ICML 2025 CO-BUILD contest: XGBoost iterations

Vladimir Pyltsov

Keywords: Forecasting, Building, XGBoost

TL;DR: This document is intended to report methodology and results for the ICML 2025 CO-BUILD contest submission.

Abstract: This document is intended to report methodology and results for the ICML 2025 CO-BUILD contest submission. The report describes the approach, including preliminary data analysis, review of the previous submissions, model selection and comparison, and discusses the findings of the results. The method approaches the contest from a data analytical perspective rather than a simulation viewpoint, treating data as a time series forecasting task. Various models are iterated, and a few data processing techniques are explored. The results show improved accuracy compared to the benchmarked value. The limitations and implications are discussed.

Submission Number: 37