Dynamic model selection in enterprise forecasting systems using sequence modeling

Abstract

Enterprise forecasting systems often involve modeling a large scale of heterogeneous time series using a pool of candidate algorithms, such as in the case of simultaneous sales forecasts of thousands of stock-keeping units. In such cases, it can be advantageous to automatically monitor and replace algorithms for each time series. We introduce TimeSpeaks, a framework that adapts sequence modeling in natural language processing to the problem of dynamic model selection in enterprise forecasting. We instantiate our framework using sequential (BiLSTM) and transformer-based (TimeXer) deep learning models to learn the temporal dependencies between candidate algorithms. We compare the performance of our framework with state-of-the-art forecasting models using two public benchmarking datasets. We further demonstrate its practical application on two retail case studies, while comparing them to alternative model selection scenarios. TimeSpeaks has superior predictive performance and scalability across different scenarios and datasets. Its ability to adapt to evolving data patterns and its minimal reliance on exogenous information make TimeSpeaks a suitable framework for large-scale enterprise forecasting applications.