Research map / Time series / Long-term forecasting

DBLoss: Decomposition-based Loss Function for Time Series Forecasting

NeurIPS 2025 · Time series › Long-term forecasting · approach: Attention for time series forecasting · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DBLoss, a decomposition-based loss function for time series forecasting that augments MSE to better capture seasonality and trend within the forecast horizon.

Problem

Standard MSE loss sometimes fails to capture seasonality or trend within the forecasting horizon, even when decomposition modules are used in the forward pass.

Approach

Replaces or augments the MSE objective with a decomposition-based loss that explicitly supervises seasonal and trend components of the forecast, so the loss itself captures these patterns rather than only the forward-pass decomposition.

Key results

• The SRSNet variant reported in the main table achieves best MSE on ETTh1 (0.404), ETTh2 (0.334), and Traffic (0.392) among listed baselines.
• Main table covers 8 multivariate datasets with horizons {96,192,336,720} averaged, compared against 10 baselines.

Limitations

• The provided text is an incomplete template; method details, limitations, and implementation are not present.
• Main table lists SRSNet as the proposed model, which does not match the DBLoss title, so the relation between the two is unclear from the text.

Experimental setup

Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Solar Traffic

Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Solar Traffic

Metrics: MSE MAE

Baselines: FEDformer Stationary DLinear TimesNet Crossformer PatchTST TimeMixer iTransformer Amplifier TimeKAN

Builds on: PatchTST DLinear

Closest papers by meaning

• Patch-wise Structural Loss for Time Series Forecasting (ICML 2025 · similarity 0.87)
• TimePerceiver: An Encoder-Decoder Framework for Generalized Time-Series Forecasting (NeurIPS 2025 · similarity 0.86)
• Parsimony or Capability? Decomposition Delivers Both in Long-term Time Series Forecasting (NeurIPS 2024 · similarity 0.85)
• Multivariate Time Series Forecasting needs Cross Variable Loss (NeurIPS 2026 · similarity 0.85)
• Loss Shaping Constraints for Long-Term Time Series Forecasting (ICML 2024 · similarity 0.85)
• Towards a General Time Series Forecasting Model with Unified Representation and Adaptive Transfer (ICML 2025 · similarity 0.85)