An Actor-Critic Method For Forecasting Collections Of Related Time Series

Description

Forecasting extensive collections of related time series (TS) is a relevant task in various domains, e.g. economic or financial predictions, public forecasting, demand models, and consumption plans. More often than not, these problems involve handling high-dimensional multivariate series which relate to each other through potentially complex patterns. Exploiting such global patterns while keeping good statistical accuracy and expressiveness level for locally calibrated predictions (i.e. each individual series) has been a widely studied objective in recent years. The literature already proposes different hybrid models, either combining the benefits of classical TS models (e.g. ARIMA) with deep learning (DL) approaches or combining more complex temporal networks, which are decomposed and regularised differently to capture local and global properties effectively. However, most of these methods are either one-dimensional or are inadequate to adapt fast to changing data distributions or relevant contextual dimensions for the real-world (e.g. cost factors). To circumvent this, we propose a new deep reinforcement learning (DRL) method – T2f. Our approach builds on traditional data-driven ensemble learning fundamentals and innovates by employing an actor-critic method based on the Twin Delayed Deep Deterministic (TD3) algorithm. T2f formulates the prediction task based on a two-step input sequence, which starts by learning a local policy from which local properties are extracted as base model features. The second step layer learns a deep factor based on the global policy, which takes these features as latent variables to control for hidden features in the time series collection. The actor’s output layer then generates the combination weights for the two input steps. We present succinct theoretical and empirical evidence for our proposal by putting it against both global and local state-of-the-art methods. This is done on two datasets, one artificial and a large retail sales report. Our results demonstrate the soundness of our solution in terms of accuracy and efficiency in balancing local and global signals dynamically. The latter is achieved in both fast adaptation to changes in the distribution and in consideration of contextually relevant dimensions.

Period	27 Jun 2023
Event title	43th International Symposium on Forecasting 2023
Event type	Conference
Conference number	43
Location	Charlottesville, United States, VirginiaShow on map
Degree of Recognition	International