Validation Geometry Is Part of the Model

Short summary

A saved label artifact containing only magnitude-filtered positive targets would have silently turned a model comparison into a dataset-artifact comparison for a cryptocurrency direction classifier. The author rebuilt the LightGBM baseline to read raw feature parquet files directly, matching a strict protocol: 46 features, 15-minute forward returns, stride-15 de-overlap, no-trade filtering, and per-asset temporal splits with purge gaps. The core lesson is that validation geometry in time-series ML is part of the model, not bookkeeping.

• •Saved label artifacts can silently change the ML task if they don't match your experimental protocol
• •Minute-scale crypto direction appears capped near 52% across seven configurations and 36M rows
• •Validation geometry—splits, purges, de-overlap—is part of the model, not administrative overhead

Generated with AI, which can make mistakes.