Specialized versus Generic XGBoost Models for Direct Multi-Step Time Series Forecasting: A Comparative Study on the ETT Benchmark

Introduction
Accurate multi-horizon forecasting of electrical system variables is a foundational challenge in smart-grid management. While gradient-boosted trees, particularly XGBoost, demonstrate consistently strong performance in tabular time series tasks, whether explicit multivariate encoding of inter-channel correlations improves or hurts predictive accuracy compared to channel-independent architectures remains under-investigated.

Methods
This study conducts a systematic empirical comparison of two single-model XGBoost paradigms for direct multi-step forecasting: a Specialized (MIMO multivariate) model and a Generic (cross-variable univariate) model. Utilizing the Electricity Transformer Temperature (ETT) benchmark, both approaches employ a 96-step lookback window to predict a 24-step horizon. The Specialized paradigm jointly encodes all seven target channels into a flattened feature vector, whereas the Generic paradigm treats each channel as an independent sequence, effectively augmenting the training set by a factor of seven. Experiments evaluate performance across in-domain, out-of-domain, and combined-dataset training configurations.

Results
Quantitative evaluation demonstrates that the Generic paradigm consistently outperforms the Specialized model in every in-domain scenario. Furthermore, combined-dataset training improves the Generic model, which achieves an R2 of 0.957 on the MUFL target. Conversely, the Specialized model undergoes catastrophic degradation in cross-domain settings, generating an R2 of -9.09 on the HULL target, whereas the Generic model retains substantially greater robustness.

Conclusions
These findings extend the channel-independence hypothesis, previously established for deep-learning models, to gradient-boosted trees, indicating that temporal pattern generalization outweighs explicit cross-variable context for this class of benchmarks. The Generic model's superior performance and robustness are primarily driven by training-set amplification, reduced input dimensionality, and implicit temporal regularization.