A Time Series Foundation Model (TSFM) takes in-context time series input to produce a forecast for it. Generic TSFM achieves this by embedding input time series using encoding map and applying a decoding map to produce forecast, where encoding-decoding maps are learnt using pre-training data. The choice of architecture, pre-training data utilized determine the out-of-sample model performance. Existing TSFMs typically require a rich pre-training dataset with large architecture leading to compute intensive trainingas well as inference. In this work, we are motivated to develop a compute efficient architecture for TSFM without sacrificing performance. Towards that, we introduce anembedding of time series that maps them, in a lossless manner, to a unit square (i.e [0, 1]^2) or equivalently a 2D image: it's intuitive and simple, i.e.computationally straightforward; it's universal, i.e. does not depend on anytraining data and works for any time series. With this embedding, we introduce an architecture where time series foundation model naturally maps to traditional supervised learning problem. The architecture has flexibility, in addition, of combining any of the existing models to produce a TSFM that is computationally efficient without sacrificing performance. The model architecture also provides ability totradeoff accuracy with compute cost of inference. We evaluate its performance in terms of forecasting accuracy as well as computational cost and compare with respect to an existing pre-trained TSFM to conclude that it achieves computational efficiency without sacrificing performance.
Biography: Devavrat Shah is an Andrew (1956) and Erna Viterbi Professor of Electrical Engineering and Computer Science at MIT. In 2013, he co-founded AI start-up Celect (part of Nike) to help retailers with allocation and fulfillment. In 2019, he is co-founder of Ikigai Labs (part of Celonis) whose mission is to transform enterprise forecasting and planning with AI. At MIT, he was the faculty director of Deshpande Center for Tech Innovation between 2020 to 2023, founding director of the Statistics and Data Science Center at MIT between2016 to 2020. He is a distinguished alumnus of his alma mater IIT Bombay. His work has been recognized through career prizes 2008 ACM Sigmetrics Rising Star, 2010 INFORMS Erlang Prize, 2024 INFORMS Applied Probability Society Markov Lecture, 2025 ACM Sigmetrics Achievement Award; paper prizes at IEEE Infocom, ACM Sigmetrics, NeurIPS, INFORMS Applied Probability Society, INFORM Management Science and Operations Management; INFORMS George B Dantzig thesis prize and test of time awards at ACM Sigmetrics. Heis Kavli Fellow of National Academy of Sciences. He was an adjunct professorat Tata Institute of Fundamental Research, Mumbai and currently adistinguished professor of Data Science and AI at Indian Institute of Technology, Guwahati.
