GitHub Repos intermediate 3 min read May 27, 2026
Public Preview Sign in free for the full digest →

TimesFM: Forecast Any Data Without Training

“A 200M-parameter Google model that forecasts any time series without training — but its own published benchmarks show ARIMA still wins on simpler datasets.”

TimesFM: Forecast Any Data Without Training
1 Views
0 Likes
0 Bookmarks
Source · github.com

“"Internal testing showed TimesFM performs on the same level of an ARIMA model...much bigger and slower" — magimas, Hacker News (https://news.ycombinator.com/item?id=47583045)”

You need to forecast 50 different time series across your company — demand by SKU, energy by facility, support tickets by category. The standard workflow is to pick a model per series type, gather enough labeled history, train and validate each model separately, and retrain whenever the underlying pattern drifts. That is weeks of ML engineering per forecasting surface, and it resets every time you add a new series. LLMs changed this for text by training once and adapting at inference — TimesFM asks whether the same bet works for numbers across unrelated domains.

time-seriesforecastingmachine-learninggoogle-researchpythonpytorchzero-shot

Think of TimesFM as an analyst who has studied millions of different patterns — Wikipedia traffic, Google search trends, retail seasonality — and now reads any new series with pattern-recognition built from that experience. You hand it a sequence of historical values (any length up to 16,000 timesteps). Internally it slices the sequence into 32-point chunks called patches, runs them through a 20-layer causal transformer, and at each step predicts the next 128 timesteps at once — 4 output points for every 1 input point, cutting the number of decoding steps needed for long horizons. A masking trick during training (randomly hiding fractions of the first patch) forced the model to handle any context length from 1 to the training maximum, not just clean multiples of 32. No fine-tuning required at inference.

01
Zero-shot forecasting — drop in any time series and get predictions without training a single parameter, saving days of model selection and fitting for each new dataset
02
4:1 output-to-input patch ratio (128 output vs. 32 input timesteps per step) — each decoding step produces 4x more predictions than it reads, cutting total inference passes for long horizons
03
Up to 16,000-timestep context window at inference — submit years of hourly history without truncation, which matters for series with long seasonal cycles
04
Continuous quantile forecasting (v2.5, September 2025) — get calibrated uncertainty intervals on every prediction without a separate training pass
05
XReg covariate module — attach known-future regressors like holidays or promotions to condition forecasts on external drivers you already know
06
LoRA fine-tuning via HuggingFace PEFT (added April 2026) — adapt the pretrained checkpoint to your domain on a single GPU without full retraining
07
Three inference backends (PyTorch, JAX/Flax, HuggingFace Transformers) plus BigQuery ML and Vertex Model Garden for teams already in the Google Cloud stack
Who it’s for

If you are an ML engineer or data scientist who repeatedly builds forecasting pipelines for new datasets — demand planning, anomaly detection, capacity forecasting — TimesFM eliminates the per-dataset training step for good-enough zero-shot baselines. Also useful for researchers benchmarking against a strong zero-shot foundation model. Not the right fit if your problem is purely multivariate (cross-series correlation), if you need interpretable models for regulated industries, or if you are on Python 3.12 where compatibility issues are actively reported in the 212 open GitHub issues.

Worth exploring

TimesFM is worth running as a zero-shot baseline before committing to supervised training — its Monash benchmark result (scaled MAE 0.68, beating N-BEATS and DeepAR in zero-shot across 18 datasets) is a genuine signal. The 20k+ stars, active maintenance through May 2026, and Google Cloud production integrations give it real staying power. However, the Darts benchmark shows ARIMA is still competitive on simpler univariate series, the 212 open issues include a reported data leakage bug in `forecast_with_covariates()`, and Python 3.12 compatibility is unresolved — run your own validation before treating it as a production forecasting layer.

Developer playbook
Tech stack, code snippet, sentiment, alternatives.
PM playbook
Adoption angles, user fit, positioning.
CEO playbook
Traction signals, ROI, build vs buy.
Deep-dive insight
Full long-form analysis, no fluff.
Easy mode
Core idea, fast — when you need the gist.
Pro mode
Technical nuance, edge cases, tradeoffs.
Read the full digest
Go beyond the preview

Deep-dive insight, Easy and Pro modes, plus action playbooks — the full breakdown is one tap away.

Underrated tools. Unfiltered takes.

Read the full digest in the Snaplyze app for deep-dive insight, Easy and Pro modes, and the playbooks you can actually use.

Install Snaplyze →