A training-time fix for uneven frequency errors
Frequency-domain forecasters split a series into components that vary at different speeds, then predict each component with its own network. BEAT's starting observation is that separating these representations does not, by itself, say how training should respond when one component fits much worse than another. Under a single shared forecasting loss, the branches can end up carrying quite different coefficient prediction errors.
BEAT's intervention is narrow. It leaves the forward prediction path alone and adds two training-time operations: a Frequency-Specific Monitor (FSM) that measures each branch's coefficient error, and a Dynamical Gradient Balancer (DGB) that turns those measurements into per-branch gradient multipliers. Because the multipliers are positive, they rescale gradient magnitude without reversing direction. The paper's claim is therefore about training emphasis, not about a new forecasting architecture.
How relative error becomes a gradient weight
FSM needs a fair comparison across branches, so it first puts targets in the same space as predictions. The ground-truth future window is normalized with the same reversible instance normalization statistics and affine parameters used for the historical input, then decomposed with the same wavelet transform. Each branch is thus scored against a target at its own resolution and on the same scale.
The discrepancy for each branch is a mean squared error between predicted and target coefficients. The reference point is the mean discrepancy across the detail branches only, excluding the approximation branch. Each branch's ratio of its own discrepancy to that reference becomes its relative error, and DGB converts the ratio into a bounded positive coefficient. Branches with higher relative error get larger gradient weights; branches with lower relative error get smaller ones. A single shared modulation-strength parameter controls how far training departs from the unmodulated case, and the ratios are recomputed at every training step, so the scheme does not lock in a fixed preference for low or high frequencies.
One detail matters for reading the ablations: the monitoring metric and the reported forecasting metric are different things. The monitor scores coefficient discrepancies in the normalized space and shapes the training signal; MSE and MAE in the results tables score the reconstructed forecast after inverse normalization.
Where the gains show up across seven datasets
The evaluation covers seven real-world datasets at four prediction lengths (96, 192, 336 and 720) against eight baselines, with each setting run three times and averaged. BEAT records the most first-place results, 26 for MSE and 30 for MAE across the 28 dataset-horizon settings plus seven average rows, counting ties. It has the lowest horizon-averaged MSE and MAE on Weather, all four ETT datasets and ECL.
The cleanest comparison is against its own base. BEAT uses WPMixer's forward architecture, and it achieves lower MSE and MAE than WPMixer in all 28 dataset-horizon settings. On ETTm2 the horizon-averaged MSE/MAE pair is 0.265/0.312 for BEAT versus 0.271/0.318 for WPMixer. Against the newer Delayformer and UP-TimeMixer, BEAT has lower horizon-averaged MSE and MAE on all seven datasets under their reported settings, for example 0.349 versus 0.377 average MSE on ETTh2, a 7.4% reduction.
The advantage is not uniform across horizons. On Traffic, BEAT ranks second on both average metrics behind iTransformer, and at prediction length 720 UP-TimeMixer obtains lower MSE on ETTm1, ETTm2 and Traffic. Average-row wins should not be read as wins at every horizon.
Which design choices the ablations actually support
Ablations on Weather and ETTh1 test sets, averaged over the four prediction lengths, isolate the components. Enabling FSM and DGB together lowers Weather MSE/MAE from 0.243/0.269 to 0.239/0.263 and ETTh1 from 0.422/0.424 to 0.415/0.419. The two components are evaluated jointly, matching their complementary roles: FSM supplies the measurements and DGB acts on them.
The modulation design is tested directly. Branch-specific gradient weighting beats a shared-loss variant that multiplies the whole loss by the average coefficient, on both datasets (0.239/0.263 versus 0.242/0.268 on Weather; 0.415/0.419 versus 0.420/0.421 on ETTh1). That supports the paper's specific claim that weights should follow each branch's own relative error rather than one multiplier for all branches.
Other choices are dataset-dependent rather than universal. Daubechies is best on Weather and Coiflets on ETTh1, and the best decomposition level is 2 on Weather and 3 on ETTh1, with levels 4 and 5 not improving results; the authors suggest deeper decomposition yields shorter coefficient sequences at coarse scales, limiting temporal modeling in those branches. For the discrepancy measure, MSE gives the lowest Weather MSE and MAE and the lowest ETTh1 MSE, tying with MAE on ETTh1 MAE.
What the comparison does not settle
The paper's own explanation for the WPMixer gap is presented as plausible rather than demonstrated: because the forward architecture is unchanged, error-dependent gradient weighting directing more training emphasis toward poorly fitted branches is offered as an interpretation, supported by the joint FSM-plus-DGB ablation. The evidence establishes that the training change coincides with lower error; it does not isolate a causal pathway beyond that ablation.
Scope is also bounded. The method is instantiated with WPMixer and a multi-level discrete wavelet transform, so the reported numbers describe that combination, not arbitrary frequency-domain forecasters. Ablations cover only Weather and ETTh1, so wavelet type and decomposition level recommendations should be treated as dataset-specific findings rather than defaults. And the gains are small in absolute terms: the Weather average MSE moves from 0.243 to 0.239.
When branch-level gradient weighting is worth the change
The decision-relevant split is between average accuracy and horizon-level behavior. If the goal is the best horizon-averaged MSE and MAE on Weather, the ETT datasets or ECL, BEAT is the stronger choice in this comparison, and it improves on WPMixer everywhere tested without touching inference: FSM and DGB run only during training, so prediction requires just normalization, decomposition, the branch networks, reconstruction and de-normalization.
If the priority is Traffic or specific long horizons, the picture changes. iTransformer leads both Traffic averages, and UP-TimeMixer wins MSE at length 720 on ETTm1, ETTm2 and Traffic. The paper's most defensible priority is therefore not "replace the architecture" but "add error-dependent gradient weighting to an existing wavelet forecaster when branches fit unevenly" — and to check per-horizon results, not just the average row, before adopting it.
Sources used in this answer
BEAT: Balanced Frequency Adaptive Tuning for Long-Term Time-Series Forecasting
BEAT adds training-only frequency-specific error monitoring and bounded per-branch gradient modulation to a wavelet forecaster, achieving the most first-place results among nine methods and lower MSE and MAE than WPMixer in all 28 dataset-horizon settings.