From XGBoost to LightGBM: How Our ML Model Adapted to Two Market Regimes

Sun, 19 Apr 2026 09:00:00 +0000

When we first wrote about SignalScope's ML backtesting pipeline in March, the model of record was XGBoost. Six weeks and hundreds of experiments later, it isn't. The current production model is a single LightGBM regressor with forty trees and depth two, trained on three-day forward returns, pulling almost all of its out-of-sample skill from a single feature. Mean information coefficient — the rank correlation between how the model ranks tickers and how they actually perform, where 0 is random and higher is better — has climbed from 0.006 on our first Ridge baseline to 0.161 today, a roughly 27x improvement. But the path between those two numbers wasn't monotonic. Twice in the last six weeks, a regime shift in the underlying data quietly invalidated our best model and forced us to start over. This is the story of how we rebuilt the pipeline three times — and what those rebuilds taught us about model stewardship in markets that keep changing underneath you.

Backtesting on myvirtualcloud.net

From XGBoost to LightGBM: How Our ML Model Adapted to Two Market Regimes