Forecasting used to follow a familiar pattern: pick a statistical method such as exponential smoothing or ARIMA, or train a gradient-boosted or deep learning model on your own history, then tune it series by series. Since 2024 a new option has arrived. Time series foundation models, pretrained on very large and diverse collections of series, can produce forecasts “zero-shot” for data they have never seen. Amazon’s Chronos family, Google’s TimesFM, Salesforce’s Moirai and others are now open source, and cloud platforms are exposing them through simple interfaces, including a SQL function in BigQuery.
The pitch is attractive: no per-series training, decent accuracy out of the box and probabilistic forecasts for thousands of series in minutes. But the research community is also pushing back, arguing that many reported wins rest on weak baselines and contaminated benchmarks. For data science and planning teams the practical question is simple: should we use these models, for which series, and how do we know they are actually better? This guide offers a decision framework and a testing playbook.
What Time Series Foundation Models Are
A time series foundation model is a neural network, usually a transformer, pretrained on a large corpus of real and synthetic series across many domains and frequencies. At inference time you pass it a window of recent history (the context) and it returns a forecast distribution for the requested horizon. No gradient updates are needed, although most models can be fine-tuned.
The field moved quickly in 2025. Amazon released Chronos-2 in October 2025, describing it as a move from univariate to “universal” forecasting. According to Amazon Science, Chronos-2 handles univariate, multivariate and covariate-informed tasks zero-shot through in-context learning. It supports past-only covariates, known future covariates such as planned promotions or weather forecasts, and categorical covariates such as holidays. Its architecture alternates attention across time within a series with “group attention” across related series, which lets it share information within groups and helps with cold-start items that have little history. Amazon reports that Chronos-2 ranked first among pretrained models on the GIFT-Eval benchmark at release and won more than 90 percent of head-to-head comparisons against its predecessor, Chronos-Bolt.
Covariate support is the important shift. Early foundation models were purely univariate, which limited them in business settings where price, promotions, calendar effects and capacity drive demand. Newer models narrow that gap, though support varies by model and by the platform that hosts it.
The Case for Caution: What the Critics Say
Benchmark leaderboards are the main evidence for foundation model accuracy, and that evidence is contested. Three issues come up repeatedly.
- Weak baselines. Forecasting researcher Rob Hyndman argues that comparisons often pit foundation models against under-tuned or non-ensembled statistical methods, and calls this the single biggest problem in the literature. Well-implemented ETS, ARIMA, Theta and simple combinations are strong and cheap.
- Leakage. Public datasets are reused so widely that test series often appear in pretraining corpora. Hyndman cites research finding that only a small fraction of the datasets used to evaluate foundation models had never appeared in any model’s training data. Indirect leakage also matters: series that share shocks such as the pandemic are correlated even if not identical.
- Unrepresentative evaluation. A NeurIPS 2025 paper on cross-frequency transfer learning reimplemented common neural forecasters, pretrained them only on proprietary and synthetic data, and evaluated on 15 forecasting competition datasets. The authors found that statistical models and their ensembles outperformed existing foundation models by more than 8.2 percent on sCRPS and more than 20 percent on MASE across datasets, when the zero-shot comparison excluded TimesFM, and that uniform averaging, small test sets and inflated horizons made foundation models look better than they are.
Benchmark builders have responded. Salesforce’s GIFT-Eval, for instance, covers 28 datasets and around 144,000 series across frequencies from seconds to years, and ships a separate pretraining corpus designed to be free of test leakage. That improves the situation but does not remove the core advice: your own data is the only benchmark that matters.
Foundation models change how forecasts are produced, not the rules for proving they are good. Test them against properly tuned statistical baselines, on your own series, with rolling-origin backtests, before they touch a plan.
Where Foundation Models Tend to Help, and Where They Do Not
Combining the published evidence with practical experience suggests a pattern.
Likely to help
- Large portfolios of short or new series. Thousands of SKUs, stores or sensors where per-series tuning is impractical and many items have little history. Cross-learning and pretrained priors help with cold starts.
- Fast prototyping. Getting a credible probabilistic baseline in hours instead of weeks, especially for teams without forecasting specialists.
- Heterogeneous frequencies and domains. Organizations forecasting web traffic, energy load and sales with one tool.
- Related series with shared dynamics. Multivariate models can exploit co-movement, for example infrastructure metrics or regional demand.
Less likely to help
- Long, stable, well-understood series where tuned ETS or ARIMA already performs well and explanations matter.
- Highly intermittent demand, unless the specific model has been validated on intermittent data; specialist methods may still win.
- Forecasts dominated by known drivers such as contracted volumes, regulatory changes or planned capacity, where a causal or driver-based model is more appropriate.
- Strict interpretability requirements, for example forecasts that feed regulated provisioning or where planners must explain every adjustment.
A Decision Framework: Five Questions
- What decision does the forecast support? Inventory replenishment, staffing and cash planning have different cost-of-error profiles. Choose evaluation metrics that reflect them, such as quantile loss at the service-level quantile for inventory.
- How many series, and how much history? Hundreds of long series favor classical methods; tens of thousands of short series favor global models, including foundation models and global gradient-boosted models.
- Which drivers matter, and are they known in advance? If promotions, prices or weather forecasts drive outcomes, require a model that accepts future covariates, and test whether it actually uses them well.
- What are the operating constraints? Latency, compute cost, data residency and the ability to run inside your platform. A model hosted only through an external API may not satisfy residency rules.
- Who must trust the forecast? If planners override forecasts frequently, interpretability and stable behavior may outweigh small accuracy gains.
Playbook: Running a Fair Forecasting Bake-Off
The most valuable thing a team can do is run a disciplined comparison on its own data. The following steps take two to four weeks for a typical portfolio.
Step 1: Assemble representative series
Sample series across volume bands, product lifecycles, intermittency levels and business units. Include new items and series affected by structural breaks. Hold back a separate set for final confirmation.
Step 2: Build strong baselines
Include seasonal naive, automatically tuned ETS and ARIMA, Theta and a simple combination of these. If you have a production model today, include it. If drivers matter, add a global gradient-boosted model with the same features you would give a foundation model. Weak baselines are the main reason published comparisons mislead, so give classical methods the same care.
Step 3: Use rolling-origin backtesting
Evaluate with multiple forecast origins over the past one to two years, at the horizon and frequency the business actually uses. A single train-test split hides instability. Make sure that every covariate is used only as it would have been known at each origin, for example the promotion calendar as it stood then, not as it was finally executed.
Step 4: Measure what matters
- Point accuracy with a scale-free metric such as MASE, plus bias.
- Probabilistic accuracy with quantile loss or CRPS, and coverage of prediction intervals.
- Results broken down by segment, not just a portfolio average.
- Forecast stability between runs, because large swings erode planner trust.
- Cost and latency per full portfolio run.
Step 5: Decide by segment
The outcome is rarely “foundation model everywhere.” More often a foundation model wins for short and new series, classical methods win for long stable series, and an ensemble of both wins overall. Route series to methods by segment, and revisit the routing quarterly.
Step 6: Productionize with monitoring
Track accuracy by segment over time, detect drift and keep a fallback method ready. Version the model and its configuration, including context length and covariates, so forecasts are reproducible for audit and post-mortems.
Worked Example: Weekly Demand for a Mid-Sized Distributor
Imagine a distributor forecasting weekly demand for 12,000 SKU-warehouse combinations, 13 weeks ahead, to drive replenishment. About a fifth of the SKUs are less than a year old, and promotions affect roughly a third of the catalog.
The team builds baselines (seasonal naive, auto-ETS, auto-ARIMA and their combination), a global LightGBM model with price, promotion and holiday features, and a covariate-aware foundation model run zero-shot. It backtests from twelve weekly origins across the previous year.
A plausible outcome looks like this. On long-history, non-promoted SKUs, the statistical combination is about as accurate as the foundation model and far cheaper, so it remains the default. On new SKUs, the foundation model clearly beats the statistical methods because it borrows patterns learned from similar series. On promoted SKUs, the gradient-boosted model and the covariate-aware foundation model are close, and an average of the two is the most accurate and the most stable. The team deploys a routing rule by segment, adds an ensemble for promoted items, and monitors quantile loss at the 90th percentile, which maps to its service-level target. The decision is defensible because it rests on the company’s own backtests rather than on leaderboard positions.
Implementation Options in 2026
Teams can access forecasting foundation models in several ways, each with trade-offs.
- Open weights on your own compute. Chronos-2 and other models are published with weights and code. This gives control over data residency and versioning, at the cost of managing GPU or CPU inference.
- Managed endpoints. Amazon provides a notebook for deploying Chronos-2 on SageMaker; other clouds offer similar paths. Good for teams already standardized on one cloud.
- In-warehouse functions. BigQuery exposes TimesFM through the AI.FORECAST SQL function, which lets analysts produce forecasts without leaving SQL. Convenient for exploration; check model versions, covariate support and limits before using it for production planning.
- Forecasting libraries. Open-source frameworks increasingly wrap foundation models alongside statistical and machine learning methods, which simplifies the bake-off described above.
Explaining Foundation Model Forecasts to Planners
Accuracy is necessary but not sufficient. Forecasts feed purchasing, staffing and financial plans, and the people who own those plans need to understand why a forecast changed. Foundation models are less transparent than exponential smoothing, where level, trend and seasonality are explicit, so teams need deliberate practices to keep planners engaged.
- Show the comparison, not just the forecast. Display the foundation model forecast next to a simple statistical baseline. When they diverge sharply, that is a prompt for review rather than blind acceptance.
- Report covariate effects through scenarios. Even when a model does not expose coefficients, you can run it with and without a planned promotion or price change and show the difference. This makes covariate use tangible.
- Explain changes between runs. Log what changed since the previous forecast: new actuals, updated covariates, a different model version. Most planner distrust comes from unexplained movement.
- Use intervals in planning conversations. Probabilistic output is one of the main advantages of modern models. Translating quantiles into service levels or downside cases helps planners use it.
These practices apply to any machine learning forecast, but they matter more for pretrained models whose internal reasoning cannot be inspected. A forecast that is slightly less accurate but consistently understood will often deliver more business value than a marginally better one that planners override.
Pitfalls to Avoid
- Trusting leaderboards over backtests. Aggregate scores hide large variation across the series you care about.
- Leaking future information through covariates. Using final actual promotions rather than planned ones inflates accuracy.
- Ignoring hierarchy and reconciliation. Forecasts at SKU, category and total levels must add up for planning. Apply reconciliation regardless of the base model.
- Neglecting cost. A small accuracy gain may not justify GPU inference for a weekly batch job when a statistical method runs on a laptop.
- No fallback. If a hosted model changes version or becomes unavailable, planning cannot stop.
- Forgetting the planner. Forecasts that swing unpredictably between runs get overridden, which erases any accuracy gain.
Outlook
Foundation models for time series are improving quickly, particularly in covariate handling and multivariate forecasting, and they are being embedded directly into data warehouses and planning tools. At the same time, the research community is raising evaluation standards, with leak-free pretraining corpora, segment-level reporting and stronger baselines. The likely end state is not a single winner but a toolbox: foundation models as strong default priors for large, sparse portfolios, classical methods for stable series, driver-based and causal models where decisions depend on known levers, and ensembles across all of them. Teams that build a repeatable backtesting harness now will be able to adopt each new model on evidence rather than hype.
Conclusion
Time series foundation models are a genuine advance, especially for organizations forecasting many short or new series with limited data science capacity. They are not an automatic upgrade over well-tuned statistical methods, and published benchmarks overstate their advantage in some settings. Decide by segment, test on your own data with rolling-origin backtests and strong baselines, measure probabilistic accuracy and stability, and keep a fallback. Do that, and foundation models become one more reliable instrument in a forecasting system rather than a gamble.
How Hanoi Technology Can Help
Hanoi Technology runs forecasting bake-offs and builds production forecasting pipelines that combine foundation models, statistical methods and driver-based planning.
Further Reading
- Introducing Chronos-2: From univariate to universal forecasting – Amazon Science
- Evaluating time series foundation model claims – Rob J Hyndman
- A More Realistic Evaluation of Cross-Frequency Transfer Learning and Foundation Forecasting Models – arXiv
- GIFT-Eval: A Benchmark for General Time Series Forecasting – Salesforce