Halo Improves Deep Time Series Forecasting Accuracy

A new research preprint on arXiv introduces Halo, a method designed to enhance the accuracy of deep time series forecasters. While heteroscedastic forecasting is conventionally used for uncertainty quantification, this work demonstrates that estimating a scale parameter also significantly improves point forecast estimates.

Halo modifies existing deep forecasting architectures by adding a second output to capture the distribution's scale and training it under the matching negative log likelihood. The authors evaluated the approach across three modern architectures — a transformer, a graph network combined with a variational autoencoder, and a single-layer convolutional network — using Gaussian and Laplacian loss functions.

Benchmarked across five electricity price markets, Halo reduced MSE and MAE errors in 28 out of 30 comparisons. It cut average MSE by 2.6% to 16.5% and average MAE by 1.7% to 11.0%. The findings also indicate that the improvement holds using hyperparameters already tuned for baseline models without requiring retuning.