Skip to content
Predict.ai

Models & methods

From ARIMA to foundation models: forecasting methods compared

Classical models are interpretable and cheap. Machine learning eats more drivers. Foundation models arrive pretrained and can go zero-shot. None of them win every series — which is why serious stacks keep a tournament, not a favorite.

Updated Aug 13, 2026·10 min read

Always a baseline

Last value. Last week. Last year. A seasonal naive. If you skip this, every other method gets a participation trophy. Baselines are not dummy work. They are the price of admission, and they should run in production as a watchdog — when the champion loses to last week three origins in a row, something is stale.

The methods, side by side

  1. 01Naive / seasonal naiveCheap
  2. 02Smoothing, ARIMALow
  3. 03Gradient boostingMedium
  4. 04Foundation (zero-shot)API / GPU
  5. 05Fine-tuned / customHighest
Complexity is a cost. Climb only when a fair backtest says the extra rung moved error enough to change a purchase order.

Smoothing and ARIMA

Exponential smoothing and ARIMA are still respectable on a single well-behaved series with a person who can read a residual. They are a slow way to lose when you have thousands of SKUs, stockouts, and a promo calendar that looks like a ransom note. Use them as candidates, not as an identity.

“Is ARIMA outdated?” is the wrong question. For one series and a statistician in the chair, no. For a retailer on Tuesday morning, usually yes.

Machine learning

Gradient boosting (and friends) eat lags, calendars, and covariates without requiring the series to sit still. They like panels: many related series, shared features. They will overfit a kitchen sink. They need the same walk-forward honesty as everyone else. Tree models do not get a free pass on leakage.

Foundation models

Pretrained on many series, then used zero-shot or fine-tuned. They help when history is short or when you want a strong first path without a custom loop. They are not magic on a one-off series with a private calendar. They still like drivers you will have at inference. Put them in the tournament. Do not replace the tournament with a blog post about tokens.

How to choose

  • Same data, same folds, same horizon for every candidate.
  • Include a naive. Include one simple statistical model.
  • Promote on the metric ops feels, not the one that flattered the lab.
  • Complexity is a cost. Climb a rung only when error moved enough to change a PO.

FAQ

What is the best forecasting model?
The one that wins a fair backtest on your series, grain, and horizon. That winner changes when the business changes. Treat 'best' as a temporary job title.
Is ARIMA outdated?
For a single well-behaved series with a statistician in the chair, it is still a respectable tool. For thousands of SKUs with promos and stockouts, it is a slow way to lose a tournament.
When do foundation models help?
Short history, many related series, or when you want a strong first path without a custom training loop. They are not magic on a one-off series with a unique promo calendar — they still like good drivers.

Keep going

Try it on your data.

Connect an outcome and its history. Predict.ai finds the drivers, races the models, and keeps the forecast live — in the workspace, over the API, and through agents.