Research map / Time series / Long-term forecasting

Dirichlet-Guided Group Forecasting for Alleviating Over-smoothing in Time Series Forecasting

NeurIPS 2026 · Time series › Long-term forecasting · approach: Attention for time series forecasting · 0 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 Dirichlet-Guided Group Forecasting (DGF), a mode-preserving time series forecaster that samples mode-conditioned predictive distributions via a Dirichlet-distributed mode-selection vector and trains them with hierarchical GRPO using accuracy, dynamics, and diversity rewards, built on Moirai backbones.

Problem

Time series forecasting models over-smooth multi-modal futures, averaging incompatible dynamic modes and losing sharp changes, oscillations and regime transitions under single-realization supervision.

Approach

Model K mode-conditioned predictive distributions plus a Dirichlet distribution over mode-selection probabilities; sample one mode-composition vector per mini-group and train with hierarchical GRPO (sample-level and Dirichlet-level advantages) plus a distribution-level diversity regularizer, instead of averaging modes.

Key results

• On Traffic, DGF with Moirai-base reaches MSE 0.295 (mean) vs 0.407 for full finetuning of Moirai-base.
• Probabilistic CRPS on Electricity improves to 0.031 with DGF on Moirai-base vs 0.046 for MSFT-adapted Moirai-base.
• Long-sequence averages: DGF with Moirai-base reports best MSE on ETTm1 (0.317) and Electricity (0.157) among compared methods.

Surprising findings

• Deterministic LSF evaluation reports the MAP-mode output, and the paper notes that alternative decision protocols are only analyzed in the appendix.

Limitations

• Requires a reasonably good warm-start backbone; reward optimization is unstable from poor initialization.
• Additional sampling and training cost versus deterministic forecasting.
• Hyperparameters (K, B, G, gamma, lambda) add tuning burden and cost.

Experimental setup

Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Electricity Weather Solar Istanbul Traffic Turkey Power Traffic

Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Electricity Weather Solar Istanbul Traffic Turkey Power Traffic

Metrics: MSE MAE CRPS MSIS DILATE mean-sample score best-sample score

Baselines: DLinear PatchTST iTransformer TimeMixer SimpleTM Moirai-small Moirai-base DeepAR TiDE N-BEATS cVAE-DILATE Variety Loss

Builds on: GRPO Moirai STRIPE

Closest papers by meaning

• Rethinking Post-Training Recipes for Multimodal Time-Series Forecasting (NeurIPS 2026 · similarity 0.88)
• TimeEmb: A Lightweight Static-Dynamic Disentanglement Framework for Time Series Forecasting (NeurIPS 2025 · similarity 0.86)
• Multivariate Time Series Forecasting needs Cross Variable Loss (NeurIPS 2026 · similarity 0.86)
• LangTime: A Language-Guided Unified Model for Time Series Forecasting with Proximal Policy Optimization (ICML 2025 · similarity 0.86)
• Learning Pattern-Specific Experts for Time Series Forecasting Under Patch-level Distribution Shift (NeurIPS 2025 · similarity 0.85)
• Auto-Regressive Moving Diffusion Models for Time Series Forecasting (AAAI 2025 · similarity 0.85)