Source-linked AI summary
Halo: Improving forecast accuracy through heteroscedastic estimation
Adam Cataldo
TL;DR
Heteroscedastic forecasting is usually used for uncertainty quantification, but evidence was limited on whether estimating scale can also improve point forecasts. Halo adds a scale output and matching negative log likelihood to existing deep forecasters, and it improves accuracy across electricity-price benchmarks while making architecture choice and retuning less decisive.
Problem
Heteroscedastic forecasting is normally motivated by uncertainty quantification, while prior evidence outside time series reported worse point estimation.
Method
Halo reuses an existing deep forecaster, adds a scale output for its implied distribution, and trains the resulting model with the matching negative log likelihood.
Results
Halo improved MSE and MAE in 28 of 30 model-market-metric comparisons across three architectures and five electricity-price markets, with average MSE reductions of 2.6% to 16.5%.
Takeaways & Limitations
Whether the scale estimate uses a dual head or parallel network matters less than estimating scale, and baseline-tuned hyperparameters are often sufficient.
Takeaways & Limitations
The authors identify testing Halo on other benchmarks and horizons, and across seeds, as the next step.
Abstract
from arXiv · showhide
Heteroscedastic forecasting, where a network estimates a scale parameter alongside a location parameter, is normally motivated by uncertainty quantification. This paper shows it also improves the point estimate, in contrast to reported negative results for heteroscedastic estimation outside time series. Halo is a modification that reuses an existing deep forecaster's architecture, giving it a second output for the scale of its implied distribution and training it under the matching negative log likelihood. Adapting three state-of-the-art models --- a transformer, a graph network paired with a variational autoencoder, and a single-layer convolutional network --- under both Gaussian and Laplacian losses demonstrates the phenomenon. On the five electricity price markets of a standard forecasting benchmark, Halo improves MSE and MAE in 28 of 30 model-market-metric comparisons, cutting average MSE by 2.6% to 16.5% and average MAE by 1.7% to 11.0%. Two findings emerge: (1) whether the scale estimate comes from a second projection head or from a full parallel network matters far less than whether the network estimates scale, and (2) the improvement holds under the hyperparameters already tuned for the point-estimate baseline, so retuning is optional.
1 INTRODUCTION
Halo modifies existing time-series deep forecasters to learn both location and scale, arguing that the scale estimate can improve point accuracy. The paper studies this approach across architectures and finds performance gains on electricity-price forecasting benchmarks.
- Motivation: Time-series forecasting differs from other deep-learning problems because data points lack semantic meaning, exogenous signals can be more informative than the target’s history, and distributions may change over time.
- Contribution: Halo adds scale estimation to an existing time-series forecaster to improve its location estimate and quantify forecast uncertainty.The method explores how to architect the location and scale outputs and how heteroscedastic estimation interacts with nonstationarity adjustments.
- Contribution: The paper argues that learning the output scale helps the network locate the point estimate itself, beyond heteroscedastic forecasting’s usual uncertainty-quantification role.
- Contribution: Halo is evaluated as a performance-boosting technique on several state-of-the-art forecasting models.
2 RELATED WORK
The related work places Halo within the development of deep neural forecasting and heteroscedastic estimation. It also contrasts the paper’s claim with a prior negative result obtained outside time-series forecasting.
- Deep forecasting: Deep forecasting has progressed from recurrent and convolutional networks to architectures such as N-BEATS and transformers.
- Heteroscedastic estimation: Heteroscedastic estimation has a long neural-network history, while probabilistic modeling remains less common than point estimation in time-series deep learning.
- Contrast with prior work: A prior study found worse point estimation with heteroscedastic models, but it used non-time-series models that were much smaller than those studied here.
3 HALO
Halo modifies an existing forecaster to estimate both location and scale, training the resulting distributional forecast with a matching negative log likelihood. The method supports probabilistic forecasts while providing evidence that learning scale can improve point estimates by accounting for variability.
- 3 HALO: Halo changes the network output from a point estimate to jointly predicted location and scale parameters for each forecast horizon.For Gaussian outputs, the network predicts mean and variance; scale estimates are constrained positive with a softplus operator.
- 3.1 HETEROSCEDASTIC LOSS: MSE and MAE training have probabilistic interpretations, corresponding to negative log likelihood under distributions with constant variance or scale.This motivates replacing point-estimate training with heteroscedastic likelihood-based training when the network predicts varying scale.
- 3.1 HETEROSCEDASTIC LOSS: The Gaussian implementation uses β-NLL with a stop-gradient term because it preserves the NLL optimum while providing more stable backpropagation updates.The paper fixes β at 0.5; no analogous modification is used for the Laplacian distribution.
- 3.1 HETEROSCEDASTIC LOSS: The assumed distribution determines the likelihood loss, so validation-set full NLL can select between Gaussian and Laplacian alternatives when the appropriate choice is unclear.The paper distinguishes full NLL model selection from objectives such as MSE or β-NLL that may share an optimum without serving as the selection criterion.
- 3.2 FROM POINT ESTIMATES TO HETEROSCEDASTIC ESTIMATES: A second projection head reuses one learned representation for location and scale, avoiding the duplicated representation learning of two parallel networks.Both architectures generally outperform the non-heteroscedastic architecture, while their relative accuracy is close to even.
- 3.3 NONSTATIONARITY ADJUSTMENTS: Halo is combined with Series Standardization by adjusting the emitted scale alongside the location when reversing the input normalization.All models use Series Standardization, and the experiments apply these adjusted outputs.
4 EXPERIMENTS
Experiments evaluate Halo across three forecasting architectures and five electricity-price markets, comparing heteroscedastic variants with baseline models and testing architecture and tuning choices.
- Experimental setup: Three state-of-the-art models test Halo across transformer, GNN–VAE, and single-layer CNN architectures under Gaussian and Laplacian losses.TimeXer and GCGNet use dual-head Halo, while CrossLinear uses parallel Halo for the headline results.
- Main results: Halo generally improves performance across five electricity-price markets, with only one market where TimeXer performs better without Halo.Table 1 compares each model with and without Halo and reports average MSE and MAE across markets.
- Hyperparameter tuning: Table 2 compares TimeXer + Halo using the original configuration, retuned hyperparameters, and the baseline model’s tuned configuration.This isolates the effect of retuning from the effect of applying Halo.
- Architecture comparison: CrossLinear experiments compare parallel Halo, dual-head Halo, and no Halo, with parallel Halo outperforming dual-head Halo on validation data.The test-set difference between the two Halo architectures is less decisive than the validation result.
- Architecture comparison: TimeXer experiments likewise compare parallel Halo, dual-head Halo, and no Halo under the configurations summarized in Table 4.For TimeXer, the less expensive dual-head architecture supports the headline results because tuning differences were less pronounced.
5 CONCLUSION
The paper introduces Halo as a heteroscedastic modification of existing forecasters and finds broad accuracy gains across models and electricity-price markets. The experiments also indicate that architecture choice and retuning are secondary to estimating scale, while broader validation remains needed.
- Conclusion: Halo improves MSE and MAE in 28 of 30 model-market-metric comparisons across three architectures and five electricity-price markets.The evaluated architectures are a transformer, a GNN paired with a VAE, and a single-layer CNN.
- Conclusion: Average MSE falls 2.6%–16.5% and average MAE falls 1.7%–11.0%, depending on the model.Every model’s average improves on both metrics.
- Practical findings: The scale-estimation architecture matters less than switching from point estimation to heteroscedastic estimation, while the dual-head design uses roughly half as many parameters as the parallel design.The two architectures finish within 1.5% on both models, making dual-head Halo preferable when parameter cost matters.
- Practical findings: Retuning is optional: validation retuning helped in two markets and hurt in three, leaving test averages within half a percent of their starting values.The improvement does not require retuning the point-estimate baseline’s hyperparameters.
- Scope: The evidence is scoped to the evaluated electricity-price setting, so testing other benchmarks, horizons, and random seeds remains a next step.The discussion contrasts these results with prior non-time-series findings and attributes the difference to time-varying volatility as a possible source of learnable scale information.
A IMPLEMENTATION DETAILS
All models are trained on a single Apple M2 Max system under a common 50-epoch budget with early stopping.
- Hardware and training: Training uses one Apple M2 Max system with a 30-core GPU, 12 CPU cores, and 32 GB shared memory.PyTorch dispatches to the GPU through Metal Performance Shaders on macOS.
- Hardware and training: Runs last up to 50 epochs with early-stopping patience of 5, placing the three models on roughly equal training footing.This budget matches GCGNet’s training budget while exceeding the reported budgets for TimeXer and CrossLinear.