TMAML: Temporal Model-Agnostic Meta-Learning for Cold-Start Time Series Forecasting

2026-10-03

AI Summary

This study addresses the generalization challenge in cold-start time series forecasting caused by scarce historical data by proposing the TMAML framework. This work introduces Model-Agnostic Meta-Learning (MAML) into time series forecasting, integrating it with the Temporal Fusion Transformer (TFT). By constructing temporally consistent meta-task windows and employing a support-query mechanism, the framework enables rapid adaptation in few-shot scenarios. As the first temporal meta-learning framework tailored for cold-start settings, experimental results demonstrate that TMAML outperforms standard Empirical Risk Minimization (ERM) on TFT in both point prediction accuracy and calibration. However, improvements in probabilistic forecasting remain limited.

Abstract

Cold-start forecasting, the task of forecasting a time series with little to no historical data, is a common challenge. Addressing it requires approaches that learn quickly from few datapoints and leverage information from related series, typically through static covariates, to generalize well to unseen series. While some global forecasting models can generate cold-start predictions by leveraging information shared across multiple series, they are not optimized for out-of-train-set generalization or adaptation from short histories. In this work, we formulate cold-start forecasting as a few-window learning problem and introduce Temporal Model-Agnostic Meta-Learning (TMAML), which tailors the model-agnostic meta-learning algorithm, originally developed for few-shot adaptation of neural networks, to deep time series forecasting. TMAML constructs meta-tasks as temporally consistent support-query windows and pairs them with a temporal meta-training and meta-testing procedure, yielding forecasting models that are explicitly optimized for cold-start forecasting. We instantiate TMAML on the Temporal Fusion Transformer (TFT) and present an initial empirical analysis of forecast accuracy and calibration across three cold-start scenarios: TMAML consistently outperforms or matches a standard ERM-trained TFT, yields better-calibrated forecasts than naive on two of the three scenarios, but does not consistently outperform naive on probabilistic forecast accuracy.

Problem

• Cold-start forecasting
• Time series forecasting
• Few-shot learning
• Out-of-distribution generalization

Innovation

• Cold-start forecasting
• Meta-learning
• Few-window learning
• Time series forecasting
• Temporal Fusion Transformer