Canonical: https://instain.ru/en/ai/google-timesfm/ Language: en Google Research the Time Series Forecast Model # TimesFM. Forecasting from time series history with backtesting 4.38 / 5 [Directory score](https://instain.ru/en/ai/google-timesfm/#editorial-rating) [Official website ](https://github.com/google-research/timesfm) The model receives the historical values of the series and builds a forecast for the selected horizon. For some versions, work with additional features and quantiles is documented. ![Google Research Repository with TimesFM Forecasting Examples](https://instain.ru/assets/images/google-timesfm-screen-90f5458337c2-1200.webp) Google Research repository with examples of TimesFM forecasting. [Open the source ](https://github.com/google-research/timesfm) ## What information was known on the day of the forecast A holiday calendar known to a shop in advance can help forecast weekly sales. A future promotion's final report cannot be included in a historical forecast. TimesFM forecasts from history, and some versions also accept covariates. Before testing, distinguish variables observed only in the past from ones known across the forecast horizon. Avoiding future data leakage matters more than tuning the model, because leakage makes evaluation look too optimistic. ## Training run on several dates Take a number of daily sales and fix several past dates on which you need to reproduce the forecast. For each date, cut off all subsequent sales and keep the horizon the same. Compare the TimesFM result with the number for the same day the previous week. Then check if adding a known calendar changes the error separately for normal weeks and holiday periods. ## Interval is necessary for solution If the version used returns quantiles, see how often the actual values fall between the selected boundaries. Too wide an interval can include almost all days, but it does little to plan a stock. Check if the spread on the rows with gaps or a sharp change in demand is growing. Save the version of the weights, the method of handling the passes and the cut-off point of the row. The rerun must use the same information that was available at each point of the test. Write down separately periods where the calendar did not help. ## Suitable tasks - Compare weekly sales forecasts for several past dates - Check the influence of a calendar known in advance Features ## Strengths - Have pre-trained weights - The documentation contains forecast options and examples of work with signs Things to consider ## Limitations - Features that use future information distort evaluation results - Check accuracy and intervals on your own series ## Selection and conditions Free access Yes Trial period Not confirmed No sign-up required Not confirmed Russian interface Not confirmed Local installation Yes API Not confirmed “Not confirmed” means there is not enough information. The feature may still be available; check the developer’s terms. Input formats Data Result Data ### Terms Code and weights are available under the terms in the corresponding Google publications. Running the model yourself requires compute resources. Features that use future information distort evaluation results. From task to first result ## Practical example Task Compare weekly sales forecasts for several past dates What to prepare The training range of daily sales and the calendar of holidays, known on the forecast date. ### Task brief or prompt Make a weekly forecast for several past dates using only the information available for each date. ### Expected result A table of forecasts and facts with error, intervals and a simple seasonal benchmark. ### How to check the result - Check the absence of future data at the entrance. - Separately compare the usual weeks and periods of shares. An editorial teaching example of a task brief, not the result of hands-on testing of the service. Check that the required mode is available, that you have rights to the source material and that the intended use of the result is permitted. Check details ## Official sources - [Official Google Research Repository ](https://github.com/google-research/timesfm) - [TimesFM Weight Card ](https://huggingface.co/google/timesfm-3.0-pytorch) This description is based on the developer’s material. Features, prices and availability may change; check the primary source before use. About the tool Developer Google Research Category [Code](https://instain.ru/en/category/code/) Access For local use, choose model weights, a compatible environment, and data with a known observation frequency. The interface varies by model version. Payment Code and weights are available under the terms in the corresponding Google publications. Running the model yourself requires compute resources. API No separate API specified [Developer terms ](https://huggingface.co/google/timesfm-3.0-pytorch) Directory score 4.38 / 5 Features 4.50 / 5 Workflows 4.20 / 5 Ecosystem 4.30 / 5 Documentation 4.50 / 5 Scores are an editorial guide to the directory. They are not user reviews or the results of personal testing. [See the methodology ](https://instain.ru/en/methodology/) Keep exploring ## For similar tasks [Full category ](https://instain.ru/en/category/code/) Code ### [Amazon Chronos](https://instain.ru/en/ai/amazon-chronos/) Open Amazon Science models for numerical time series 4.30 / 5 the Time Series Forecast Model Similar tasks; compare terms, formats and limitations