Google quietly open-sourced the prediction model TimesFM — just feed it data and it predicts the future.
Sales curves. Market prices. Web traffic. Energy demand. Crypto volatility.

These seemingly unrelated things share one thing in common: they’re all data that changes over time — time series.

In the past, if you wanted to predict any of these, you either had to hire a statistics PhD to write formulas or spend weeks training a model.

Google Research

💡 What You Will Learn

Google quietly open-sourced the prediction model TimesFM — just feed it data and it predicts the future. Sales curves. Market prices. Web traffic. Energy demand. Crypto volatility. These seemingl

📜 Table of Contents

• 01. A "Time Intuition" Trained on 100 Billion Time Points
• 02. Zero-Shot Prediction: Just Feed It Data
• 03. How Is It Different from Large Language Models?
• 04. Latest 2.5 Version: Smaller, Then Bigger
• 05. What Scenarios Are Most Worth Using?
• 06. Personal Take: Google's "Quiet Release" Strategy Is Worth Watching
• Let's Chat in the Comments

Google Quietly Open-Sources TimesFM: Feed It Data, It Predicts the Future

Sales curves. Market prices. Web traffic. Energy demand. Crypto volatility.

These seemingly unrelated things share one thing in common: they're all data that changes over time — time series.

In the past, predicting these required either a statistics PhD to write formulas or weeks to train a model.

Google Research quietly dropped something on GitHub a couple days ago called TimesFM (Time Series Foundation Model), and it just smashed that barrier to the floor.

It's already racked up 21,000+ Stars on GitHub and blown up.

01. A "Time Intuition" Trained on 100 Billion Time Points

TimesFM isn't some early-stage toy project.

Google trained it on 100 billion real-world time points — mainly from Google Trends search data and Wikipedia page view stats.

What scale are we talking about?

• If recorded once daily, that's 270 million years of continuous data
• Covers almost every domain: e-commerce, finance, energy, social networks
• Learns "the patterns of time" from data, not just memorizing specific scenarios

After all that, TimesFM's model parameters are only 200 million — less than a fraction of a tiny 7B model.

Smaller parameters mean you can run it on an ordinary computer.

02. Zero-Shot Prediction: Just Feed It Data

TimesFM's most insane capability is called zero-shot prediction.

This isn't the "switch scenarios and re-fine-tune" you're used to. It's straight-up: download the model from HuggingFace, give it your data, and it predicts the next 12, 24, or even 1000 steps.

model = timesfm.TimesFM_2p5_200M_torch.from_pretrained(
    "google/timesfm-2.5-200m-pytorch"
)
result = model.forecast(horizon=12, inputs=[your_data])

Four lines of code. Done.

And it doesn't just spit out a single "predicted value." The new 2.5 version adds a quantile prediction head (30M parameters) — it tells you the 10% to 90% confidence range of the forecast. In other words, it knows how uncertain its own predictions are.

Compared to llmtime (powered by GPT-3.5), TimesFM, despite being orders of magnitude smaller, performs better on most benchmarks.

03. How Is It Different from Large Language Models?

Here's a really interesting design choice.

LLMs break text into tokens; TimesFM breaks time series into patches (contiguous blocks of time points). It takes 32 time points as input and predicts 128 time points — the output patch is 4x larger than the input patch.

This is completely different from LLMs, which predict one token at a time.

Google's intuition: time series have patterns that don't need word-by-word generation like writing an article. Looking at 128 steps at once captures long-term trends better.

What does this mean?

Traditional time series models accumulate errors with each step. The longer the output, the more the prediction drifts. TimesFM's "big-step" approach reduces error accumulation at the architectural level.

04. Latest 2.5 Version: Smaller, Then Bigger

In March, TimesFM updated to version 2.5 with several major changes:

• Parameters dropped from 500M to 200M — smaller
• Context length increased from 2048 to 16,000 — can see longer history
• Supports up to 1000-step predictions — paired with the quantile head
• Removed frequency markers — no need to tell the model if data is daily or monthly; it figures it out
• Restored covariate support (XReg) — external factors can also be inputs

In June, the PyPI package was updated to timesfm=2.0.0, making installation easier:

pip install timesfm[torch]

Done.

In May, they added a LoRA fine-tuning tutorial — even though zero-shot is already strong, fine-tuning can make it even better for specific business data.

And now it's embedded in Google BigQuery and Google Sheets — people who don't code can use this model for predictions right in Excel.

05. What Scenarios Are Most Worth Using?

Let's talk about a few real-world scenarios:

Scenario 1: E-commerce Seller You have a year of GMV data and want to know how much inventory to stock next month. Before, you'd spend ages tuning ARIMA or Prophet parameters. Now, just throw it at TimesFM and get a prediction in 1 minute.

Scenario 2: Individual Investor You're tracking a cryptocurrency's historical price. You're not a professional quant team, but you want to see "where it might head next." Run the model locally — no internet needed, no data sent to anyone else.

Scenario 3: Content Creator You're tracking the "search trend" of a breaking news story. TimesFM's training data includes tons of Google Trends data, giving it a natural advantage for this kind of prediction.

My recommendation: If your data volume is small (a few thousand rows or less) and the frequency is daily or hourly, go straight to zero-shot mode. Its performance is already close to or even better than traditional supervised models. If you have large data or a specialized scenario, add LoRA fine-tuning.

06. Personal Take: Google's "Quiet Release" Strategy Is Worth Watching

TimesFM isn't something Google just dropped out of nowhere.

Accepted by ICML in 2024, version 2.5 launched in September 2025, and it's been quietly iterating for over half a year — no press conference, no PR release, no hype-driven marketing. GitHub just silently climbed to 20k stars.

This is very Google's style: TensorFlow, BERT, T5 — all released first in academic circles, let the community blow up, then roll out into commercial products.

Now TimesFM is already in BigQuery, AlloyDB, Sheets — it's no longer just an open-source toy.

But honestly, I think TimesFM's most valuable aspect isn't its accuracy — it's that it democratizes prediction.

For a long time, time series forecasting was the domain of statisticians and data scientists. You needed to know ARIMA, ETS, seasonal decomposition, tune parameters, check residuals, run white noise tests.

TimesFM tells you: download a model, input data, get results.

For most non-core prediction scenarios (spotting trends, rough estimates), this accuracy is good enough. And the truly high-stakes scenarios (risk control, high-frequency trading) wouldn't use a general-purpose model anyway.

Getting 80% of people to save 80% of their time on predictions — that's its biggest value.

Let's Chat in the Comments

What new use cases do you think will emerge once time series prediction is "democratized"?

If you're an e-commerce operator / product manager / data analyst, what scenarios would you use TimesFM for?

Feel free to share in the comments.