TimesFM 2.5: A Practical Read for Forecasting Teams

time-seriesforecastingtimesfmgoogle-researchbigquery

Most teams do not need another forecasting demo. They need a model they can evaluate quickly across many series, with clear uncertainty output and reasonable operational overhead.

That is why TimesFM is worth attention.

Google Research positions TimesFM as a pretrained time-series foundation model for forecasting, and the current open repo points to the ICML 2024 paper, Hugging Face checkpoints, and productized BigQuery integration paths.

What TimesFM is, and what it is not

TimesFM is forecasting-first. It predicts future values from historical sequences.

Think demand, sales, traffic, utilization, pageviews, sensor streams, and similar numeric series.

It is not a causal model. It will not prove intervention effects by itself. It is also not a general anomaly root-cause system, even though TimesFM-backed anomaly workflows exist in BigQuery via AI.DETECT_ANOMALIES.

What changed in TimesFM 2.5

As of April 6, 2026, the public google-research/timesfm README lists TimesFM 2.5 as the latest open version and calls out the core changes versus 2.0.

AreaTimesFM 2.0TimesFM 2.5
Parameter count500M200M
Max context204816k
Quantile supportNot the same continuous setupContinuous quantile forecasting up to horizon 1k via optional quantile head
Frequency indicatorPresentRemoved
Covariates (XReg)Transition periodAdded back on October 29, 2025

That is a meaningful shift. The model got smaller while context support expanded by a lot.

API shape is simple enough for fast trials

The current usage pattern is straightforward in the repo and model card:

import numpy as np
import timesfm

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

model.compile(
    timesfm.ForecastConfig(
        max_context=1024,
        max_horizon=256,
        normalize_inputs=True,
        use_continuous_quantile_head=True,
        fix_quantile_crossing=True,
    )
)

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

The practical win is not just the point forecast. Quantiles give you planning bands, which are usually what business users actually need.

Why the long context window matters

A 16k context window is not just a bigger number for benchmarks.

In real planning systems, older seasonal patterns still matter. Annual cycles, promo periods, shutdown windows, and prior event patterns can all be relevant long after a short context model has already discarded them.

Longer context does not guarantee better forecasts, but it makes certain classes of series feasible without heavy handcrafted feature engineering.

Open model versus official product

There is an important contract boundary.

The open TimesFM repo explicitly says it is not an officially supported Google product. Separately, Google offers TimesFM in BigQuery as an official supported path through AI.FORECAST.

If your team needs managed support and governance controls, that distinction matters.

Where I think this fits best

TimesFM 2.5 looks strongest as a high-leverage baseline and experimentation engine:

  • many related univariate series
  • limited appetite for per-series training pipelines
  • a need for range-based planning with quantiles
  • teams that want to test quickly before committing to custom model stacks

That does not remove the need for benchmarking discipline.

What to benchmark before production

If you are evaluating TimesFM, run a clean baseline ladder first.

BaselineWhy keep it in the bake-off
Seasonal naiveHard to beat on stable seasonal data if your model is overcomplicated
ARIMA/ETS style classical modelsStrong interpretable references for many operational datasets
XGBoost or LightGBM on lag featuresOften competitive with low operational complexity
GluonTS-style deep baselinesGood learned baseline family for comparison across horizons
TimesFM 2.5Foundation-model option with strong zero-shot workflow and quantiles

The wrong move is betting on novelty before beating simple baselines on your own error metrics and service constraints.

Final note

My read is that TimesFM is one of the more serious open forecasting options for teams that want a foundation-model path instead of bespoke model training per dataset.

TimesFM 2.5 makes that story stronger, especially with the smaller footprint, long context support, and quantile output.

But the production bar is still the same as always: beat simple baselines, validate on your horizons, and prove operational reliability before rollout.

Sources

Contact

Questions, feedback, or project ideas. I read every message.