Time-Series Foundation Model
A model pretrained specifically on large-scale time-series corpora. It processes continuous numerical sequences with patch-based input and outputs future numerical forecasts, with zero-shot or few-shot forecasting capability.
Timeline (last 90 days)
In zero-shot mode, time-series foundation models cannot use market-specific structural information such as trading volume, macroeconomic events or market sentiment.
The models currently ranking at the top of major time-series leaderboards (such as Kronos-2 and TimeSafe M2.5) are all time-series foundation models, not approaches that stuff an LLM into the forecasting pipeline.
A combination in which the time-series foundation model handles numerical forecasting and the LLM reads the relevant text, with the two results combined by weighted averaging, outperformed the strongest single time-series foundation model in the experiments.
Time-series foundation models have zero-shot or few-shot forecasting capability and avoid the information loss caused by serializing numerical values into text.
The portion of time-series model input exceeding the context window is directly truncated rather than processed in chunks or raising an error; truncation occurs after tokenization and before model computation.
Even with the correct context length, time-series foundation models still struggle with distribution shift: if the training data does not contain future states, the model will interpolate confidently but may be wildly wrong; a longer context window cannot make up for this flaw.
Time-series foundation models like Mistral's Timestral process input within a fixed context window; when the sequence length exceeds the window, the model simply discards the oldest tokens, all without any error, warning or logging.
All Facts (7)
In zero-shot mode, time-series foundation models cannot use market-specific structural information such as trading volume, macroeconomic events or market sentiment.
50%UnverifiedThe models currently ranking at the top of major time-series leaderboards (such as Kronos-2 and TimeSafe M2.5) are all time-series foundation models, not approaches that stuff an LLM into the forecasting pipeline.
50%UnverifiedA combination in which the time-series foundation model handles numerical forecasting and the LLM reads the relevant text, with the two results combined by weighted averaging, outperformed the strongest single time-series foundation model in the experiments.
50%UnverifiedTime-series foundation models have zero-shot or few-shot forecasting capability and avoid the information loss caused by serializing numerical values into text.
50%UnverifiedThe portion of time-series model input exceeding the context window is directly truncated rather than processed in chunks or raising an error; truncation occurs after tokenization and before model computation.
50%UnverifiedEven with the correct context length, time-series foundation models still struggle with distribution shift: if the training data does not contain future states, the model will interpolate confidently but may be wildly wrong; a longer context window cannot make up for this flaw.
50%UnverifiedTime-series foundation models like Mistral's Timestral process input within a fixed context window; when the sequence length exceeds the window, the model simply discards the oldest tokens, all without any error, warning or logging.
50%