Google Researchers Rajat Sen and Yichen Zhou discuss the significance of time-series forecasting in various domains. Deep learning (DL) models have been successful in forecasting multivariate, time-series data because they perform well in a variety of settings. They introduce TimesFM, a single forecasting model pre-trained on a large time-series corpus of 100 billion real world time-points. The model adapts to variable context and horizon lengths, and is smaller yet performs close to state-of-the-art supervised approaches. They discuss the need for large volume of legitimate time series data for pretraining and share the zero-shot evaluation results of TimesFM, which performs better than most statistical methods and DL models.
**Introduction to Time-Series Forecasting and Foundation Model for Time-Series Forecasting
Time-series forecasting is crucial in many areas such as retail, finance, manufacturing, healthcare and natural sciences. In this blog post, Google Researchers Rajat Sen and Yichen Zhou talk about the significance of time-series forecasting and introduce a foundation model for time-series forecasting, called TimesFM.
**A Decoder-Only Foundation Model for Time-Series Forecasting
TimesFM is a single forecasting model pre-trained on a large time-series corpus of 100 billion real world time-points. The model adapts to variable context and horizon lengths and is smaller yet performs close to state-of-the-art supervised approaches.
**Pretraining Data for TimesFM
Google Research has spent a significant amount of time and effort creating and assessing training datasets for TimesFM. They discuss the need for a large volume of legitimate time series data for pretraining and share their findings, including the use of synthetic data and real-world data to improve the model’s performance.