Halo improves point forecasts by also estimating their uncertainty

In heteroscedastic forecasting, a network estimates a scale parameter for its output distribution alongside the usual location, or point, parameter, normally to quantify uncertainty. A new paper reports that this also improves the point estimate itself, in contrast to negative results reported for heteroscedastic estimation outside time series.

The paper's method, called Halo, reuses an existing deep forecaster's architecture rather than replacing it. It gives the network a second output for the scale of its implied distribution and trains the whole model under the negative log likelihood that matches that distribution. The authors adapt three state-of-the-art forecasting architectures this way, a transformer, a graph network paired with a variational autoencoder, and a single-layer convolutional network, and test each one under both a Gaussian and a Laplacian loss.

On five electricity price markets from a standard forecasting benchmark, Halo improves both mean squared error (MSE) and mean absolute error (MAE) in 28 of the 30 model-market-metric combinations tested. Averaged across models and markets, MSE falls by 2.6% to 16.5% and MAE by 1.7% to 11.0%, depending on which model and market are compared.

Two further findings narrow down what drives the gain. First, it barely matters whether the scale output comes from a small second projection head added to the existing network or from a full separate parallel network: what matters is that the model estimates a scale at all, not how it does so. Second, the improvement shows up under the hyperparameters already tuned for each model's original point-estimate version, so adopting Halo did not require a fresh round of hyperparameter tuning in the paper's own tests.

Key facts

Halo gives an existing deep forecasting model a second output for the scale of its implied distribution, alongside its usual point output, and trains the model under the matching negative log likelihood loss.
The method is tested on three state-of-the-art forecasters, a transformer, a graph network paired with a variational autoencoder, and a single-layer convolutional network, under both Gaussian and Laplacian losses.
On five electricity price markets, Halo improves MSE and MAE in 28 of 30 model-market-metric comparisons.
Averaged across models and markets, MSE falls by 2.6% to 16.5% and MAE by 1.7% to 11.0%.
The gains hold under the hyperparameters already tuned for each model's original point-estimate version, so adopting Halo did not require retuning in the paper's tests.

Why it matters

Heteroscedastic forecasting, where a network estimates the scale of its output distribution alongside the point prediction, is normally used only for uncertainty quantification, that is, to know how confident the model is. Prior work outside time series has reported negative results when that scale estimate was expected to also help the point prediction. This paper's central claim is that within time series forecasting the opposite holds: adding the scale estimate, and training under the loss that matches it, also sharpens the point forecast itself. That reframes an output normally computed only for uncertainty quantification as a possible source of accuracy gains rather than a separate, secondary concern.

Who it affects

The direct audience is researchers and engineers building deep learning forecasters for time series data, the kind of problem the paper tests through electricity price prediction. Because Halo reuses an existing model's architecture rather than replacing it, teams that already have a point-estimate forecasting model built are best placed to try it without a full rebuild.

How to use it

Concretely, Halo adds a second output head to an existing deep forecasting network, one that predicts the scale of the implied output distribution rather than just its location, and trains the whole model under the negative log likelihood loss that matches that distribution; the paper tests both a Gaussian and a Laplacian version. It adapts three quite different architectures this way, a transformer, a graph network paired with a variational autoencoder, and a single-layer convolutional network, which suggests the trick is not tied to one architecture family. A practical detail for adoption: the improvement held under the hyperparameters already tuned for each model's original point-estimate version, so the paper's own tests did not need a fresh tuning pass to see the gain.

How solid is it

The claim rests on 30 model-market-metric comparisons: three adapted models across five electricity price markets from an unnamed standard forecasting benchmark, each scored on MSE and MAE. Halo improves both metrics in 28 of those 30 comparisons, with average MSE down 2.6% to 16.5% and average MAE down 1.7% to 11.0%. The source gives only these relative reductions, not the underlying absolute error values, so how large the errors were to begin with cannot be judged from the text. The pattern holds across three architecturally different models and two loss families, a reasonable breadth for one paper, but the test is confined to a single domain, electricity price forecasting, and a single, unnamed benchmark. The paper is posted on arXiv, and the text gives no submission or publication date.

Risks and caveats

The source names no authors or institutions, so the work's provenance cannot be checked from the text alone. It also names neither the benchmark nor the three underlying model architectures beyond generic descriptions, a transformer, a graph network paired with a variational autoencoder, a single-layer convolutional network, which limits independent verification. Only relative reductions in MSE and MAE are reported, not absolute error figures. Two of the 30 model-market-metric comparisons did not improve, so the effect is not universal even within the one domain tested. Because the tests cover only electricity price forecasting on one benchmark, they do not by themselves establish whether the technique carries over to other time series problems. The text does not mention code or data availability for reproducing the results.

“This paper shows it also improves the point estimate, in contrast to reported negative results for heteroscedastic estimation outside time series.”

— the paper's authors