How Time Series Forecasting Models Are Trained: Moirai and Contrastive Predictive Coding, Hands-On

Time series foundation models such as Moirai are pretrained on large collections of series from many domains and then used as zero-shot forecasters. This hands-on workshop shows how such models are trained. Working in Google Colab, participants train a small-scale version of Moirai (Woo et al., 2024) and a small-scale Contrastive Predictive Coding model (van den Oord et al., 2018), which learns representations of time series by predicting the future in latent space. Along the way, we cover the main ingredients of training, from data preparation to probabilistic evaluation, with GIFT-Eval (Aksu et al., 2024) as the reference benchmark. Participants leave with notebooks they can reuse and extend. Prerequisites: Python and PyTorch basics. Attendees bring a laptop; everything runs in Google Colab.