Research map / Time series / Long-term forecasting
GMP-AR: Granularity Message Passing and Adaptive Reconciliation for Temporal Hierarchy Forecasting
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 GMP-AR, a temporal hierarchy forecasting framework combining granularity message passing for base forecasts, node-weighted adaptive projection reconciliation, and a plug-and-play task-based optimization module, evaluated on three public datasets and deployed in Alipay payment traffic management.
Problem
Forecasting time series at multiple temporal granularities while keeping forecasts coherent (aggregation constraints) and accurate, without sacrificing accuracy during reconciliation, and adapting to real-world task targets and constraints.
Approach
Uses top-down proportion inputs and child-distribution features plus CNN-based child-to-parent fusion to enrich per-node temporal representations, then reconciles with a node-dependent weighted projection (adaptive reconciliation) that is learned by the network, and finally applies an optimizer for task targets and constraints.
Key results
- GMP-AR reaches b-MAPE 0.1499 on Electricity versus 0.1765 for the best baseline SLOTH and 0.1761 for DeepAR-TDProb, and MAPE 0.2158 vs 0.2370 for DeepAR-TDProb.
- GMP-AR achieves the best MAPE on Traffic (0.3798) and Exchange Rate (0.0082) among all compared methods.
- On Alipay, GMP-AR with the optimization module improves MAPE by 3.8% over the online DeepAR baseline without loss in trend accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GMP with adaptive reconciliation instead of the strongest reported baseline | temporal hierarchy forecasting · Electricity | b-MAPE: 0.1761(0.0055) -> 0.1499(0.0021) | Helped |
| use GMP with adaptive reconciliation instead of the strongest reported baseline | temporal hierarchy forecasting · Electricity | MAPE: 0.1319 -> 0.2158(0.003) | Hurt |
| use GMP with adaptive reconciliation instead of the strongest reported baseline | temporal hierarchy forecasting · Traffic | b-MAPE: 0.4621(0.0026) -> 0.4289(0.0202) | Helped |
| use GMP with adaptive reconciliation instead of the strongest reported baseline | temporal hierarchy forecasting · Traffic | MAPE: 0.4292(0.0377) -> 0.3798(0.0066) | Helped |
| use GMP with adaptive reconciliation instead of the strongest reported baseline | temporal hierarchy forecasting · Exchange Rate | b-MAPE: 0.0092(0.0006) -> 0.0085(0.0021) | Helped |
| use GMP with adaptive reconciliation instead of the strongest reported baseline | temporal hierarchy forecasting · Exchange Rate | MAPE: 0.0087(0.0006) -> 0.0082(0.0021) | Helped |
| add CNN child feature fusion instead of direct temporal features (ablation) | GMP · three public datasets | relative forecast improvement: 8.55% | Helped |
| add Attention child feature fusion instead of direct temporal features (ablation) | GMP · three public datasets | relative forecast improvement: 22.76% | Helped |
| use CNN rather than Attention for child feature fusion (ablation) | GMP · three public datasets | relative forecast improvement: 14.21% | Helped |
| add child distribution modeling to vanilla temporal inputs (ablation) | GMP · three public datasets | relative forecast improvement: 4.34% | Helped |
| add top-down proportion inputs to vanilla temporal inputs (ablation) | GMP · three public datasets | relative forecast improvement: 5.79% | Helped |
| add both child distribution modeling and top-down proportion inputs (ablation) | GMP · three public datasets | relative forecast improvement: 13.47% | Helped |
Surprising findings
- Statistical THIEF methods outperform deep and multivariate HTS baselines on overall MAPE for the Electricity dataset.
- The model with best bottom-level b-MAPE does not necessarily achieve best overall MAPE, attributed to differing level contributions.
Limitations
- Assumes all nodes at the same level have the same number of children, so the tree structure is fixed per level.
- Vanilla GMP-AR trades off trend accuracy (about 5% lower) when optimizing only for MAPE, and the Alipay application is proprietary.
- Node-dependent weights learned by the network lack a theoretical guarantee of forecast improvement beyond empirical results.
Experimental setup
Datasets: Electricity Traffic Exchange Rate Alipay payment traffic
Benchmarks: Electricity Traffic Exchange Rate
Metrics: MAPE b-MAPE trend accuracy
Baselines: THIEF DeepAR NBeats Autoformer HierE2E SHARQ COPDeepAR SLOTH TDProb BU Proj
Builds on: TDProb HierE2E closed-form projection DeepAR GRU COPDeepAR
Closest papers by meaning
- TimeTok: Granularity-Controllable Time-Series Generation via Hierarchical Tokenization (NeurIPS 2026 · similarity 0.84)
- Rethinking the Power of Timestamps for Robust Time Series Forecasting: A Global-Local Fusion Perspective (NeurIPS 2024 · similarity 0.83)
- MG-TSD: Multi-Granularity Time Series Diffusion Models with Guided Learning Process (ICLR 2024 · similarity 0.83)
- Learning Optimal Projection for Forecast Reconciliation of Hierarchical Time Series (ICML 2024 · similarity 0.83)
- TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting (ICML 2025 · similarity 0.83)
- Dirichlet-Guided Group Forecasting for Alleviating Over-smoothing in Time Series Forecasting (NeurIPS 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.