Research map / Time series / Long-term forecasting
Improving Time Series Forecasting via Instance-aware Post-hoc Revision
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 PIR, a model-agnostic post-hoc plugin that estimates per-instance forecasting error and revises backbone forecasts using local covariate/exogenous context and globally retrieved similar series.
Problem
Time series forecasters perform poorly on specific instances due to instance-level variations (distribution shifts, missing data, long-tail patterns), even when average accuracy looks strong.
Approach
A post-forecasting module that predicts the backbone's per-instance MSE as an uncertainty signal, then revises forecasts with a local Transformer over forecasted covariates and exogenous info plus a global retrieval-based weighted average of similar historical series, combined via learned residual gates.
Key results
- PIR reduces average MSE by 8.99% for PatchTST and 25.87% for SparseTSF across 48 settings, and by 3.47% for iTransformer and 2.34% for TimeMixer.
- On PEMS04 with PatchTST, MSE drops from 0.206 to 0.140 (32.04% improvement) when PIR is added, averaged over horizons.
- Estimated per-instance error correlates with actual MSE with R-squared 0.9067 (PatchTST) and 0.7500 (iTransformer) on ETTm1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add PIR post-forecasting revision | PatchTST · Long-term and short-term forecasting benchmarks, 48 experimental settings | MSE reduction: 8.99% | Helped |
| Add PIR post-forecasting revision | SparseTSF · Long-term and short-term forecasting benchmarks | MSE reduction: 25.87% | Helped |
| Add PIR post-forecasting revision | iTransformer · Long-term and short-term forecasting benchmarks | MSE reduction: 3.47% | Helped |
| Add PIR post-forecasting revision | TimeMixer · Long-term and short-term forecasting benchmarks | MSE reduction: 2.34% | Helped |
| Add local revision (ablation) | PIR with PatchTST · Electricity and PEMS03 | MAE: 0.270 -> 0.253; 0.192 -> 0.183 | Helped |
| Add global revision (ablation) | PIR with PatchTST · Electricity and PEMS03 | MAE: 0.257 -> 0.253; 0.201 -> 0.183 | Helped |
| Add the auxiliary uncertainty-error constraint (ablation) | PIR · ETTm1 | forecasting accuracy | Helped |
| Increase backbone depth to match PIR's added layers (ablation) | iTransformer · Electricity and Weather | forecasting accuracy: 0.147 vs 0.145; 0.177 vs 0.170 (MSE, horizon 96) | Mixed |
| Train PIR without a pretrained forecasting backbone (ablation) | PIR · Electricity and Weather | forecasting accuracy: 0.145 -> 0.164; 0.170 -> 0.217 (MSE, horizon 96) | Hurt |
| Replace cosine-similarity retrieval with LSH (ablation) | PIR · ETTh1 and Traffic | inference time: 0.024 -> 0.415; 0.079 -> 87.957 | Hurt |
Surprising findings
- Channel-dependent backbones (iTransformer, TimeMixer) gain less from PIR than channel-independent ones, since they already use covariate information.
- Training PIR from scratch without a pretrained backbone degrades performance noticeably.
Limitations
- Gains are smaller for channel-dependent backbones that already use covariate information.
- Retrieval cost grows with dataset size (LSH on Traffic was 87.96 s), and the authors only suggest sampling or dimensionality reduction as remedies.
- No statistical significance or error bars are reported.
Experimental setup
Datasets: ETT Electricity Solar Weather Traffic PEMS03 PEMS04 PEMS07 PEMS08 Energy (Time-MMD) Health (Time-MMD)
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Electricity Solar Weather Traffic PEMS03 PEMS04 PEMS07 PEMS08
Metrics: MSE MAE
Baselines: PatchTST SparseTSF iTransformer TimeMixer TFT TiDE NBEATSx TimeXer RAFT RATD RATSF
Builds on: PatchTST iTransformer SparseTSF TimeMixer RAFT
Closest papers by meaning
- fev-bench: A Realistic Benchmark for Time Series Forecasting (NeurIPS 2026 · similarity 0.87)
- An Analysis of Linear Time Series Forecasting Models (ICML 2024 · similarity 0.87)
- TimeEmb: A Lightweight Static-Dynamic Disentanglement Framework for Time Series Forecasting (NeurIPS 2025 · similarity 0.87)
- Cherry-Picking in Time Series Forecasting: How to Select Datasets to Make Your Model Shine (AAAI 2025 · similarity 0.87)
- ARM: Refining Multivariate Forecasting with Adaptive Temporal-Contextual Learning (ICLR 2024 · similarity 0.86)
- AutoTimes: Autoregressive Time Series Forecasters via Large Language Models (NeurIPS 2024 · similarity 0.86)
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.