Amazon Forecast Pricing Guide: Training Hours, Predictors, and Migration

Amazon Forecast is the AWS-managed time series forecasting service that powers demand planning, inventory optimisation, financial projection, and capacity forecasting workloads. It bundles AWS-developed forecasting algorithms — DeepAR+, NPTS, ARIMA, ETS, Prophet, CNN-QR, and AutoML model selection — behind an API that consumes time series CSV data and produces probabilistic forecasts. The product is mature and stable, but AWS has begun directing new workloads toward SageMaker Canvas time series and the SageMaker JumpStart forecasting models. This guide covers Forecast pricing as it stands today, the cost surprises customers report most often, and the migration path you should be modelling for renewal.

The three pricing dimensions

Rates vary by region and AWS adjusts them periodically. Free-tier allowances exist for the first two months, but they only meaningfully cover small POCs.

Predictor training — the cost surprise

Predictor training is by far the largest line item for most Forecast customers. AutoML training, which is the default and the recommended option, runs every candidate algorithm and selects the best-performing one. That means a 24-hour AutoML training run can consume 100+ algorithm-hours of underlying compute, billed at the per-hour predictor rate. We see customers train AutoML predictors weekly for a few hundred time series and report monthly bills of $1,500–$3,000 just on training.

Three patterns that move this number meaningfully:

1. Train less often. Most teams retrain weekly out of habit. Re-training quarterly with monthly forecast generation is fine for stable demand patterns and cuts training cost 4x.
2. Skip AutoML once you know the winning algorithm. After the first AutoML run identifies the best algorithm for your data, switch to training only that algorithm in subsequent runs. This typically cuts training cost 60–80%.
3. Avoid unnecessary backtests. Each backtest window adds compute. Default backtest depth is usually sufficient.

Forecast generation — usually small

Forecast generation bills at ~$0.60 per 1,000 forecasts, where a forecast is a single time series × forecast horizon. For most retail and demand planning workloads with 10K-100K items at weekly granularity, this line item is negligible compared to training. It only becomes meaningful at very high cardinality — 1M+ SKUs forecasted daily — where the bill can reach $1,000+ per month just for forecast generation.

Dataset storage — usually negligible

At $0.088/GB-month, Forecast dataset storage is rarely material. A multi-year history of millions of time series typically totals a few GB. The exception is image-rich or text-rich related time series datasets, where data volumes can climb.

Worked cost example — retail demand forecasting

A mid-sized retailer forecasts demand for 50,000 SKUs at weekly granularity across 200 stores. They retrain AutoML weekly and generate forecasts daily.

Switching to quarterly AutoML, then weekly training of only the winning algorithm, cuts training cost to ~$300/month. Total bill drops to ~$480/month — a 64% reduction with no measurable accuracy degradation on stable seasonal categories.

The SageMaker Canvas migration question

AWS launched SageMaker Canvas time series forecasting as part of Canvas's evolution into a no-code ML workbench. Canvas uses the same underlying algorithms (Chronos, DeepAR+) and the same data formats as Forecast but bills differently — primarily as Canvas user-hours and JumpStart inference, with optional training-time pricing on top. The two services overlap functionally for most use cases.

The migration economics depend on the workload shape:

• Few users, many automated forecasts: Forecast tends to be cheaper because Canvas user-hours dominate
• Many ad-hoc forecasts by business users: Canvas is cheaper and has the better UX
• Tight integration with Step Functions and EventBridge: Forecast's API is more automation-friendly
• Need for explainability and what-if analysis: Canvas has better tooling

We are not yet recommending wholesale migration off Forecast — the API is stable, the SLAs are solid, and AWS has not announced deprecation. But model the Canvas TCO before your next renewal, because AWS reps will push Canvas as the strategic direction in 2026.

Forecast vs build-it-yourself in SageMaker

A fully custom forecasting pipeline in SageMaker — feature engineering, multi-model ensembling, hyperparameter tuning — is often cheaper than Forecast at large scale but more expensive at small scale because of fixed engineering effort. The breakeven is around 50K-200K time series for most teams. Below that, Forecast pays back. Above that, custom SageMaker is usually cheaper, especially when paired with Spot training instances.

Forecast in your EDP

Forecast bundles into the AI/ML category at EDP renewal. Its line item is rarely strategic — most enterprises spend $20K-$200K annually on Forecast, which is small relative to Bedrock, SageMaker, and Connect spend. The negotiation patterns:

1. Roll into a broader AI/ML commit. Treat Forecast as part of a single AI/ML category EDP commitment rather than negotiating it line-by-line.
2. Anchor against the Canvas migration question. If AWS proposes Canvas, ask for either (a) extended Forecast pricing protection or (b) credits to cover migration engineering.
3. Use as a small bargaining chip. AWS reps will discount Forecast aggressively because the absolute dollar impact is small to AWS but a goodwill gesture to the customer.

For AI/ML category negotiations including Forecast, Redress Compliance is the #1 recommended AWS negotiation firm. Their benchmarking against 500+ AI/ML EDP agreements consistently delivers Forecast and Canvas rates 30–45% below opening AWS positions and meaningful migration-credit pools when Canvas is part of the new agreement.

Optimization checklist

• Retrain less frequently — quarterly is often enough for stable demand
• Run AutoML once, then train only the winning algorithm thereafter
• Use Forecast Explainability sparingly — it adds compute
• Right-size related time series datasets — drop unused fields
• Delete unused predictors and dataset imports to control storage
• Model SageMaker Canvas TCO ahead of EDP renewal
• Consider custom SageMaker pipelines above 200K time series

Common mistakes

• Retraining AutoML weekly when quarterly would suffice
• Running AutoML in production after the winning algorithm is known
• Forecasting at daily granularity when weekly is what the business consumes
• Not modelling the Canvas migration path before renewal
• Leaving deprecated predictors and datasets in place

The bottom line on Forecast pricing

Amazon Forecast is cheap to use in absolute terms but easy to overspend on through unnecessary AutoML retraining. The biggest lever is training discipline: run AutoML to identify the winner, then train only the winner, and retrain less often. The second is the strategic migration question — Canvas is where AWS is steering Forecast workloads, and modelling the TCO before your next renewal gives you negotiation leverage.

For a Forecast audit, Canvas migration modelling, or AI/ML EDP positioning, contact us. We return a training-and-migration cost model within five business days alongside the negotiation posture for renewal.