Time-Series LLMs, Explained with t0-alpha

Time-series foundation models like t0-alpha operate by converting numerical sequences into patches for processing by a causal transformer. These models generate probabilistic forecasts by emitting a set of quantiles rather than a single point prediction, a method evaluated by metrics like CRPS and MASE. The author reproduces the GIFT-Eval benchmark for t0-alpha, confirming its reported performance and placing it within a competitive cluster of similar models. While t0-alpha outperforms classical baselines and some larger models, the analysis suggests future gains may come from areas beyond simple transformer variations.