Summary
Forecasting future values of a time series is hard when there’s little or no past data available, a problem known as cold-start forecasting. The authors treat this as a quick learning challenge, using a method called Temporal Model-Agnostic Meta-Learning (TMAML) adapted from another field to quickly learn from short historical windows. TMAML trains forecasting models on small, time-consistent data chunks to better handle new series with limited history. Experiments show TMAML generally provides better or equally good forecasts and improved confidence calibration compared to standard training approaches, although it does not always yield better probabilistic accuracy than simple methods.
What this means in practice
- •For inventory planners: Generate forecasts for new products with minimal historical sales data using models that adapt quickly to short time series.
- •For energy demand analysts: Produce better-calibrated short-term forecasts for newly installed or recently changed measurement sites with limited data histories.
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.