Google Open‑Sources TimesFM: A Foundation Model for Plug‑and‑Play Time‑Series Forecasting

Overview: Google Research open‑sourced TimesFM, a foundation model for time‑series forecasting that works out‑of‑the‑box on many prediction tasks without task‑specific retraining.

Why TimesFM matters

TimesFM is pre‑trained on massive, diverse public time‑series datasets, learning generic temporal patterns. When a new sales or traffic series is presented, the model can be used directly for inference, often matching or surpassing custom‑built models.

Technical architecture and key improvements

TimesFM uses a decoder‑only Transformer architecture, similar to large language models such as GPT, which generates future sequences from historical context. The latest 2.5 release introduces three notable upgrades:

Smaller yet stronger: Parameters reduced from 500 M to 200 M while context length increased from 2 048 to 16 000 tokens, enabling longer history handling.

Continuous quantile prediction: An optional 30 M‑parameter quantile head outputs predictions for percentiles from 10 % to 90 %, providing uncertainty estimates useful for risk assessment.

Simplified usage: The model no longer requires an explicit “frequency” label, allowing it to infer data patterns autonomously.

“We demonstrate that a single time‑series forecasting model can achieve strong zero‑shot performance across a wide range of datasets, comparable to the best models trained individually for each dataset.” – TimesFM paper

Quick start: forecasting in five minutes

With a clear API and Hugging Face hosting, TimesFM can be run with a few commands. First clone the repository and install dependencies (the example uses the uv package manager):

git clone https://github.com/google-research/timesfm.git
cd timesfm
uv venv
source .venv/bin/activate
uv pip install -e .[torch]

Then load the pretrained model and run a forecast with a few lines of Python:

import timesfm
import numpy as np

# Load the pretrained model
model = timesfm.TimesFM_2p5_200M_torch.from_pretrained(
    "google/timesfm-2.5-200m-pytorch"
)

# Configure forecasting parameters
model.compile(
    timesfm.ForecastConfig(
        max_context=1024,
        max_horizon=256,
        use_continuous_quantile_head=True,  # enable quantile prediction
    )
)

# Example input series
point_forecast, quantile_forecast = model.forecast(
    horizon=12,
    inputs=[
        np.linspace(0, 1, 100),               # series 1
        np.sin(np.linspace(0, 20, 67)),       # series 2
    ],
)

# point_forecast.shape -> (2, 12)
# quantile_forecast.shape -> (2, 12, 10)

Applicable scenarios and target audience

TimesFM is well‑suited for rapid prototyping or handling many heterogeneous forecasting tasks, such as retail sales, energy load, financial indicators, network monitoring, and IoT sensor data.

Data scientists / analysts: Need a strong baseline model to validate ideas quickly without extensive hyperparameter tuning.

Full‑stack / backend developers: Want to embed forecasting capabilities in products without deep time‑series expertise.

Researchers and students: Want to study state‑of‑the‑art foundation‑model techniques and build on them.

Conclusion and outlook

Google’s open‑source TimesFM marks the arrival of foundation‑model approaches in time‑series analysis, bringing unprecedented convenience and strong zero‑shot performance. Ongoing development plans include a faster Flax implementation and restored covariate support. For anyone working on prediction problems, TimesFM is a heavyweight project worth adding to the toolbox.