Google's TimesFM-3 finally natively forecasts many time series at once, zero-shot
Google's time-series foundation models were univariate until now, which is exactly backwards for real forecasting. TimesFM-3 (330M params, pretrained on 1T+ time points) fixes that with native zero-shot multivariate support, jointly predicting coevolving series, including a lookahead trick so the model can peek at known future covariates like promotions and weather. No fine-tuning required, one forward pass on retail, finance, observability and healthcare workloads.