What Is TimesFM 3? Google's Time Series Forecasting Model Explained
TimesFM 3 is Google's zero-shot time series forecasting model. Here's how its covariate-aware forecasting works and why it matters for builders.
What is TimesFM 3?
TimesFM 3 is a time series foundation model built by Google Research. It’s a 330 million parameter decoder only transformer, pre-trained on more than a trillion real and synthetic time points, designed to forecast new time series data zero shot. That means no training or fine-tuning on your specific dataset. You feed it a sequence of past values and it produces a forecast, the same way a large language model completes text without being retrained on your particular sentence.
The model is the latest in Google’s TimesFM line, and its main upgrade over earlier versions is that it can take future known events, called covariates, into account when forecasting, not just the historical pattern of the numbers themselves.
TL;DR
- TimesFM 3 is a 330 million parameter decoder only transformer from Google Research, pre-trained on over a trillion time points for zero-shot forecasting.
- It supports covariate-aware forecasting, meaning it can factor in known future events (promotions, holidays, weather flags) rather than just extrapolating from past values alone.
- A demo forecasting daily active users around a Pride season event showed the blind univariate forecast completely missing the spike, while the covariate-informed forecast tracked the actual surge closely.
- In that same test, forecast error (measured as MAE) dropped roughly threefold, from an error of about 3,000 users a day down to about 1,000, just by adding one extra signal column.
- The model is small enough to run on modest hardware, using well under 2 GB of VRAM in the demo, and can even run on CPU.
- It’s available through Hugging Face and can be installed from source, making it accessible to anyone comfortable with a basic Python and GPU (or CPU) setup.
- Use cases span retail demand, energy load, staffing, and web traffic forecasting, essentially anywhere you know something concrete about the future ahead of time.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
How does covariate-aware forecasting work?
Traditional time series forecasting, the “univariate” approach, looks only at a sequence of past numbers and tries to project them forward. It can pick up patterns like weekly cycles or general trends, but it has no way of knowing that a holiday sale, a marketing campaign, or a public event is about to happen. If that event isn’t reflected in the historical pattern already, the model has no way to anticipate it.
Covariates fix this by giving the model an extra signal alongside the main series. In the demo covered here, the primary series was daily active users on a dating app, and the covariate was a simple flag marking which days fell during a Pride season promotional window. The model could see, ahead of time, which future days were “special,” even though it had never seen this specific spike happen before.
The architectural idea is straightforward: the model takes in the historical series plus the future-known flag as an additional input channel, and it adjusts its forecast for the periods where that flag is active. This is different from feeding in more historical variables that also need forecasting themselves. A covariate here is something you already know the answer to in advance, like a scheduled promotion date or a holiday calendar, so the model can use it directly rather than having to predict it.
What did the demo actually show?
The video walked through a synthetic dataset representing daily active users on a dating app in the lead-up to and during Sydney’s Pride season. Two forecasts were generated from the same model and the same historical data:
The first was blind, using only past user counts. The model correctly learned the underlying weekly rhythm of app usage, but had no information about the upcoming event. Its forecast line stayed flat through the entire Pride window, missing the surge entirely.
The second forecast used the identical history, plus one additional column: a binary flag marking which days were part of the Pride event. With that single extra signal, the model’s forecast climbed into the surge and tracked the actual observed values closely.
The difference in accuracy was measured using mean absolute error (MAE). The blind forecast was off by roughly 3,000 users a day during the event window. The covariate-informed forecast cut that error down to about 1,000 users a day, a threefold improvement from one added column of information.
Why does zero-shot forecasting matter?
Most forecasting workflows historically required training a model on your specific dataset, tuning it to the shape and seasonality of your particular series, and retraining whenever conditions changed. That’s a real cost in time and engineering effort, especially for teams that need forecasts across many different series (product lines, regions, sensors) that don’t each justify a custom model.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
A zero-shot foundation model like TimesFM 3 changes that equation. Because it was pre-trained on a very large and diverse set of real and synthetic time series, it has already learned general patterns common to time series data such as trends, seasonality, and noise. You can point it at a new dataset it has never seen and get a usable forecast immediately, without any training step. This mirrors the shift large language models brought to text: instead of building a custom model per task, one general-purpose pre-trained model handles many tasks out of the box.
Is TimesFM 3 practical to run yourself?
Yes, and notably it doesn’t require heavy hardware. In the demonstration, running TimesFM 3 with a covariate-based forecast consumed under 2 GB of VRAM on a GPU, and it’s small enough to run on CPU as well, including on a laptop. That’s a meaningful difference from the assumption that any transformer-based foundation model needs a serious GPU cluster to be usable.
Getting started involves cloning the model’s repository, installing prerequisites (the demo used the uv Python package manager), and downloading the pretrained weights from Hugging Face. From there, running a forecast is a matter of loading a CSV with your historical series (and any covariate columns you want to include), splitting it into a history window and a holdout window to test against, and calling the model.
What kinds of problems fit this model?
The pattern that makes TimesFM 3 useful is any situation where you have historical numeric data recorded over time, and you also know something concrete about the future that would affect that data. Examples raised in the source material include retail demand forecasting around promotions, energy load forecasting around known weather patterns, staffing forecasts around scheduled events, and web traffic forecasting around planned campaigns or holidays.
The common thread: if you can flag a future date or period ahead of time, whether it’s a holiday, a scheduled promotion, or an expected weather event, TimesFM 3 can incorporate that flag directly into its forecast without any additional training. The gap between a forecast that “knows” versus one that doesn’t is exactly the gap the model is built to close.
Frequently Asked Questions
What is TimesFM 3 in simple terms?
It’s a Google Research model that predicts future values in a time series (like daily sales, app usage, or energy demand) without needing to be trained on your specific data first. It can also incorporate known future events to sharpen those predictions.
How is TimesFM 3 different from earlier forecasting models?
Its key upgrade is covariate-aware forecasting: the ability to factor in future known events (like a scheduled promotion or holiday) as an extra signal, rather than relying purely on historical patterns in the numbers themselves.
Do I need a powerful GPU to run TimesFM 3?
No. In testing, the model used under 2 GB of VRAM and can also run on CPU, meaning even a modest machine or laptop can handle it.
What does “zero-shot” mean for this model?
It means the model can generate a forecast for a brand new dataset it has never seen before, without any fine-tuning or retraining step. You simply feed it your historical data and get a forecast back.
What kinds of covariates can I use with TimesFM 3?
Anything you know about the future ahead of time and can represent as a signal alongside your main series, such as promotion dates, holidays, scheduled events, or known weather patterns.
