Explainable Machine Learning for BTCIRT Limit Order Book Dynamics

Problem

Dense LOB forecasting demos assume regular sampling and exact short horizons, but Nobitex BTCIRT snapshots arrive with median gaps near a minute — so fixed 10/30/60-second claims can be scientifically misleading.

Context

Sparse LOB snapshots require separating next-observation, next-change, and strict-horizon questions, storing actual delays, and documenting underpowered analyses honestly.

Approach

1. Audited sparse LOB snapshots and redesigned Studies A / B / C with delay tracking and leakage controls.
2. Compared price-only vs LOB feature families with XGBoost and CatBoost under chronological validation.
3. Explained models with SHAP, permutation importance, and ablation; published a full PDF report.

Solution

An explainable multi-study pipeline whose claims match what sparse LOB data can support, with underpowered pilots labeled instead of oversold.

Outcome

Development-test results for next-observation and next-change studies, LOB incremental-value tables, and a complete PDF with figures. Strict fixed-horizon pilots remain underpowered on this sample.