
Time-series forecasting with AI predicts the next values of anything you measure over time: weekly sales, prices, call volume, server load. In 2026 you have a new option next to classic statistics and machine learning. Pretrained foundation models like Google's TimesFM 3.0 can forecast a series they have never seen, with no training on your data. They are strongest when you have many series and short histories, like a catalog full of new products. Classic methods still win when the drivers of demand are well known and you need to explain every number.
Whichever model you pick, accuracy depends far more on your data than on the algorithm. Below we compare the main approaches and look at what changed with foundation models this year. Then we get to the part that decides results: the data you need and how to test a forecast before anyone plans around it.
The short version
Start with a simple baseline, test every model on your own history, and judge it by what forecast errors cost you. A foundation model is often the fastest way to a decent first forecast; it isn't automatically the best one.
Where Forecasting Pays Off
Forecasting earns its keep when a decision has to be made before the outcome is known, and a wrong guess costs money in either direction.
| Use case | What gets forecast | What a better forecast saves |
|---|---|---|
| Inventory and purchasing | Units sold per product and location | Fewer stockouts and less cash tied up in slow stock |
| Pricing | Competitor prices, demand at a given price | Fewer reactive price cuts, better margin |
| Staffing | Calls, tickets, orders per hour | Fewer idle shifts and fewer missed service levels |
| Cash planning | Revenue and payments by week | Less emergency borrowing, better timing of spend |
| Infrastructure | Traffic and server load | Capacity bought ahead of peaks instead of during them |
If none of these decisions is made regularly in your business, a forecasting project will struggle to show a return, no matter how good the model is.
The Main Approaches Compared
| Classic statistics (ETS, ARIMA) | Machine learning with features (gradient boosting) | Foundation models (TimesFM, Chronos-2) | |
|---|---|---|---|
| Needs training on your data | Yes, per series | Yes, one model across series | No, works zero-shot; fine-tuning optional |
| History needed | Several seasons per series | Moderate, pooled across series | Little; works on short histories |
| Uses external drivers (price, promotions, weather) | Limited | Yes, its main strength | Newer models support them |
| Explaining a forecast | Easy | Feature importance, reasonably clear | Hard |
| Setup effort | Low | Medium to high | Low to start, higher to run well in production |
| Best for | A few stable series | Retail and demand with known drivers | Many series, new items, fast first results |
The approaches aren't rivals so much as stages. A classic model is the baseline. Gradient boosting is where many mature demand-planning systems end up. A foundation model is the fastest way to a credible forecast on day one, and in many setups a strong extra input to the other two.
What Changed With Foundation Models In 2026
Google released TimesFM 3.0 on August 31, 2026. It has 330 million parameters, was pretrained on more than a trillion time points, and is the first TimesFM that forecasts several related series together instead of one at a time. Google reports it as the top-ranked pretrained model on three public benchmarks (GIFT-Eval, fev-bench and TIME) for both point and probabilistic forecasts. Earlier versions also gained external driver support and LoRA fine-tuning during the year.
There is a licensing catch business teams should know about. The TimesFM code and weights up to version 2.5 are open under Apache 2.0, while the 3.0 weights are released for non-commercial use only. For commercial production use, Google offers TimesFM 3.0 through BigQuery ML, where its rollout finished in September 2026. If you want to run a model on your own servers, version 2.5 or Amazon's Chronos-2 are the open options to start from.
What this means in practice: for a business with thousands of products and patchy history, a reasonable forecast is now days of work instead of months. It doesn't remove the work that decides accuracy, which is the data and the testing.
Where Classic Methods Still Win
The most-cited real-world test is the M5 forecasting competition on Walmart sales data, where most top solutions were built on gradient-boosted trees fed with prices and promotions alongside the calendar. When you know what moves your demand and have the data for it, a well-built feature model is hard to beat.
Classic approaches also win when someone has to defend the number. A purchasing manager can follow why an exponential smoothing model predicts a December spike. A foundation model's forecast is harder to explain, and that matters when a forecast drives a six-figure order. For a handful of stable series, a simple statistical model is cheaper to run and often just as accurate.
The Data You Need
- Enough history at the right grain. Daily data for daily decisions, ideally covering at least one full seasonal cycle. Foundation models tolerate short histories better, but no model can learn a yearly pattern from three months.
- The drivers that actually move the number. Your own prices and promotions, holidays, weather for some businesses, and competitor prices in retail. Competitor data usually has to be collected from the web; our guide to web scraping for business covers how that works.
- Clean records of what really happened. Sales during a stockout show what you had, not what customers wanted. Unless those periods are flagged, every model will underestimate demand.
- A single, reliable pipeline. Forecasts are only as fresh as the data feeding them. If the numbers arrive by spreadsheet once a month, that is the first thing to fix.
Test Accuracy Before You Trust It
Backtest On Your Own History
Pretend it's six months ago, forecast the next period, compare with what happened, move the window forward and repeat. This rolling backtest is the only accuracy number that means anything for your business. Public benchmark rankings tell you which models are worth trying, not which one wins on your data.
Beat A Naive Baseline First
Every candidate should beat a seasonal naive forecast, which simply repeats last year's or last week's value. It sounds trivial, and it's surprisingly hard to beat on some series. A model that can't clear it isn't worth deploying.
Measure In Money, Not Just Percent
Choose a metric that matches the decision. Weighted error across products (WAPE) is usually more useful than average percentage error, which explodes on slow sellers. Better still, translate errors into cost: an extra unit of stock costs one thing, a lost sale costs another. Ask for forecast ranges too, since the ordering decision depends on the uncertainty as much as on the single number.
What A Forecasting Project Costs
Most of the budget goes into data rather than the model. With us, discovery takes up to five business days and costs $1,000–$1,500; for forecasting, that's where we check whether your history is long and clean enough. Environment preparation, including data pipelines, runs about $2,000. A first working version, an MVP that forecasts your key series and shows the backtest, typically lands at $2,000–$10,000 depending on how many sources need to be connected. After launch, models need monitoring and periodic retraining; our support for AI systems runs $700–$1,500+ a month. The full stage breakdown is in our AI development cost guide.
How Our Team Builds Forecasting Systems
The hard part of forecasting is usually the data, and data is our home ground. Large-scale data collection and processing is one of our strongest areas: we build the pipelines and ETL that keep prices and sales history flowing, including price monitoring for RetailShake and competitive intelligence for LiveTrend. Those systems produce exactly the kind of time series that forecasting runs on.
We build system-first. The model is one component. Around it sit the data pipeline and backtesting, plus the dashboard where planners actually use the numbers. Our backend work runs on Python and PostgreSQL, deployed on AWS, Azure or Google Cloud, so a TimesFM model in BigQuery ML or a self-hosted open model both fit the same architecture.
Frequently Asked Questions
Sometimes. Foundation models do very well on public benchmarks and on series with short histories, but a gradient-boosting model with the right drivers often wins on mature retail data. The only reliable answer comes from backtesting both on your own history.
Not by downloading the weights, which are licensed for non-commercial use. Commercial use goes through Google's BigQuery ML. Versions up to 2.5 are open under Apache 2.0 and can run on your own infrastructure.
Ideally at least one full seasonal cycle at the grain you decide on, such as a year of daily sales for yearly patterns. Foundation models can produce forecasts from much shorter histories, but they can't invent a seasonal pattern the data never showed.
There's no universal number. A forecast is good when it beats a naive baseline on your own data and reduces the cost of wrong decisions. Compare error rates in the context of each product's volume and the cost of being over or under.