2025-07-25 –, Main Room 6
EvoTrees.jl is a Julia implementation of gradient-boosted trees, a state-of-the-art algorithm class for tabular data.
This talk provides an overview of EvoTrees.jl implementation and key features. We present recent advancements, including new loss functions, benchmarks against popular libraries, and planned improvements, such as enhanced GPU support and auto-diff for custom loss functions.
Key topics covered:
1. Key steps in gradient-boosted trees algorithm: binarization, build histogram, best split search, prediction.
2. EvoTrees approach to reconcile ease of research and achieving performance comparable to its C++ peers
3. Minimal benchmark against XGBoost, LightGBM and CatBoost
4. Beyond gradient-based learning: mean-absolute error and volatility-adjusted losses
5. Future development paths: improved GPU acceleration, auto-diff support for custom and multi-target losses and enhanced support of categorical variables.
Jeremie is the Head of Science at Evovest. He's responsible for the development of systematic investment strategies in global equities. Over 8 years at Intact in actuarial and R&D roles, he drove the development of analytical solutions in predictive risk modeling, price optimization and telematics.
He also worked as an applied research scientist at ElementAI and as a consulting actuary at WTW.