Toto-2.0: Advancements in Time Series Forecasting

nkafr · reddit · 2026-07-09

The post introduces Toto-2.0, a multivariate time series forecasting model released by Datadog research, emphasizing its goal to scale time series forecasting "like an LLM." Key designs include replacing autoregressive decoding with Contiguous Patch Masking, using arcsinh normalization to preserve small fluctuations, and a NorMuon optimizer better suited for pinball loss.

The post also mentions the model uses u-µP hyperparameter transfer: tuning on a 10-million parameter proxy model before transferring to multiple target scales. Links for further discussion and tutorials are provided.

Original post →

More from Research

Research channel →