Frozen time-series forecasters improve by updating memories regularly
Retrieval Is Not Enough: Refreshing Memory for Frozen Time-Series Forecasters
Machine Learning
Summary
Time-series forecasting tries to predict future data points based on past observations. The paper shows that using a fixed memory of past data collected before training can limit prediction quality because it ignores new observations after deployment. The authors propose FreshCast, a method that keeps the forecasting model fixed but continuously refreshes the memory with recent data and adjusts how much to trust this memory. This approach improves prediction accuracy by making sure the memory stays relevant and corrects specific forecasting errors.
What this means in practice
- •For demand planners: Improve sales and inventory predictions by continuously updating historical reference data without retraining models.
- •For energy load forecasters: Maintain more accurate electricity consumption forecasts by incorporating recent observations into memory for frozen forecasting models.
Authors
Chao He, Jianyu Xu, Xinyi Guo, Ruiqi Liu, Haobin Ding, Ruiqi He, Dongqing Song
Abstract
Retrieval-augmented time-series forecasting uses the continuations of historical segments similar to the current context as references for a forecaster. Most existing methods build the retrieval memory once from the training segment, leaving observations revealed after deployment unavailable as references, and generally do not calibrate how much the retrieved information should influence a frozen forecaster. We identify two key determinants of retrieval utility for a frozen forecaster: whether the history still reflects the current state, and whether the correction it induces aligns with the forecaster's residual errors, an alignment that can shift between validation and deployment when the memory becomes stale. We propose FreshCast, a plug-in retrieval framework that keeps the forecaster frozen, continuously updates a non-parametric memory with new observations, forms a memory forecast through relational kernel regression, and calibrates its weight in closed form on the validation segment. Under a simplified generative model, we characterize the optimal combination gain through the second-order relation between forecaster error and memory correction, and show that a sufficiently long look-back can make periodic memory information redundant. Across seven benchmarks and ten forecasting architectures, FreshCast reduces average MSE for every evaluated forecaster and input length, by 14.6% and 5.6% at input lengths 96 and 720, and achieves lower MSE than the evaluated retrieval-augmented and online baselines in their comparison settings. Ablations show that freezing the memory at the end of training removes most of the gain, identifying post-training observations as a primary source of improvement. For a frozen forecaster, useful historical references must remain timely and provide information that helps correct its remaining errors.