastic RSI with Keltner Channel confirmation beats RSI(2) by adding volatility-envelope context: Buy when StochRSI(14,14,3,3) K-line < 0.05 AND price is below lower Keltner Channel (EMA20 - 2*ATR14). Exit at Keltner midline (EMA20) or PT5%/SL1%. This dual-filter selects oversold conditions that ... - ⏳ **[intermarket_filters]** (confidence=0.7800) VIX regime filter applied to RSI(2) signals substantially improves Sharpe by avoiding mean-reversion entries during volatility regime transitions: Only take RSI(2)<10 buy signals when VIX is either (a) below 20 (calm market, high mean-reversion probability) or (b) above 35 AND declining (panic peak,... - ⏳ **[adaptive_indicators]** (confidence=0.7200) Kaufman Adaptive Moving Average (KAMA) trend filter combined with RSI(2) entries outperforms fixed SMA(200) filter: Use KAMA(10,2,30) as the trend filter — only take RSI(2)<10 buys when price > KAMA. KAMA adapts its smoothing to market efficiency ratio, becoming fast in trends and slow in chop, prov... - ⏳ **[composite_signal]** (confidence=0.7600) RSI(2) enhanced with consecutive-down-day streak and volume surge creates a 'capitulation score' that isolates the highest-return mean reversion entries: Score = (100 - RSI2)/100 * streak_factor * volume_factor. Only enter when score > 0.8. This should produce fewer but much higher quality trades wi... ### Round 3 - ⏳ **[cross_asset_validation]** (confidence=0.8200) RSI(2) mean-reversion is NOT purely survivorship bias - it is profitable OOS on 7/8 tested assets (SPY, QQQ, IWM, EFA, EWJ, GLD, XLE), with only EEM showing negative OOS return (-10.18%). However, SPY has the highest Sharpe (1.90) by a wide margin, suggesting SPY-specific alpha exists beyond generic... - ⏳ **[cross_asset_validation]** (confidence=0.7200) The alternative PT5/SL1 configuration shows stronger cross-asset robustness than PT3/SL2, particularly on non-US equity assets. This suggests wider profit targets with tighter stops capture the fat-tailed recovery dynamics better across diverse assets. - ⏳ **[ml_alpha]** (confidence=0.7000) RandomForest ML filtering of RSI(2) signals improves win rate from 52.2% to 60.9% but filters out 90.7% of trades, resulting in only 23 OOS trades with lower total return (8.66% vs 47.23%). The ML model is too conservative to be practically useful at the 0.60 probability threshold, and zero trades p... - ⏳ **[ml_alpha]** (confidence=0.6800) Feature importance analysis reveals ATR%, distance from SMA200, and recent momentum (5d/10d returns) are the most predictive features for RSI(2) trade outcomes - more so than VIX level or RSI(14). This suggests that volatility regime and trend context matter more than the oversold depth itself. - ⏳ **[risk_audit]** (confidence=0.9300) XLE Sharpe=2.52 and GLD Sharpe=2.55 under PT5/SL1 are statistically unreliable due to extremely low trade counts — likely fewer than 20-25 OOS trades each, making these Sharpe estimates noise rather than signal - ⏳ **[risk_audit]** (confidence=0.9000) SPY Sharpe discrepancy across rounds (R1=1.68, R2=1.27, R3=1.90) reveals inconsistent test methodology — different train/test splits, different RSI(2) thresholds, or different Sharpe annualization are producing incomparable numbers - ⏳ **[risk_audit]** (confidence=0.9100) Cross-asset sample sizes are critically low for EFA, EEM, EWJ, and IWM — these assets likely have fewer than 30 RSI(2)<10 events in the OOS period, making any statistical inference unreliable - ⏳ **[risk_audit]** (confidence=0.8800) The Random Forest ML model with 12 features trained on ~138 RSI(2)<10 samples is massively overparameterized and the win rate improvement from 52.2% to 60.9% is almost certainly overfitting, despite walk-forward structure - ⏳ **[risk_audit]** (confidence=0.8500) RSI(2) alpha is fundamentally 'buy the dip in uptrending assets with structural mean-reversion' — and the key question is whether this edge persists post-2022 as interest rates changed the regime from ZIRP/QE to restrictive policy - ⏳ **[risk_audit]** (confidence=0.8700) Transaction cost impact is negligible at 4 trades/year, but the real cost is opportunity cost — sitting in cash 95% of the time while SPY earns ~10% annually means the strategy must significantly outperform during its brief deployment periods to justify its existence - ⏳ **[risk_audit]** (confidence=0.8600) The cross-asset validation's 'PASS' verdict (7/8 profitable) masks a gradient of edge quality that, when properly adjusted for sample size and trend beta, reduces to 'RSI(2) adds marginal value in trending markets' — a much weaker conclusion than presented - ⏳ **[risk_audit]** (confidence=0.8400) The PT/SL parameter instability worsens in R3: PT5/SL1 outperforms PT3/SL2 on non-US assets while PT3/SL2 outperforms on SPY — this means there is no single optimal exit configuration, and the 'best' parameters are asset-dependent and likely period-dependent - ⏳ **[risk_audit]** (confidence=0.8500) Overall Round 3 risk assessment: Cross-asset validation provides partial but not conclusive evidence against survivorship bias. ML filtering is a dead end. The strategy's real-world utility is as an overlay or tactical signal, not a standalone system. - ⏳ **[cross_asset_validation]** (confidence=0.9000) Cross-asset RSI(2) mean reversion validation: Test RSI(2)<10 with PT3%/SL2% on QQQ, IWM, EFA, EEM, and EWJ (Nikkei proxy) to determine if the signal is a genuine short-term mean-reversion anomaly or SPY survivorship bias. If RSI(2) achieves Sharpe>0.5 on at least 3 of 5 non-SPY assets, the signal is... - ⏳ **[ml_ensemble]** (confidence=0.6500) Random Forest ensemble combines RSI(2) with 10+ features to produce a probability-calibrated entry signal: Only enter when RF P(5-day-return > 1%) > 0.60, replacing the binary RSI(2)<10 threshold. Walk-forward trained with expanding window, retrained quarterly. Expected: fewer but higher-quality ent... - ⏳ **[pairs_trading]** (confidence=0.7000) SPY-QQQ pairs mean reversion: When the 5-day rolling SPY/QQQ return spread exceeds 2 standard deviations (z-score), go long the underperformer and short the outperformer. This is a market-neutral strategy that hedges away directional risk and addresses the survivorship bias concern since alpha comes... - ⏳ **[ml_standalone]** (confidence=0.5500) Gradient Boosting classifier trained on a rich feature set predicts next-5-day SPY direction with >55% accuracy on walk-forward OOS data, generating a standalone ML signal with Sharpe > 0.5. Use sklearn GradientBoostingClassifier since XGBoost is unavailable. - ⏳ **[sentiment_alternative_data]** (confidence=0.6000) Put/Call ratio from options data as a contrarian sentiment indicator: Buy SPY when equity put/call ratio > 1.2 (extreme fear) AND RSI(2) < 15, combining sentiment extreme with price extreme. This dual-filter should isolate genuine capitulation events with 70%+ win rate. ### Round 4 - ⏳ **[risk_audit]** (confidence=0.8800) RSI(2)<20 OOS Sharpe=0.885 is genuinely good but NOT exceptional — SPY buy-and-hold annualized Sharpe is ~0.5-0.7 depending on period, so the strategy outperforms by only 0.2-0.4 Sharpe units after accounting for time-in-market opportunity cost - ⏳ **[risk_audit]** (confidence=0.9000) Std=0.395 across 5 windows with per-window Sharpe [0.492, 0.325, 1.277, 1.182, 1.148] reveals a concerning bimodal pattern — the first two windows (pre-2015) dramatically underperform the last three (post-2013), suggesting a regime shift rather than a stable edge - ⏳ **[risk_audit]** (confidence=0.9300) The RSI(2)<10 collapse from Sharpe 1.2-1.9 (single split) to 0.011 (WFO) is the MOST IMPORTANT finding of Round 4 — it retroactively invalidates all specific Sharpe estimates from Rounds 1-3 while preserving the directional insight that RSI(2) has mean-reversion edge - ⏳ **[risk_audit]** (confidence=0.8700) 410 total OOS trades (82 per window * 5 windows) finally reaches statistical significance — the 95% CI on Sharpe 0.885 is approximately [0.60, 1.17], which excludes both zero and buy-and-hold Sharpe of ~0.46 - ⏳ **[risk_audit]** (confidence=0.9100) Fixed PT3%/SL2% exits were NOT validated out-of-sample — they were hard-coded from Round 2's single-split optimization, making them a hidden in-sample parameter that biases the WFO results upward - ⏳ **[risk_audit]** (confidence=0.8500) Window 2 (OOS 2007-2012, Sharpe=0.325) is the closest proxy for 'regime where strategy fails' — it does NOT fail catastrophically (no negative Sharpe) but delivers sub-par returns that would test psychological commitment to the strategy - ⏳ **[risk_audit]** (confidence=0.9200) The WFO-optimized threshold (Sharpe=0.089) versus fixed RSI(2)<20 (Sharpe=0.885) confirms a fundamental principle: in low-dimensional parameter spaces with limited data, fixed conservative parameters dominate adaptive optimization — this should guide ALL future parameter choices - ⏳ **[risk_audit]** (confidence=0.8400) The Round 4 Sharpe annualization method (sqrt(252/avg_hold_days)) may inflate results by 20-40% — the per-trade return distribution should be examined for autocorrelation and fat tails that violate the IID assumption - ⏳ **[risk_audit]** (confidence=0.8600) Fat-tail dependency test STILL not performed after 4 rounds of audit requests — removing the top 5% of trades (~21 of 410) could plausibly reduce Sharpe from 0.885 to below 0.5 given the asymmetric PT3%/SL2% structure - ⏳ **[risk_audit]** (confidence=0.8700) Overall Round 4 risk assessment scorecard: 4 PASS, 2 FAIL, 3 PARTIAL — material improvement from Round 3 (2 PASS, 3 FAIL, 3 PARTIAL) driven by WFO implementation, but critical gaps remain in exit validation and fat-tail testing - ⏳ **[entry_timing]** (confidence=0.8800) Relaxed RSI(2) threshold ladder: Test RSI(2)<15, <20, <25 alongside <10 baseline. RSI(2)<10 yields ~4 trades/year; RSI(2)<20 should yield ~20-30 trades/year based on Round 1 data showing RSI(2)<15 has n=562 vs n=381 for <10 (48% more trades with only 8% lower avg return: 0.68% vs 0.74%). Crucially, ... - ⏳ **[entry_timing]** (confidence=0.8200) Multi-signal OR-gate: Combine RSI(2)<10, Bollinger Band lower touch (close < BB(20,2) lower), and volume spike (volume > 2x SMA20) as INDEPENDENT entry triggers (any one triggers a trade, not all required). Each signal fires on different market conditions — RSI on momentum, BB on deviation, volume o... - ⏳ **[entry_timing]** (confidence=0.8500) Multi-asset RSI(2) rotation across SPY+QQQ+IWM simultaneously: Instead of waiting for RSI(2)<10 on SPY alone, trade RSI(2)<10 signals on ANY of the three major US equity ETFs. This triples the signal universe while diversifying across market-cap segments. Round 3 confirmed RSI(2) works on 7/8 ETFs t... - ⏳ **[entry_timing]** (confidence=0.8000) Regime-adaptive RSI threshold: Use VIX level to dynamically set the RSI(2) entry threshold — RSI(2)<20 in low-VIX (VIX<18, calm markets), RSI(2)<10 in medium-VIX (18-30), RSI(2)<5 in high-VIX (>30). Low-VIX markets have smaller price swings so RSI(2) rarely reaches <10, yet 2-day pullbacks in calm m... - ⏳ **[entry_timing]** (confidence=0.6800) Day-of-week + RSI(2) interaction: RSI(2)<15 entries on Monday and Tuesday have historically higher win rates than Wednesday-Friday entries, because weekend/Monday panic selling creates deeper but more transient oversold conditions that mean-revert by mid-week. By using RSI(2)<15 (relaxed from <10) b... ### Round 5 - ⏳ **[risk_audit]** (confidence=0.9500) RSI exit has look-ahead bias: exit executes at same close where RSI signal fires. Fixing to next-day exit drops RSI>50 Sharpe from 2.282 to 1.256 (45% drop). RSI>70 drops from 2.001 to 1.711 (14% drop). The 2.28 Sharpe is inflated by this bias. - ⏳ **[risk_audit]** (confidence=0.9000) WFO windows have 20-60% overlap in OOS periods, artificially reducing Sharpe dispersion. Overlapping std=0.987 vs non-overlapping std=0.103. The reported std=0.317 understates true uncertainty. - ⏳ **[risk_audit]** (confidence=0.7500) Per-trade Sharpe annualization is approximately correct for this strategy but daily equity curve Sharpe provides a more conservative and realistic estimate. The annualization method is not the primary issue — look-ahead bias is. - ⏳ **[risk_audit]** (confidence=0.9200) Fat tail dependency remains severe even after fixing look-ahead bias. Strategy performance is concentrated in a small number of large winners. - ⏳ **[risk_audit]** (confidence=0.8000) Trade overlap is not checked — the script allows entering new positions while existing positions are still open, effectively running a portfolio not a single-position strategy. This inflates trade count and may inflate Sharpe. - ⏳ **[risk_audit]** (confidence=0.9300) After correcting look-ahead bias, using non-overlapping WFO windows, and adding 10bp round-trip transaction costs, RSI(2)<20 entry + RSI(2)>50 exit Sharpe drops from claimed 2.282 to approximately 0.79. The strategy still shows positive alpha but is not a 2+ Sharpe system. - ⏳ **[exit_rules]** (confidence=0.8800) Time-based exit (N-day holding period, NO PT/SL) is the most robust exit rule for RSI(2)<20 mean reversion: exit at close of day N after entry, where N in {3,5,7,10,15,20}. This eliminates PT/SL parameter sensitivity entirely. Academic literature strongly supports fixed holding periods for mean-reve... - ⏳ **[exit_rules]** (confidence=0.8300) ATR-normalized exits (PT = K1 * ATR(14), SL = K2 * ATR(14)) adapt to volatility regime and should produce more stable OOS Sharpe than fixed percentage exits across the 5 WFO windows — specifically, they should reduce the Sharpe dispersion from std=0.395 to std<0.25 by auto-scaling exits to current m... - ⏳ **[exit_rules]** (confidence=0.8000) RSI-based exit (exit when RSI(2) crosses above threshold T, where T in {50, 60, 70, 80, 90}) is the theoretically purest mean-reversion exit because it exits when the SIGNAL itself says the mean reversion is complete, rather than using an arbitrary price target. Connors' original research recommends... - ⏳ **[exit_rules]** (confidence=0.7800) Trailing stop exit (initial stop = 1.5% below entry, trail up by ratcheting stop to max(current_stop, highest_close - 1.5*ATR(14))) combined with max_hold=15 days captures more upside than fixed PT by letting winners run while protecting gains. This specifically addresses the fat-tail dependency con... - ⏳ **[exit_rules]** (confidence=0.8200) Hybrid exit combining time limit + trailing stop + hard SL floor: Enter RSI(2)<20, set hard SL at 3% below entry (catastrophe floor), use trailing stop of 1.5*ATR after any gain, force exit at day 10. This 3-layer exit addresses ALL identified risks simultaneously: the SL floor prevents catastrophic... - ⏳ **[exit_rules]** (confidence=0.9200) Fat-tail dependency audit (REQUIRED by R4 auditor, unfulfilled for 4 rounds): Compute Sharpe after removing top 5% and top 10% of trades by P&L, AND after removing all trades that hit PT within 2 days ('lucky' trades). Apply to BOTH the existing PT3%/SL2% setup and the new exit rules tested in this ... - ⏳ **[risk_audit]** (confidence=0.9500) RSI exit has look-ahead bias: exit executes at same close where RSI signal fires. Fixing to next-day exit drops RSI>50 Sharpe from 2.282 to 1.256 (45% drop). RSI>70 drops from 2.001 to 1.711 (14% drop). The 2.28 Sharpe is inflated by this bias. - ⏳ **[risk_audit]** (confidence=0.9000) WFO windows have 20-60% overlap in OOS periods, artificially reducing Sharpe dispersion. Overlapping std=0.987 vs non-overlapping std=0.103. The reported std=0.317 understates true uncertainty. - ⏳ **[risk_audit]** (confidence=0.7500) Per-trade Sharpe annualization is approximately correct for this strategy but daily equity curve Sharpe provides a more conservative and realistic estimate. The annualization method is not the primary issue — look-ahead bias is. - ⏳ **[risk_audit]** (confidence=0.9200) Fat tail dependency remains severe even after fixing look-ahead bias. Strategy performance is concentrated in a small number of large winners. - ⏳ **[risk_audit]** (confidence=0.8000) Trade overlap is not checked — the script allows entering new positions while existing positions are still open, effectively running a portfolio not a single-position strategy. This inflates trade count and may inflate Sharpe. - ⏳ **[risk_audit]** (confidence=0.9300) After correcting look-ahead bias, using non-overlapping WFO windows, and adding 10bp round-trip transaction costs, RSI(2)<20 entry + RSI(2)>50 exit Sharpe drops from claimed 2.282 to approximately 0.79. The strategy still shows positive alpha but is not a 2+ Sharpe system. - ⏳ **[exit_rules]** (confidence=0.8800) Time-based exit (N-day holding period, NO PT/SL) is the most robust exit rule for RSI(2)<20 mean reversion: exit at close of day N after entry, where N in {3,5,7,10,15,20}. This eliminates PT/SL parameter sensitivity entirely. Academic literature strongly supports fixed holding periods for mean-reve... - ⏳ **[exit_rules]** (confidence=0.8300) ATR-normalized exits (PT = K1 * ATR(14), SL = K2 * ATR(14)) adapt to volatility regime and should produce more stable OOS Sharpe than fixed percentage exits across the 5 WFO windows — specifically, they should reduce the Sharpe dispersion from std=0.395 to std<0.25 by auto-scaling exits to current m... - ⏳ **[exit_rules]** (confidence=0.8000) RSI-based exit (exit when RSI(2) crosses above threshold T, where T in {50, 60, 70, 80, 90}) is the theoretically purest mean-reversion exit because it exits when the SIGNAL itself says the mean reversion is complete, rather than using an arbitrary price target. Connors' original research recommends... - ⏳ **[exit_rules]** (confidence=0.7800) Trailing stop exit (initial stop = 1.5% below entry, trail up by ratcheting stop to max(current_stop, highest_close - 1.5*ATR(14))) combined with max_hold=15 days captures more upside than fixed PT by letting winners run while protecting gains. This specifically addresses the fat-tail dependency con... - ⏳ **[exit_rules]** (confidence=0.8200) Hybrid exit combining time limit + trailing stop + hard SL floor: Enter RSI(2)<20, set hard SL at 3% below entry (catastrophe floor), use trailing stop of 1.5*ATR after any gain, force exit at day 10. This 3-layer exit addresses ALL identified risks simultaneously: the SL floor prevents catastrophic... - ⏳ **[exit_rules]** (confidence=0.9200) Fat-tail dependency audit (REQUIRED by R4 auditor, unfulfilled for 4 rounds): Compute Sharpe after removing top 5% and top 10% of trades by P&L, AND after removing all trades that hit PT within 2 days ('lucky' trades). Apply to BOTH the existing PT3%/SL2% setup and the new exit rules tested in this ... ### Round 6 - ⏳ **[risk_audit]** (confidence=0.9500) RSI(2) vs B&H Sharpe difference is statistically ZERO. Block bootstrap (10,000 iterations, 21-day blocks) yields 95% CI of [-0.24, +0.27] for the Sharpe difference, with only 56.7% probability that RSI beats B&H. On truly non-overlapping windows (7yr each, no shared OOS data), avg RSI Sharpe=0.573 v... - ⏳ **[risk_audit]** (confidence=0.9200) RSI(2)'s drawdown advantage is REAL and produces superior Calmar and Sortino ratios in crisis periods. Avg RSI Calmar=0.389 vs B&H=0.282 (+38%), avg RSI Sortino=0.533 vs B&H=0.703 (B&H still wins Sortino due to upside capture). The drawdown advantage is valuable but ONLY for investors who prioritize... - ⏳ **[risk_audit]** (confidence=0.9300) LEVERAGE OVERLAY IS DANGEROUS: RSI(2)-conditioned leverage (1.5x on RSI<20, 0.75x on RSI>80) achieves Sharpe 0.62-0.67 (better than B&H's 0.51) but with max DD of -44% to -47%. During the GFC, the overlay draws down -45% vs B&H -52%. The leverage 'buys the dip' INTO crashes — RSI<20 fires repeatedly... - ⏳ **[risk_audit]** (confidence=0.8500) Seasonal (Sell-in-May + Turn-of-Month) strategy shows modest risk-adjusted improvement over B&H in bear markets but underperforms in bull markets, similar to RSI(2). Full-period Seasonal Sharpe=0.549 vs B&H=0.513. It is NOT a reliable standalone strategy post-2014. - ⏳ **[risk_audit]** (confidence=0.9600) After 6 rounds of rigorous investigation, the honest verdict on RSI(2) for SPY is: it is a STATISTICALLY MARGINAL timing signal with ZERO alpha over buy-and-hold on Sharpe, but with a REAL drawdown-reduction benefit (avg DD -22% vs -40%) that makes it useful ONLY as a position-sizing overlay for dra... - ⏳ **[risk_audit]** (confidence=0.8800) The proposed multi-signal ensemble (RSI overlay + momentum + seasonal + VIX) faces a fundamental correlation problem: ALL four signals are LONG SPY, creating a false sense of diversification. Estimated portfolio Sharpe benefit of sqrt(N) diversification requires LOW correlation between signals, but ... - ⏳ **[strategy_pivot]** (confidence=0.8200) RSI(2) as a LEVERAGE OVERLAY on buy-and-hold: Stay 100% invested in SPY at all times (capture full B&H return). When RSI(2)<20, increase position to 150% (1.5x leverage via futures or margin). When RSI(2)>80, reduce to 75% (de-risk when overbought). This way we capture the B&H Sharpe of 0.68-0.85 as... - ⏳ **[strategy_pivot]** (confidence=0.7800) Dual-momentum regime switching: Use 12-1 month momentum (Antonacci 2012) as primary signal for SPY vs T-Bills allocation, with RSI(2) as secondary timing within the SPY-allocated periods. When 12-1mo SPY return > 12-1mo T-Bill return, invest in SPY and use RSI(2) for position sizing overlay. When mo... - ⏳ **[strategy_pivot]** (confidence=0.7200) Seasonal + turn-of-month strategy: Invest in SPY only during historically positive seasonal windows — November through April ('Sell in May'), plus the last 3 and first 3 trading days of each month ('Turn of Month' effect). Cash during May-October non-TOM days. This exploits two of the most robust ca... - ⏳ **[strategy_pivot]** (confidence=0.7000) VIX mean-reversion strategy: When VIX spikes above its 90th percentile (historically ~25-30), go long SPY and hold for 10 trading days. VIX mean-reversion is one of the most robust phenomena in finance — extreme fear reverts faster than extreme complacency. This is similar to RSI(2)<20 but uses an E... - ⏳ **[strategy_pivot]** (confidence=0.7500) Multi-signal ensemble portfolio: Allocate capital equally across 3-4 uncorrelated timing signals, each running independently on SPY. Signal 1: RSI(2) leverage overlay (always invested, tilt on RSI). Signal 2: 12-1 month momentum (in/out monthly). Signal 3: Seasonal/TOM calendar. Signal 4: VIX mean-r... ### Round 7 - ⏳ **[risk_audit]** (confidence=0.9300) VolTarget Sharpe advantage over B&H is NOT statistically significant. Block bootstrap (10,000 iterations, 21-day blocks) 95% CI: [-0.1611, 0.2616]. P(VolTarget > B&H) = 0.6885. The CI includes zero, meaning the 0.07 Sharpe improvement cannot be distinguished from noise at 95% confidence. However, th... - ⏳ **[risk_audit]** (confidence=0.9200) The VolTarget CAGR sacrifice (7.73% vs 11.23% = -3.5% annual) is JUSTIFIED for risk-constrained investors. Calmar ratio improves from 0.204 to 0.286 (+40%). Ulcer Performance Index improves from 0.935 to 1.08. The question is investor utility function: (a) Return-maximizer → B&H always wins (10.7% C... - ⏳ **[risk_audit]** (confidence=0.9000) SMA(200) whipsaws are manageable but real. 6.3 signal changes per year, 50.0% reverse within 5 days (pure cost drag). Despite whipsaws, SMA(200) achieves the best Calmar (0.357) because the whipsaw cost (3.2% CAGR) is far outweighed by the drawdown avoidance (max DD -22.4% vs -55.2%). Main risk: a c... - ⏳ **[risk_audit]** (confidence=0.9100) ALL overlays show strong regime dependency. 2003-2009 (crash): best=Momentum_12_1 (Sharpe 0.683), worst=RSI2_Overlay (Sharpe 0.3). 2010-2019 (bull): best=B&H (Sharpe 0.933), worst=RSI2_Overlay (Sharpe 0.532). 2020-2025 (mixed): best=SMA200 (Sharpe 0.907), worst=RSI2_Overlay (Sharpe 0.705). In the 20... - ⏳ **[risk_audit]** (confidence=0.9000) Cash drag is the primary cost of binary (0/100) strategies. Momentum spends 14.9% of time in cash, losing 1.66% CAGR vs B&H. SMA200 spends 18.9% in cash, losing 3.24% CAGR. This is why VolTarget (continuous sizing, never fully in cash) achieves higher CAGR than SMA200/Momentum despite similar Sharpe... - ⏳ **[risk_audit]** (confidence=0.9500) HONEST VERDICT after 7 rounds: There is NO free lunch in SPY timing. Every overlay sacrifices 1.8-3.4% annual CAGR to reduce drawdown by 21-33 percentage points. The Sharpe improvements (0.02-0.07) are NOT statistically significant at 95% confidence. The REAL value is in drawdown reduction and behav... - ⏳ **[final_verdict]** (confidence=0.9500) RSI(2) as position sizing overlay reduces Sharpe by 25% vs B&H — definitively harmful - ⏳ **[final_verdict]** (confidence=0.8500) VolTarget 10% is best risk-adjusted overlay but improvement is not statistically significant - ⏳ **[final_verdict]** (confidence=0.9000) SPY daily timing alpha is near-zero — this is a well-arbitraged market ### Round 8 - ⏳ **[risk_audit]** (confidence=0.9500) Hourly signals are catastrophically worse than daily — noise dominates - ⏳ **[risk_audit]** (confidence=0.9200) SPY timing alpha does not exist at any tested timeframe - ⏳ **[risk_audit]** (confidence=0.8800) Research should pivot to options or structural risk premia - ⏳ **[final_conclusion]** (confidence=0.9300) 8-round research conclusion: SPY equity timing has no exploitable alpha - ⏳ **[final_conclusion]** (confidence=0.8500) Only actionable finding: VolTarget or SMA(200) for drawdown reduction ### Round 9 - ⏳ **[final_risk_assessment]** (confidence=0.9000) Combined Soft strategy provides 4x drawdown reduction at cost of 3.56% CAGR — appropriate for risk-averse investors - ⏳ **[final_risk_assessment]** (confidence=0.9200) B&H remains optimal for long-horizon investors who can tolerate 50%+ drawdowns - ⏳ **[methodology]** (confidence=0.9500) Research integrity validated: Sharpe degraded from 1.68 (R1) to 0.634 (R9) through systematic bias correction - ⏳ **[actionable_conclusions]** (confidence=0.9200) 9-round research delivers three actionable conclusions for SPY traders - ⏳ **[future_directions]** (confidence=0.8800) Future research should target structural risk premia, not equity timing ### Round 10 - ⏳ **[final_results]** (confidence=0.9200) Final strategy comparison table (all biases corrected, 4-window non-overlapping WFO + holdout) - ⏳ **[sign_off]** (confidence=0.9300) Final risk sign-off: research conclusions are robust and honestly reported - ⏳ **[final_verdict]** (confidence=0.9500) FINAL VERDICT: SPY equity timing has no exploitable alpha at daily or hourly frequency - ⏳ **[recommendation]** (confidence=0.9000) Recommended strategy depends on risk tolerance - ⏳ **[methodology]** (confidence=0.9200) Key methodological lesson: backtest Sharpe inflation averages 2.7x from standard biases ## 負面結果（已排除的方向） - **R3** [expiration_choice] RSI(2) mean-reversion is NOT purely survivorship bias - it is profitable OOS on 7/8 tested assets (SPY, QQQ, IWM, EFA, EWJ, GLD, XLE), with only EEM s... 失敗原因：Composite score 0.2417 below baseline 分數：0.2417，交易數：0 - **R4** [position_sizing] RSI(2)<20 OOS Sharpe=0.885 is genuinely good but NOT exceptional — SPY buy-and-hold annualized Sharpe is ~0.5-0.7 depending on period, so the strategy... 失敗原因：Composite score 0.2300 below baseline 分數：0.2300，交易數：0 - **R6** [regime_filter] RSI(2) vs B&H Sharpe difference is statistically ZERO. Block bootstrap (10,000 iterations, 21-day blocks) yields 95% CI of [-0.24, +0.27] for the Shar... 失敗原因：Composite score 0.2221 below baseline 分數：0.2221，交易數：0 - **R9** [entry_timing] Combined Soft strategy provides 4x drawdown reduction at cost of 3.56% CAGR — appropriate for risk-averse investors; B&H remains optimal for long-hori... 失敗原因：Composite score 0.1932 below baseline 分數：0.1932，交易數：0 ## 方法評分 | Agent | 方法 | 成功率 | 提出數 | 確認數 | 樣本數 | |-------|------|--------|--------|--------|--------| | optimizer | entry_exit_signals | 0.00% | 3 | 0 | 3 | | researcher | final_verdict | 0.00% | 3 | 0 | 3 | | optimizer | validation | 0.00% | 1 | 0 | 1 | | risk-auditor | risk_framework | 0.00% | 12 | 0 | 12 | | researcher | mean_reversion_rsi2 | 0.00% | 1 | 0 | 1 | | researcher | multi_timeframe_trend_reversion | 0.00% | 1 | 0 | 1 | | researcher | ml_random_forest_signals | 0.00% | 1 | 0 | 1 | | researcher | volume_confirmed_mean_reversion | 0.00% | 1 | 0 | 1 | | researcher | trend_following_adaptive_exit | 0.00% | 1 | 0 | 1 | | optimizer | technical_indicators | 0.00% | 6 | 0 | 6 | | risk-auditor | risk_audit | 0.00% | 10 | 0 | 10 | | researcher | advanced_indicators | 0.00% | 2 | 0 | 2 | | researcher | intermarket_filters | 0.00% | 1 | 0 | 1 | | optimizer | exit_rules | 0.00% | 1 | 0 | 1 | | researcher | composite_signal | 0.00% | 1 | 0 | 1 | | optimizer | technical_indicators | 0.00% | 12 | 0 | 12 | | risk-auditor | risk_audit | 0.00% | 20 | 0 | 20 | | researcher | advanced_indicators | 0.00% | 4 | 0 | 4 | | researcher | intermarket_filters | 0.00% | 2 | 0 | 2 | | researcher | adaptive_indicators | 0.00% | 2 | 0 | 2 | | researcher | composite_signal | 0.00% | 2 | 0 | 2 | | optimizer | cross_asset_validation | 0.00% | 2 | 0 | 2 | | optimizer | ml_alpha | 0.00% | 2 | 0 | 2 | | risk-auditor | risk_audit | 0.00% | 9 | 0 | 9 | | researcher | cross_asset_validation | 0.00% | 1 | 0 | 1 | | researcher | ml_ensemble | 0.00% | 1 | 0 | 1 | | researcher | pairs_trading | 0.00% | 1 | 0 | 1 | | researcher | ml_standalone | 0.00% | 1 | 0 | 1 | | researcher | sentiment_alternative_data | 0.00% | 1 | 0 | 1 | | risk-auditor | risk_audit | 0.00% | 10 | 0 | 10 | | researcher | entry_timing | 0.00% | 5 | 0 | 5 | | risk-auditor | risk_audit | 0.00% | 6 | 0 | 6 | | researcher | exit_rules | 0.00% | 6 | 0 | 6 | | risk-auditor | risk_audit | 0.00% | 12 | 0 | 12 | | researcher | exit_rules | 0.00% | 12 | 0 | 12 | | risk-auditor | risk_audit | 0.00% | 6 | 0 | 6 | | researcher | strategy_pivot | 0.00% | 5 | 0 | 5 | | risk-auditor | risk_audit | 0.00% | 6 | 0 | 6 | | researcher | methodology | 0.00% | 1 | 0 | 1 | | risk-auditor | risk_audit | 0.00% | 3 | 0 | 3 | | researcher | final_conclusion | 0.00% | 2 | 0 | 2 | | risk-auditor | final_risk_assessment | 0.00% | 2 | 0 | 2 | | risk-auditor | methodology | 0.00% | 1 | 0 | 1 | | researcher | actionable_conclusions | 0.00% | 1 | 0 | 1 | | researcher | future_directions | 0.00% | 1 | 0 | 1 | | optimizer | final_results | 0.00% | 1 | 0 | 1 | | risk-auditor | sign_off | 0.00% | 1 | 0 | 1 | | researcher | final_verdict | 0.00% | 1 | 0 | 1 | | researcher | recommendation | 0.00% | 1 | 0 | 1 | | researcher | adaptive_indicators | 0.00% | 1 | 0 | 1 | ## 建議與後續行動 - 156 個假說仍待驗證，可增加輪次或手動檢視。