Feature explorer
214 closed signals (97W / 58L / 59 expired) · baseline win rate 62.6% · prompt versions v21, v22, v23, v24
• Exploratory analysis on a small sample (~hundreds of closed signals, ~100 features). With this many comparisons, several features will look 'significant' by chance — treat rankings as hypotheses, not conclusions.
• Win rates exclude expired signals from the denominator (won/lost only); n_expired is reported separately.
• Citation win rate measures association between the LLM mentioning a feature and the outcome — not the feature's causal predictive power.
• Signals are not independent (overlapping market regimes, correlated assets), so confidence intervals are, if anything, too narrow.
• Tables are sorted by the Wilson lower bound, not the point estimate, to penalize tiny samples.
Tier 1 — win rate when the analyst cites a feature
| Feature | Cited | W / L / Exp | WR when cited | 95% CI | CI vs baseline |
|---|---|---|---|---|---|
| hurst_exp_10d | 85 | 45/16/24 | 73.8% | 61.6%–83.2% | |
| distance_5d_low | 8 | 6/0/2 | 100.0% | 61.0%–100.0% | |
| top_trader_pos_ratio_1h | 28 | 14/3/11 | 82.3% | 59.0%–93.8% | |
| top_trader_long_pct | 40 | 23/8/9 | 74.2% | 56.8%–86.3% | |
| binance_top_trader_pos_ratio_1h | 6 | 5/0/1 | 100.0% | 56.5%–100.0% | |
| fear_greed_25 | 5 | 5/0/0 | 100.0% | 56.5%–100.0% | |
| binance_oi_tls.top_trader_account_ratio_1h | 10 | 8/1/1 | 88.9% | 56.5%–98.0% | |
| mtf_alignment | 90 | 44/21/25 | 67.7% | 55.6%–77.8% | |
| adx_14_1h | 36 | 22/8/6 | 73.3% | 55.5%–85.8% | |
| adx_20 | 53 | 26/11/16 | 70.3% | 54.2%–82.5% | |
| rolling_z_20d | 117 | 55/30/32 | 64.7% | 54.1%–74.0% | |
| fear_greed_index | 97 | 45/26/26 | 63.4% | 51.8%–73.6% | |
| distance_from_extrema | 68 | 31/16/21 | 66.0% | 51.7%–77.8% | |
| top_trader_account_ratio_1h | 28 | 14/5/9 | 73.7% | 51.2%–88.2% | |
| gold_spot_24h_change_pct | 25 | 14/5/6 | 73.7% | 51.2%–88.2% | |
| binance_xagusdt_change_24h_pct | 6 | 4/0/2 | 100.0% | 51.0%–100.0% | |
| oi_change_24h_pct | 38 | 20/9/9 | 69.0% | 50.8%–82.7% | |
| fear_greed_12 | 10 | 8/2/0 | 80.0% | 49.0%–94.3% | |
| avg_daily_range_pct_7d | 13 | 6/1/6 | 85.7% | 48.7%–97.4% | |
| fear_greed_29 | 7 | 6/1/0 | 85.7% | 48.7%–97.4% | |
| oi_5m_z_24h | 25 | 12/5/8 | 70.6% | 46.9%–86.7% | |
| adx_1h | 17 | 9/3/5 | 75.0% | 46.8%–91.1% | |
| realized_vol_percentiles | 34 | 16/8/10 | 66.7% | 46.7%–82.0% | |
| mtf_alignment_full_bearish | 30 | 16/8/6 | 66.7% | 46.7%–82.0% | |
| delta_funding_rate | 18 | 7/2/9 | 77.8% | 45.3%–93.7% | |
| top_trader_positioning | 73 | 31/22/20 | 58.5% | 45.1%–70.7% | |
| binance_oi_tls.oi_15m_z_24h | 5 | 3/0/2 | 100.0% | 43.9%–100.0% | |
| ema_stack_4h | 12 | 5/1/6 | 83.3% | 43.6%–97.0% | |
| atr_pct_4h | 8 | 5/1/2 | 83.3% | 43.6%–97.0% | |
| binance_oi_tls | 20 | 9/4/7 | 69.2% | 42.4%–87.3% | |
| ema_stack_5m_15m_1h_4h | 8 | 6/2/0 | 75.0% | 40.9%–92.8% | |
| distance_from_5d_low | 13 | 7/3/3 | 70.0% | 39.7%–89.2% | |
| taker_buy_sell_ratio_1h | 45 | 17/13/15 | 56.7% | 39.2%–72.6% | |
| binance_oi_tls.top_trader_pos_ratio_1h | 13 | 8/4/1 | 66.7% | 39.1%–86.2% | |
| liquidation_pressure | 23 | 9/5/9 | 64.3% | 38.8%–83.7% | |
| recent_drishti_signals_same_regime | 17 | 10/6/1 | 62.5% | 38.6%–81.5% | |
| distance_from_extrema_5d_low | 7 | 4/1/2 | 80.0% | 37.5%–96.4% | |
| ema_stack_1h | 9 | 5/2/2 | 71.4% | 35.9%–91.8% | |
| liquidation_pressure.long_short_liq_ratio | 22 | 8/5/9 | 61.5% | 35.5%–82.3% | |
| adx_14 | 19 | 8/6/5 | 57.1% | 32.6%–78.6% | |
| rsi_5m | 12 | 5/3/4 | 62.5% | 30.6%–86.3% | |
| rv_24h | 7 | 3/1/3 | 75.0% | 30.1%–95.4% | |
| adx_15m | 7 | 3/1/3 | 75.0% | 30.1%–95.4% | |
| ema_stack_5m_15m | 6 | 3/1/2 | 75.0% | 30.1%–95.4% | |
| rsi_4h | 25 | 10/11/4 | 47.6% | 28.3%–67.6% | |
| calibration_short_wr | 14 | 5/4/5 | 55.6% | 26.7%–81.1% | |
| regime_v19.adx_20 | 9 | 5/4/0 | 55.6% | 26.7%–81.1% | |
| funding_z_score_30d | 35 | 10/13/12 | 43.5% | 25.6%–63.2% | |
| distance_from_extrema.5d_low | 8 | 4/3/1 | 57.1% | 25.1%–84.2% | |
| adx_14_4h | 23 | 7/8/8 | 46.7% | 24.8%–69.9% | |
| rsi_1h | 23 | 8/10/5 | 44.4% | 24.6%–66.3% | |
| daily_move | 7 | 3/2/2 | 60.0% | 23.1%–88.2% | |
| liquidation_pressure_4h | 21 | 6/8/7 | 42.9% | 21.4%–67.4% | |
| multi_timeframe_technicals | 6 | 2/1/3 | 66.7% | 20.8%–93.8% | |
| adx_4h | 5 | 2/1/2 | 66.7% | 20.8%–93.8% | |
| recent_signals_attribution | 15 | 4/5/6 | 44.4% | 18.9%–73.3% | |
| realized_vol_24h | 6 | 3/3/0 | 50.0% | 18.8%–81.2% | |
| top_trader_positioning.long_pct | 8 | 2/2/4 | 50.0% | 15.0%–85.0% | |
| binance_xagusdt_change_24h | 7 | 2/2/3 | 50.0% | 15.0%–85.0% | |
| long_short_liq_ratio_4h | 5 | 2/2/1 | 50.0% | 15.0%–85.0% | |
| this_asset_30d_wr | 5 | 2/2/1 | 50.0% | 15.0%–85.0% | |
| daily_move.avg_daily_range_pct_7d | 15 | 3/5/7 | 37.5% | 13.7%–69.4% | |
| distance_from_5d_high | 13 | 3/6/4 | 33.3% | 12.1%–64.6% | |
| total3_mcap_5d_chg | 8 | 2/3/3 | 40.0% | 11.8%–76.9% | |
| gold_token_premium_pct | 7 | 2/3/2 | 40.0% | 11.8%–76.9% | |
| atr_pct_1h | 6 | 2/3/1 | 40.0% | 11.8%–76.9% | |
| oi_15m_z_24h | 13 | 2/4/7 | 33.3% | 9.7%–70.0% | |
| binance_oi_tls.oi_5m_z_24h | 10 | 2/4/4 | 33.3% | 9.7%–70.0% | |
| macro.gold | 6 | 2/4/0 | 33.3% | 9.7%–70.0% | |
| regime_v19 | 5 | 1/1/3 | 50.0% | 9.4%–90.5% | |
| distance_from_extrema.5d_high | 10 | 1/3/6 | 25.0% | 4.6%–69.9% | |
| fear_greed_index.value | 7 | 1/3/3 | 25.0% | 4.6%–69.9% | |
| fear_greed_22 | 6 | 1/3/2 | 25.0% | 4.6%–69.9% | |
| oi_notional_24h | 6 | 1/3/2 | 25.0% | 4.6%–69.9% | |
| rv_24h_pctile | 6 | 1/4/1 | 20.0% | 3.6%–62.5% | |
| macro_dxy | 6 | 1/4/1 | 20.0% | 3.6%–62.5% | |
| regime_v19.hurst_exp_10d | 14 | 1/5/8 | 16.7% | 3.0%–56.4% | |
| macro.vix | 8 | 1/5/2 | 16.7% | 3.0%–56.4% | |
| taker_history_24h | 9 | 1/6/2 | 14.3% | 2.6%–51.3% | |
| binance_funding_rate | 7 | 0/2/5 | 0.0% | 0.0%–65.8% | |
| microstructure.taker_buy_sell_ratio_1h | 5 | 0/3/2 | 0.0% | 0.0%–56.1% |
Tier 2 — win rate by feature value
n per bucket: low=25 · mid=22 · high=17 (buckets with n<10 suppressed)
Tier 3 — statistical importance
Logistic regression (won vs lost), 64 signals × 65 features. Cross-validated AUC 0.796 ± 0.077
Bar = mean ΔAUC when the feature is shuffled (30 repeats). Green = positive logit coefficient (higher value → win), red = negative.