Mechanism diagram · Tabular Tree
How it works
Standard LightGBM with mean-squared-error loss per coin.
The model is trained per-coin (independent) or as a panel; either way the loss is regression, not ranking.
Output is a per-coin forecast.
Pros and cons on this universe
Pros
- Fast, parallelisable, no GPU needed.
- Strong on older CPCV consolidated cards (SR_net mean ~+2.2 — but on a different window and different metric).
Cons / failure modes
- Near-zero universe IC (+0.023).
- Superseded by LGBM Rank for cross-sectional ranking tasks.
- Regression loss does not align with the cross-sectional decision the strategy makes.