Lab · ML Experiments

ML — Pattern Discovery

Inverted workflow: find conditional edges in BTC data first, build strategies second.
55 experiments
Overview & Experiments 55 Synthesis Roadmap Lookahead Audit

Direction Stage 1.5 — GBM with macro/sentiment features

Dropped
2026-05-17 directionstage-1.5macrocore
Hypothesis
Adding Fear & Greed sentiment, funding-rate positioning, and ETH/SOL cross-asset features lifts the baseline GBM's IC above +0.02 and the long-short spread above 15 bps — closing the gap toward a tradeable directional edge.
Verdict
**MACRO ADDS NOTHING** — IC +0.0161 vs Stage 1 +0.0148 (lift +0.0013), L/S spread +3.6 bps. Sentiment + funding + cross-asset don't extract a directional edge on this data. The honest answer: vol-targeting is the ceiling; pivot to live-trading + accumulating better data (Binance snapshot collector) for future iterations.
n_windows
21
pooled_IC
+0.0161
gates_total
3
gates_passed
1
pooled_IC_CI
[-0.0010, +0.0337]
long_short_bps
+3.5769
n_features_macro
14
n_features_total
47
IC_lift_vs_stage1
+0.0013
long_short_t_stat
+0.9293
positive_IC_rate_pct
+85.7143
spread_lift_vs_stage1
-1.6231

Direction Stage 1.5 — GBM with macro/sentiment features

2026-05-17 · status: dropped · 12.3s

Hypothesis: Adding Fear & Greed sentiment, funding-rate positioning, and ETH/SOL cross-asset features lifts the baseline GBM's IC above +0.02 and the long-short spread above 15 bps — closing the gap toward a tradeable directional edge.

Verdict: MACRO ADDS NOTHING — IC +0.0161 vs Stage 1 +0.0148 (lift +0.0013), L/S spread +3.6 bps. Sentiment + funding + cross-asset don't extract a directional edge on this data. The honest answer: vol-targeting is the ceiling; pivot to live-trading + accumulating better data (Binance snapshot collector) for future iterations.

Key metrics

metric value
n_features_total 47
n_features_macro 14
n_windows 21
pooled_IC +0.0161
pooled_IC_CI [-0.0010, +0.0337]
IC_lift_vs_stage1 +0.0013
long_short_bps +3.5769
long_short_t_stat +0.9293
spread_lift_vs_stage1 -1.6231
positive_IC_rate_pct +85.7143
gates_passed 1
gates_total 3

Approach

Same GBM (HistGradientBoosting, 200 iter, depth 4, lr 0.05) and same walk-forward scheme as Stage 1 — only the feature set is bigger.

  • Base features (33): OHLCV-derived from direction_baseline
  • Macro features (14): fng_value, fng_chg_7d, fng_z_30d, fng_extreme, fr, fr_z_30d, fr_cum_3d, fr_cum_7d, eth_btc_log_ratio, eth_btc_ret_4h, eth_btc_corr_24h, btc_eth_vol_spread, sol_btc_log_ratio, sol_btc_ret_4h

Source historical depth: - Fear & Greed: 2018-02 → today (daily) - Funding rate: 2023-05 → today (8h settlements) — NaN in earlier windows - ETH/SOL: 2020-01 → today (1m); SOL starts 2020-09

HistGBR handles NaN natively (treats missing as its own split direction), so windows before funding data exists still train cleanly — they just can't use the funding features.

Stage 1 reference: pooled IC +0.0148, L/S spread +5.2 bps, stability 76.2%.

1. Walk-forward IC

  • Pooled Spearman IC (non-overlapping 4h): +0.0161 (n = 11,496, 95% CI [-0.0010, +0.0337])
  • Per-window IC: min -0.0221, max +0.0964, mean +0.0253
  • Positive-IC windows: 18 / 21 (85.7%)
  • Delta vs Stage 1: IC +0.0013 | stability +9.5 pp
window test_start test_end n IC
1 2021-01-05 2021-04-05 2160 0.0514
2 2021-04-05 2021-07-05 2184 0.001
3 2021-07-05 2021-10-05 2208 0.0246
4 2021-10-05 2022-01-05 2208 0.0433
5 2022-01-05 2022-04-05 2160 -0.0146
6 2022-04-05 2022-07-05 2184 0.0046
7 2022-07-05 2022-10-05 2208 0.0205
8 2022-10-05 2023-01-05 2208 0.0058
9 2023-01-05 2023-04-05 2160 0.0252
10 2023-04-05 2023-07-05 2184 0.056
11 2023-07-05 2023-10-05 2208 -0.0221
12 2023-10-05 2024-01-05 2208 0.0833
13 2024-01-05 2024-04-05 2184 0.0964
14 2024-04-05 2024-07-05 2184 0.0095
15 2024-07-05 2024-10-05 2208 0.0332
16 2024-10-05 2025-01-05 2207 0.0202
17 2025-01-05 2025-04-05 2160 0.0034
18 2025-04-05 2025-07-05 2184 -0.0154
19 2025-07-05 2025-10-05 2208 0.0039
20 2025-10-05 2026-01-05 2208 0.0757
21 2026-01-05 2026-04-05 2160 0.0254

2. Quintile spread (raw 4h returns)

q n mean_target_norm mean_bps se_bps t_stat
0 2300 0.0106 -0.15 2.56 -0.06
1 2299 -0.0336 -2.15 2.4 -0.9
2 2299 -0.0011 -1.36 2.65 -0.51
3 2299 0.0318 3.45 2.71 1.27
4 2299 0.0073 3.42 2.87 1.19
  • Long-short spread Q4 - Q0: +3.6 bps (SE ≈ 3.85, t ≈ +0.93)
  • Delta vs Stage 1: -1.6 bps

3. Permutation importance (last test window)

Computed by shuffling each feature's values and measuring R² drop. Macro features are highlighted via the is_macro column. Higher = more informative.

feature importance std is_macro
fr_cum_7d 0.01173 0.00431 True
kurt_1d 0.0038 0.00244 False
range_4h 0.00298 0.00058 False
rv_240m 0.00279 0.00067 False
rv_60m 0.00255 0.00041 False
fr_z_30d 0.00218 0.00343 True
hour_cos 0.00184 0.00318 False
dow 0.00152 0.00045 False
dd_1d 0.00124 0.00146 False
pk_vol_240m 0.00124 0.00114 False
hour_of_day 0.0012 0.00097 False
log_vol 0.00113 0.00017 False
ret_5m 0.00095 0.00119 False
fng_z_30d 0.00069 0.00088 True
vol_z_1d 0.00053 0.00086 False
vol_z_1h 0.00027 0.00071 False
hour_sin 0.00015 0.00049 False
is_weekend 0 0 False
fng_extreme 0 0 True
ret_15m -6e-05 4e-05 False
  • Macro features total importance: -0.00769 (-0.8% of positive importance)
  • Base features total importance: -0.06495 (-6.5%)

per-window IC

quintile

Production gates

gate pass? actual

| pooled IC > 0.020 | FAIL | +0.0161 |

| quintile spread > 15 bps | FAIL | +3.6 bps |

| positive IC in ≥ 70% windows | OK | 85.7% |

Passed: 1/3

Lifts vs Stage 1: IC +0.0013, spread -1.6 bps.