Demand forecasting with foundation models
Predict sales, inventory, and capacity needs using zero-shot time series models. No training step, no per-SKU model management, no GPU infrastructure to operate.
{
"model": "amazon/chronos-bolt-base",
"inputs": [{
"item_id": "SKU-1042",
"target": [312, 287, 341, 298, 376, 355, 402],
"start": "2026-03-01T00:00:00Z"
}],
"parameters": {
"prediction_length": 28,
"freq": "D",
"quantiles": [0.1, 0.5, 0.9]
}
}28-day demand forecast with uncertainty intervals for inventory planning.
Scale
Forecast thousands of SKUs through one endpoint
Speed
No model training — send history, get forecasts immediately
Confidence
Quantile forecasts for safety stock and capacity planning
How foundation models change demand forecasting
Classical demand forecasting requires a trained model per SKU. Foundation models generalize across series with no fitting step.
No per-SKU training
Foundation models forecast any series zero-shot. Add new products or locations without retraining. Cold-start SKUs get forecasts from day one.
Built-in uncertainty
Quantile forecasts give you prediction intervals out of the box. Use them for safety stock calculations, capacity buffers, and risk-aware planning.
Compare models on your data
Test Chronos, TimesFM, and Moirai on your actual demand patterns. The API uses one request shape so you can swap models without rewriting integration code.
Compare modelsGetting started with demand forecasting
- 1
Pick a representative SKU
Start with one product or location that has at least a few weeks of daily history. Send the series to the API and evaluate the forecast against held-out data.
- 2
Evaluate model fit
Try 2–3 models on the same series. Compare accuracy, prediction intervals, and latency. Use the playground for interactive comparison.
- 3
Scale to your catalog
Once you have a model that works, use the batch endpoint to forecast your full SKU catalog. Integrate the results into your inventory or planning system.
Frequently Asked Questions
Forecast your first SKU
Send a demand series to the API and get back point forecasts with prediction intervals. No training required.