Research map / Time series / Causal and generative time series

SteerCast: Retrieval-Based Latent Steering for Decoder-Only Time Series Forecasting

NeurIPS 2026 · Time series › Causal and generative time series · approach: Probabilistic and conformal 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 SteerCast, which stores retrieval keys and latent steering vectors (ground-truth vs. model-predicted continuation differences) from a training set and injects retrieved, gated steering into a fine-tuned decoder-only forecaster's hidden states at inference, without parameter updates.

Problem

Fine-tuned decoder-only time series forecasters still make persistent errors in rare regimes, distribution shifts, and long horizons, and improving them usually requires extra training or scaling; prepending retrieved histories is costly given limited context windows.

Approach

Build a memory of mean-pooled hidden-state keys and per-layer latent steering vectors defined as the difference between hidden states under ground-truth and model-predicted continuations; at test time, average the steering vectors of k nearest neighbors and inject a norm-bounded, cosine-gated update into all layers at every autoregressive step.

Key results

• Under the various look-back setting, SteerCast achieves the lowest or tied-lowest MSE on 25 of 30 dataset-backbone entries and reduces MSE by 7.4% on average relative to RAF.
• Under the fixed look-back (512) setting, SteerCast reduces MSE by 5.1% over FT, 5.3% over RAF, and 2.7% over RAFT on average.
• With Time-MoE on ETTh1, SteerCast average MSE 0.394 vs. 0.402 for RAFT (fixed setting, from averaged tables); output-level ensembling averages 0.613 vs SteerCast 0.542 across nine datasets.

Surprising findings

• Mean-pooled retrieval keys outperform last-token keys at every horizon, contrary to the receptive-field intuition.
• Euclidean retrieval outperforms cosine retrieval, with the gap widening at longer horizons.

Limitations

• Requires storing a memory bank whose size scales with dataset size and channel count (e.g., 39 GB for Traffic).
• Restricted to decoder-only autoregressive forecasters because the injection relies on per-step latent rollout.
• Gains are sometimes absent or negative (e.g., Births under some settings, Saugeen and Sunspots for some backbones), and hyperparameters (k, lambda) are tuned per dataset on validation.

Experimental setup

Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Exchange Weather Illness US Births SaugeenDay Sunspots

Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Exchange Weather Illness US Births SaugeenDay Sunspots

Metrics: MSE MAE

Baselines: Fine-tune (FT) RAF RAFT output-level ensembling

Builds on: in-context vectors RAFT RAF Time-MoE Timer-XL TimesFM

Closest papers by meaning

• TimePerceiver: An Encoder-Decoder Framework for Generalized Time-Series Forecasting (NeurIPS 2025 · similarity 0.89)
• Retrieval Augmented Time Series Forecasting (ICML 2025 · similarity 0.88)
• TS-RAG: Retrieval-Augmented Generation based Time Series Foundation Models are Stronger Zero-Shot Forecaster (NeurIPS 2025 · similarity 0.87)
• Towards a General Time Series Forecasting Model with Unified Representation and Adaptive Transfer (ICML 2025 · similarity 0.86)
• OLinear: A Linear Model for Time Series Forecasting in Orthogonally Transformed Domain (NeurIPS 2025 · similarity 0.86)
• A decoder-only foundation model for time-series forecasting (ICML 2024 · similarity 0.86)