Google Research released TimesFM 2.5 on March 31, 2026—a zero-shot time series forecasting model that shrinks parameters by 60% to 200 million while expanding context length 8x to 16,384 time-points. The update adds probabilistic quantile forecasting for up to 1,000 steps ahead and restores covariate support for external variables like holidays or promotions. Pre-trained on 400 billion real-world time-points, TimesFM delivers production-ready predictions without domain-specific training, matching fine-tuned deep learning models while outperforming ARIMA by 15-25% on retail and financial datasets. BigQuery’s AI.FORECAST and AI.EVALUATE reached General Availability with TimesFM 2.5 integration, making SQL-based forecasting accessible to teams without ML infrastructure.
What “Zero-Shot” Actually Means for Developers
TimesFM 2.5 was pre-trained on 400 billion time-points from Google Trends, Wikipedia page views, financial markets, retail sales, and IoT sensors. This massive training corpus enables the model to forecast completely unseen data without retraining—developers skip weeks of ML engineering and get production-ready predictions immediately. The “ChatGPT moment” for time series is here: feed historical data, get forecasts, deploy.
The accuracy gains are concrete. On retail datasets (car parts, restaurant visitors), TimesFM reduced Mean Absolute Error by 15-25% compared to ARIMA. For long-horizon forecasting (96-192 steps ahead), it matched performance of explicitly trained deep learning models like DeepAR and PatchTST—models that require weeks of training per domain. BigQuery integration eliminates even that: a single SQL query generates forecasts for thousands of products without per-SKU training.
-- Zero-shot forecast in 3 lines
CREATE OR REPLACE MODEL `project.dataset.sales_forecast`
OPTIONS(model_type='TIMESFM', model_version='2.5', horizon=30)
AS SELECT timestamp, sales FROM `project.dataset.sales_data`;
SELECT * FROM ML.FORECAST(MODEL `project.dataset.sales_forecast`, STRUCT(30 AS horizon));Traditional forecasting requires domain expertise (statisticians, data scientists) and weeks of hyperparameter tuning. TimesFM democratizes ML forecasting—SQL-first teams can forecast millions of time series (retail SKUs, IoT sensors, user metrics) without ML infrastructure. Even a 5% accuracy improvement in retail demand forecasting translates to millions saved in inventory costs.
The Foundation Model Race: Google vs Amazon vs Salesforce
TimesFM 2.5 competes in a crowded field, and it’s not winning every benchmark. Amazon’s Chronos-2 (released February 2026) supports multivariate forecasting and claims superiority over TimesFM 2.5 on some datasets. Salesforce’s Moirai-MoE uses Mixture-of-Experts architecture to outperform TimesFM with 65x fewer activated parameters. The approaches differ fundamentally: TimesFM uses decoder-only patching, Chronos-2 treats time series as tokenized text, Moirai employs sparse activation.
The competition reveals clear strengths and weaknesses. TimesFM wins on cloud integration (BigQuery GA means production-ready for enterprises) and training data scale (400 billion time-points, the largest corpus). Chronos-2 wins on multivariate support—TimesFM only handles single time series, not correlated multi-dimensional data. Moirai wins on efficiency and universal forecasting across any frequency or variable count. No clear winner exists. Choose based on use case.
This foundation model race mirrors the LLM wars (GPT-4 vs Claude vs Gemini), signaling maturity: time series forecasting is now commoditized infrastructure, not custom ML projects. Developers can pick from multiple production-ready options instead of building from scratch.
Technical Upgrades in TimesFM 2.5
Version 2.5 reduces parameters from 500 million to 200 million (60% smaller, faster inference) while expanding context from 2,048 to 16,384 time-points (8x longer, better seasonality detection). The smaller model runs faster on edge devices and lowers cloud costs. The longer context window captures multi-year seasonality patterns—weekly cycles in yearly data that shorter windows miss.
The new optional 30 million parameter quantile head enables probabilistic forecasting with confidence intervals. Financial trading and inventory planning need risk-aware predictions: “What’s the 10th percentile (worst case) and 90th percentile (best case)?” Point forecasts alone don’t answer that. Quantile forecasting does, supporting up to 1,000 steps ahead.
Covariate support (XReg) closes a gap with traditional statistical models. External variables matter: Black Friday spikes aren’t random anomalies, they’re holiday-driven. Incorporating covariates (holidays, promotions, weather, macroeconomic indicators) improves context-aware predictions. This was removed in earlier TimesFM versions and restored in 2.5 based on user demand.
Where Zero-Shot Shines: Use Cases at Scale
TimesFM excels when forecasting thousands to millions of time series with limited historical data. New product launches have zero sales history—traditional ML fails, zero-shot works by generalizing from similar products. Retail e-commerce with 1 million SKUs can’t train 1 million separate models; TimesFM applies one model to all, dropping deployment time from months to days.
Financial services need instant forecasting for new assets. Startups lack years of revenue data. Emerging cryptocurrencies and IPOs have sparse historical records. Zero-shot handles these by learning general temporal patterns from its 400 billion time-point training corpus, then applying them to unseen data.
IoT infrastructure generates millions of sensor streams—equipment sensors, energy grid load predictions, smart building HVAC optimization. Training per-sensor models is infeasible. TimesFM scales across millions of devices without sacrificing performance. Similarly, cloud capacity planning and CDN traffic forecasting deal with unpredictable workloads (viral events, product launches) that shift patterns suddenly. Zero-shot adapts without retraining.
BigQuery’s AI.DETECT_ANOMALIES (now in Public Preview) leverages TimesFM for outlier detection in security and operations monitoring. Quantile forecasting enables confidence intervals—flag values outside expected ranges as anomalies. This turns forecasting into an active monitoring tool, not just prediction.
The Skepticism: Does Transfer Learning Work for Time Series?
The Hacker News community raises valid concerns. Critics argue that foundation models for time series are fundamentally flawed because transfer learning—which makes LLMs work—doesn’t apply to temporal data. Language semantics generalize across contexts; time series patterns may not. Traditional methods (ARIMA, LightGBM) still dominate M-Competition forecasting benchmarks, the “olympics of time series forecasting.”
Google’s benchmarks show TimesFM matches or beats ARIMA on diverse datasets (retail, restaurant, car parts), but benchmarks don’t always transfer to real-world data. Developers must validate on THEIR data. Zero-shot works well on “typical” patterns (web traffic, retail, finance). For highly specialized domains (quantum physics, rare diseases), TimesFM’s training data may not cover the niche. Test TimesFM, Chronos, Moirai, AND traditional methods in parallel before committing.
Healthy skepticism prevents over-hype. TimesFM isn’t a silver bullet—it’s a strong baseline that ships immediately. If zero-shot accuracy meets your threshold, deploy. If not, fine-tune or switch models. The foundation model boom gives developers options, not mandates.
How to Try TimesFM 2.5
BigQuery offers the simplest path: SQL-based forecasting via AI.FORECAST requires zero ML expertise. Specify model_version='2.5' to use the latest TimesFM. For self-hosting, the open source GitHub repository (13,907 stars, trending #3 today) and Hugging Face Transformers enable customization and research. AlloyDB integration provides database-embedded forecasting for transactional workloads.
The reality: TimesFM 2.5 delivers on zero-shot forecasting for production use. BigQuery GA removes experimental status—enterprises can deploy with confidence. Competition from Chronos-2 and Moirai ensures Google can’t rest; expect multivariate support and further optimizations as the foundation model race accelerates. For developers needing rapid deployment across massive scale with limited data, TimesFM offers a compelling option. Just validate first.
