Wednesday 26 March 2025
Researchers have made a significant breakthrough in time series forecasting, a crucial task for predicting future events and trends in various fields such as finance, weather, and healthcare. The new approach, called AdapTS, leverages online feedback to improve forecasts by adapting to changing patterns in the data.
Traditionally, foundation models (FMs) are trained on large datasets and then deployed without further adjustments. However, this can lead to suboptimal performance when faced with novel or dynamic data distributions. AdapTS addresses this issue by learning from up-to-date feedback, allowing FMs to adapt their forecasts in real-time.
The team behind AdapTS used a combination of two key components: the AdapTS-Forecaster and the AdapTS-Weighter. The former is a lightweight forecaster that learns online on each dataset, while the latter combines the forecasts of the FM with those of the AdapTS-Forecaster using a weighted average.
In experiments, AdapTS outperformed traditional FMs in various datasets, including traffic, weather, and cloud usage patterns. For example, when forecasting traffic levels, AdapTS accurately captured daily and weekly periodicities that were missed by the FM alone. Similarly, on cloud data, AdapTS successfully predicted large spikes in usage that the FM failed to anticipate.
AdapTS’s ability to learn from online feedback enables it to pick up on dataset-specific features that may not be well-represented in pre-trained FMs. This is evident in the Traffic dataset, where AdapTS correctly identified a decrease in traffic during weekends, which was overlooked by the FM. In contrast, the AdapTS-Forecaster, trained online on this data, successfully captured this trend.
The implications of AdapTS are far-reaching. By allowing FMs to adapt to changing patterns, it can improve forecasting accuracy and reduce errors in various applications. This is particularly important in fields where accurate predictions can have significant consequences, such as finance or healthcare.
While AdapTS shows promising results, there are still limitations to be addressed. For instance, the approach assumes a rolling window of data, which may not always be feasible in practice. Additionally, the performance of AdapTS can degrade when faced with complex or non-stationary data distributions.
Despite these challenges, AdapTS represents a significant step forward in time series forecasting.
Cite this article: “Adaptive Time Series Forecasting with Online Feedback: A Breakthrough in Predictive Accuracy”, The Science Archive, 2025.
Time Series Forecasting, Online Feedback, Adaptive Learning, Foundation Models, Forecasting Accuracy, Data Distributions, Weighted Average, Periodicities, Traffic Patterns, Cloud Usage.
