The Google research team recently launched TimesFM (Temporal Foundation Model) 2.0, a pre-trained model specifically designed for time series forecasting. This model aims to enhance the accuracy of time series predictions and promote the development of artificial intelligence through open-source and scientific sharing.
TimesFM 2.0 is a powerful model capable of handling univariate time series forecasting with up to 2048 time points and supports arbitrary prediction time spans.
It is worth noting that while the model's training maximum context length is 2048, it can handle longer contexts in practical applications. The model focuses on point predictions and experimentally offers 10 quantile heads, although these have not been calibrated after pre-training.
In terms of data pre-training, TimesFM 2.0 includes a combination of multiple datasets, incorporating the pre-training set from TimesFM 1.0 as well as additional datasets from LOTSA. These datasets cover various fields, such as residential electricity load, solar power generation, traffic flow, and provide a rich foundation for model training.
With TimesFM 2.0, users can more easily perform time series forecasting, advancing various applications, including retail sales, stock trends, website traffic, environmental monitoring, and intelligent transportation.
Model link: https://huggingface.co/google/timesfm-2.0-500m-pytorch
Key Points:
🌟 TimesFM 2.0 is a new temporal forecasting model launched by Google, focused on improving the accuracy of time series predictions.
🔧 The model supports predictions with up to 2048 time points and can handle arbitrary prediction time spans.
📊 Users can freely choose prediction frequencies based on different time series characteristics, enhancing prediction flexibility.
