Introducing TimesFM-3, a state-of-the-art time series foundation model that enables accurate multivariate time series forecasting in a single forward pass, significantly outperforming other forecasting models across major benchmarks.
More on the blog
I used ChatGPT to help me understand it; I will paste the same below
Forecasting simply means predicting what comes next using what happened before.
Example:
Electricity use for the last 2 years → predict electricity use tomorrow.
Data recorded in chronological order is called a time series.
Examples include stock prices, temperature, sales, electricity consumption, traffic, heart rate, commodity prices, etc.
Suppose a shop wants to predict tomorrow’s sales.
Historically, it would collect its sales data and build a mathematical/statistical model specifically for that data:
Past sales → ARIMA / statistical model → tomorrow’s sales.
Later, machine learning and deep learning models such as XGBoost, LSTMs and Transformers became popular.
They could consider much more information:
Sales + weather + promotions + holidays → ML model → future sales.
But there was a major inconvenience:
You generally had to train a model specifically for your problem.
That means collecting data, cleaning it, training the model, testing it and tuning it.
This is similar to the idea behind ChatGPT.
Instead of every person training their own language model, a huge model is pretrained on enormous amounts of data.
You then give it a new problem it hasn’t specifically been trained for.
Google applied this idea to time-series forecasting with TimesFM.
Your data → already-trained TimesFM → forecast
When it can forecast new data without being specifically trained on that dataset, this is called zero-shot forecasting.
TimesFM-3 is Google’s latest time-series foundation model.
The major advancement is that it is multivariate.
Previous TimesFM models mainly worked like:
Past crude-oil prices → predict crude oil
TimesFM-3 can instead consider related variables together:
Crude oil
Brent
Natural gas
Dollar index
Inventory
Volatility
↓
TimesFM-3
↓
Future forecasts
It tries to understand both how each variable changes through time AND how the variables relate to each other.
Target: what you want predicted — e.g. crude-oil price.
Past information: things you historically observed — e.g. volume or inventory.
Known future information: things you already know will happen — e.g. holidays, scheduled promotions or calendar effects.
It can also predict several targets together.
Forecasting isn’t:
“Oil will definitely be $75.”
TimesFM-3 produces probability ranges.
For example:
| Forecast | Price* |
|---|---|
| Lower scenario | $71 |
| Most likely/median | $75 |
| Upper scenario | $81 |
*Illustrative numbers.
Technically, it outputs nine quantiles from the 10th to 90th percentile.
This is valuable because real-world forecasting is inherently uncertain.
TimesFM-3 has 330 million parameters and was pretrained on more than 1 trillion real and synthetic time points.
It also introduces Contiguous Patch Masking, allowing it to predict the whole requested future horizon in a single forward pass rather than repeatedly predicting small chunks.
Google reports strong results on major forecasting benchmarks, although benchmark superiority does not mean it will automatically predict financial markets profitably.
| Industry | Example |
|---|---|
| Finance | Prices, volatility, volume |
| Retail | Future sales/demand |
| Energy | Electricity demand/prices |
| Manufacturing | Production/equipment |
| Logistics | Inventory/demand |
| IT | Server/network load |
| Healthcare | Patient measurements |
| Science | Physical/environmental measurements |
Traditional statistics
Your data → build mathematical model → forecast
↓
Machine learning
Lots of variables → train your own ML model → forecast
↓
Foundation forecasting models
Huge pretrained model → give it your new data → forecast immediately
↓
TimesFM-3
Multiple related time series → pretrained model → multiple future predictions + probability ranges, without training a new model for your dataset.
That’s the core significance of TimesFM-3.