# AI use cases for Timesfm (google-research/timesfm) Repo: [google-research/timesfm](https://github.com/google-research/timesfm) | Use cases: 3 | Page: https://dedupedai.com/use-cases/google-research-timesfm/ TimesFM is Google Research's pretrained time series foundation model. Version 3.0 forecasts univariate and multivariate series zero shot, supports past only and past and future covariates, and returns point forecasts plus nine quantiles. It installs from PyPI with a PyTorch backend or an MLX backend for Apple silicon, and the repo includes a LoRA fine-tuning example. This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL. - Forecast a series without training anything: https://dedupedai.com/use-cases/google-research-timesfm/forecast-a-series-zero-shot-with-timesfm/ - Forecast several related series with drivers: https://dedupedai.com/use-cases/google-research-timesfm/forecast-with-covariates-using-timesfm/ - Run forecasts on a Mac with no GPU server: https://dedupedai.com/use-cases/google-research-timesfm/run-timesfm-on-apple-silicon-with-mlx/ ## How to forecast a series without training anything with Timesfm Repo: [google-research/timesfm](https://github.com/google-research/timesfm) | Difficulty: Intermediate | Time: 45 minutes | Basis: From the README Load the TimesFM 3.0 checkpoint and predict a horizon with quantiles straight from your history. **Who it is for:** An analyst who needs a forecast today and has no model. **The problem:** Fitting a model per series is slow and needs tuning. **The outcome:** Point forecasts and nine quantiles for each series you pass. ### Steps 1. Install it with `pip install timesfm[torch]` 2. Create a ModelConfig with checkpoint_path set to google/timesfm-3.0-pytorch and build a TimesFM3Evaluator as in the README 3. Call forecaster.predict_batch with your list of 1D NumPy arrays, a horizon and return_quantiles=True 4. Read outputs[0].forecast and outputs[0].quantiles, which have shapes like (12,) and (12, 9) ### Prompt to paste into your agent ```text Using google-research/timesfm, install `pip install timesfm[torch]`, forecast the next 12 steps of each series in with TimesFM3Evaluator and return_quantiles, and write the medians and deciles to a new CSV. ``` ### Caveat Weights up to version 2.5 remain Apache 2.0 if you need a permissive option. Risk, platform: TimesFM 3.0 pretrained weights are under a non-commercial license and the README says commercial or production use of self-hosted weights is not permitted. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/google-research-timesfm/forecast-a-series-zero-shot-with-timesfm/ ## How to forecast several related series with drivers with Timesfm Repo: [google-research/timesfm](https://github.com/google-research/timesfm) | Difficulty: Intermediate | Time: 1 hour | Basis: From the README Give TimesFM 3.0 multiple target variates plus past only and past and future covariates in one call. **Who it is for:** A planner whose demand depends on price, weather or promotions. **The problem:** Univariate models ignore the drivers you already know. **The outcome:** A joint forecast across all target variates, with quantiles. ### Steps 1. Install it with `pip install timesfm[torch]` 2. Shape your targets as a 2D array of (num_variates, context_length) as the README shows 3. Pass past_only_covariates and past_future_covariates to predict_batch alongside contexts and horizon 4. Check that outputs[0].forecast has shape (num_variates, horizon) and quantiles adds the nine deciles ### Prompt to paste into your agent ```text With google-research/timesfm, forecast for 24 steps using my known promotions as a past and future covariate, and report the median and the 10th and 90th percentiles. ``` ### Caveat Past and future covariates must cover context plus horizon, as in the README shapes. Risk, platform: The 3.0 weights are non-commercial; production use goes through Google Cloud services per the README. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/google-research-timesfm/forecast-with-covariates-using-timesfm/ ## How to run forecasts on a Mac with no GPU server with Timesfm Repo: [google-research/timesfm](https://github.com/google-research/timesfm) | Difficulty: Intermediate | Time: 45 minutes | Basis: From the README Use the MLX native backend to run TimesFM 3.0 locally on Apple silicon without PyTorch. **Who it is for:** An analyst on a MacBook who does not want a cloud GPU. **The problem:** The default path assumes CUDA, which a laptop does not have. **The outcome:** Local forecasts that match the PyTorch backend closely, at laptop speed. ### Steps 1. Install the MLX extra with `pip install timesfm[mlx]` 2. Create the forecaster with TimesFM3Forecaster.from_pretrained on google/timesfm-3.0-pytorch from timesfm3.mlx 3. Call predict with your context, a horizon and return_quantiles=True, or predict_batch for many series in one pass 4. Compare the shapes and values against the PyTorch backend numbers quoted in the README ### Prompt to paste into your agent ```text Using google-research/timesfm, set up the MLX backend on my Mac with `pip install timesfm[mlx]` and forecast 128 steps ahead with quantiles, reporting how long it takes. ``` ### Caveat The published benchmark numbers are for the 330M model on an Apple M4 Max, so your speed will differ. Risk, platform: Contexts longer than the global context of 15,360 are truncated to their most recent points before decoding. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/google-research-timesfm/run-timesfm-on-apple-silicon-with-mlx/