Google TimesFM 2.5: Zero-Shot Time Series AI — Complete Guide

On March 31, 2026, Google Research released TimesFM 2.5 — a 200-million-parameter zero-shot time series forecasting foundation model that delivers production-ready predictions without any domain-specific training. Pre-trained on 400 billion real-world time-points spanning retail, finance, healthcare, and industrial telemetry, TimesFM 2.5 introduces a context window 8x larger than its predecessor (16,384 time-points), a probabilistic quantile forecasting head, and restored support for external covariates like holidays and promotional events. The headline benchmark result: TimesFM 2.5 outperforms ARIMA by 15–25% on retail and financial forecasting datasets while matching fully fine-tuned deep learning models — with zero task-specific training data required. For developers and data scientists who have spent years managing training pipelines, data labeling overhead, and retraining cycles, TimesFM 2.5 represents the same paradigm shift that large language models brought to text: a single pre-trained model that works out of the box across wildly different domains.

Try it yourself: Free CAGR Calculator — free, no signup, runs in your browser.

Why Time Series Foundation Models Took So Long

Large language models arrived first because text is the most abundant structured data format in existence — the internet provided a natural pre-training corpus of near-infinite scale. Time series data is different. It lives in silos: retail demand data sits in ERP systems, financial time series in proprietary trading platforms, industrial sensor data in SCADA systems. None of it is scraped and indexed. Building a general-purpose foundation model required Google to curate a specialized dataset of 400 billion real-world time-points from sufficiently diverse domains to achieve genuine zero-shot generalization.

The architectural challenge was equally non-trivial. Standard transformer architectures were designed for discrete token sequences. Time series data is continuous, multivariate, and irregularly sampled across domains. TimesFM’s solution — introduced in the original 2024 paper and refined in subsequent versions — is a patch-based tokenization approach: instead of treating individual time-points as tokens, TimesFM groups contiguous time-points into “patches” and processes each patch as a single token in a decoder-only autoregressive architecture. This design captures local temporal patterns within patches while the transformer’s attention mechanism captures long-range dependencies across the full time series — effectively the same insight that enabled transformers to handle long documents by chunking sentences into paragraphs.

What Changed in TimesFM 2.5

The 2.5 release ships four meaningful improvements over the 2.0 version:

Context length expanded from 2,048 to 16,384 time-points. The 8x context expansion is the most practically significant change. A daily sales time series at 2,048 context points covers roughly 5.6 years of history. At 16,384 points, you capture 44 years — enough to detect multiple business cycles, long-term seasonality, and macro trend shifts that shorter-context models systematically miss. For weekly data, the 16K context covers over 300 years, which is effectively unlimited for any real business forecasting use case.

Parameter count reduced 60% to 200 million. Despite the larger context window, TimesFM 2.5 achieves better performance with fewer parameters through architectural efficiency improvements. At 200M parameters, the model runs comfortably on a single consumer-grade GPU or on CPU for batch inference. It is available as google/timesfm-2.5-200m-pytorch on Hugging Face, installable via pip install timesfm.

Optional 30M quantile head for probabilistic forecasting. The 2.0 release produced point estimates only — a single predicted value per future time-step. The 2.5 quantile head produces a full distribution of outcomes at configurable percentile levels (P10, P20, P30, P50, P70, P80, P90). This is the capability that makes TimesFM genuinely useful for inventory planning, supply chain management, and financial risk modeling, where the tails of the forecast distribution matter as much as the median prediction. The quantile head supports up to 1,000 steps of horizon, covering approximately 2.7 years of daily data in a single inference pass.

Restored XReg (external regressor) support. TimesFM 2.0 temporarily removed covariate support during an architecture revision. Version 2.5 restores it via XReg, which applies a linear ridge regression correction using external covariates on top of the model’s base forecast. This means you can condition TimesFM’s predictions on known future events: holiday schedules, promotional calendars, planned price changes, or macroeconomic indicators. The correction is applied post-hoc, preserving the zero-shot base forecast while incorporating structured domain knowledge where it is available.