ML Edge Report
After 660 GPU configurations on RTX 4080 SUPER, Strategy Arena measured an ML edge on volatility regime prediction: Brier OOS 0.1215 on FLOKI 15min, versus baseline 0.2500. Five-minute direction remains at the random-walk floor.
The measured win: FLOKI volatility
Brier OOS
0.1215
0.1215
Baseline
0.2500
0.2500
Edge
+0.1285
+0.1285
CI90
[0.1182, 0.1247]
[0.1182, 0.1247]
FLOKI / 15min / volatility_next_24h / mlp_shallow / mega.
Secondary wins
| Target | Asset | TF | Arch | Features | Brier | Baseline | Edge | Caveat |
|---|---|---|---|---|---|---|---|---|
| breakout_next_10 | COMP | 15min | mlp_deep | mega | 0.0839 | 0.1338 | +0.0498 | Rare class; baseline already low |
| magnitude_next_1h | FLOKI | 5min | cnn1d | volatility | 0.2196 | 0.2500 | +0.0304 | Marginal edge |
Confirmed negative result
Five-minute direction: best observed result Brier OOS 0.2474, baseline 0.2463, edge -0.0011. We confirm this target remains at the random-walk floor in this run.
Anti-2CV transparency: we publish what does not work too. ML layers should be read as regime context, calibration, risk gating and filtering, not directional oracles.
Methodology
660 configs
Broad exploration + refinement
Broad exploration + refinement
Walk-forward
Temporal split, 48-bar embargo
Temporal split, 48-bar embargo
CI bootstrap
100 OOS resamples
100 OOS resamples
RTX 4080 SUPER
1.05h GPU compute
1.05h GPU compute
- Mandatory comparison versus train-climatology baseline.
- Target thresholds computed on train only to avoid leakage.
- Multiple-testing caution: we keep edge versus baseline, not absolute Brier alone.
Interpretation
Volatility regime prediction works. Five-minute direction prediction does not. This matches the literature: volatility clusters, while high-frequency direction approaches a random walk.
What we are not claiming: not yet a live tradable edge, not yet generalized across every asset, and the breakout target carries class-imbalance caveats.