Demand Forecasting with Amazon Forecast - End of New Sign-Ups, Migration to SageMaker Canvas, and How Time Series Forecasting Works
Amazon Forecast stopped accepting new users on July 29, 2024 (existing users can continue to use it). This article explains migration to time series forecasting in SageMaker Canvas, the official successor, as well as Forecast's model building / probabilistic forecasts / billing structure.
End of New Sign-Ups and Where to Migrate
Amazon Forecast stopped accepting new customers as of July 29, 2024. AWS officially states that existing users can continue to use the service as before and that it will continue to improve security / availability / performance, but that it does not plan to add new features. If you are starting time series forecasting from now on, the official successor is the time series forecasting feature of Amazon SageMaker Canvas, and the official AWS blog post "Transition your Amazon Forecast usage to Amazon SageMaker Canvas" presents the approach and steps for migration. The forecast intervals in SageMaker Canvas are 1 minute / 5 minutes / 15 minutes / 30 minutes / 1 hour / 1 day / 1 week / 1 month / 1 year. Its input concept of a timestamp column / target column / item ID column is also almost the same as Forecast, so you can bring your data design over as is. In addition, it supports comparing algorithms on the model leaderboard and automation through the AutoML API. This article is kept as operational material for existing Forecast environments and as a historical record of the design Forecast adopted. Read the mechanisms and figures covered below as a reference for operational decisions in existing environments and for mapping when migrating to SageMaker Canvas.
Challenges of Demand Forecasting and the Position of Forecast
Demand forecasting is an important challenge in many industries, such as retail, logistics, manufacturing, and energy. Traditional statistical methods (moving averages, exponential smoothing, ARIMA) are relatively easy to implement, but they have limits in capturing complex patterns such as seasonality / trends / the influence of external factors. Deep learning-based methods (DeepAR, Transformer) are highly accurate but require ML expertise and model tuning. Forecast is a service that became available in August 2019. Its defining trade-off was that you only had to import time series data, and AutoML would compare multiple algorithms and choose the most accurate model through backtesting. A design that returns P10, P50, and P90 prediction intervals through probabilistic forecasts and supports risk-aware decision-making by quantifying uncertainty was also at the core of managed forecasting services of this generation. The idea of treating a forecast not as "a single number" but as "a distribution" has been carried over to its successor, SageMaker Canvas, so understanding Forecast's design is also directly useful for design after migration.
Data Preparation and Building Forecast Models
Using Forecast starts with importing data into a dataset group. The target time series (required) is the data to be forecast and consists of three columns: timestamp, item ID, and value (units sold, demand quantity). The related time series (optional) contains external factors such as weather data, promotion information, and price changes. Item metadata (optional) contains item attributes such as category, brand, and color, and is used to share patterns among similar items. You place the data in S3 in CSV format and specify the attribute type and data type of each column in a schema definition (JSON). The granularity of timestamps is specified as a combination of an integer and a unit. DataFrequency on the dataset side and ForecastFrequency on the forecast side use a common format. Minutes take the range 1 to 59, hours 1 to 23, days 1 to 6, weeks 1 to 4, months 1 to 11, and years 1. Specifications such as 15min (every 15 minutes) and 2W (every other week) are also valid (values listed in the API reference as of August 2026). However, you cannot specify a value that overlaps with the next larger unit, such as 60 minutes = 1 hour. You set it according to your business requirements and match it to the unit of the forecast horizon. A dataset group can hold one dataset of each type, up to three datasets, and the target time series can have up to 13 columns, made up of 3 required columns + 10 additional forecast dimension columns. Because of these constraints, you settle the dimension design first. AutoML builds models with six algorithms, DeepAR+, Prophet, ETS, NPTS, ARIMA, and CNN-QR, and automatically selects the most accurate model through backtesting.
Probabilistic Forecasts and What-If Analysis
Forecast provides not only point forecasts (a single forecast value) but also probabilistic forecasts (which include the uncertainty of the forecast). By default, it generates forecasts at three quantiles: P10 (a 10% probability that demand falls below this value), P50 (the median), and P90 (a 90% probability that demand falls below this value). The quantiles are not fixed to the default values. According to the official API reference, you can specify up to five quantiles from 0.01 to 0.99 in steps of 0.01 with ForecastTypes, and you can also specify mean, which returns the average value (values listed as of August 2026). In inventory management, you can use P50 to decide normal order quantities and P90 to set safety stock. It supports decisions that match business requirements, such as ordering on a P90 basis for products where stockout risk cannot be tolerated and on a P50 basis for products where you want to minimize inventory costs. However, the number of quantiles you specify is proportional to the number of forecast data points generated and directly affects the billing described later, so the practical key point is to choose only the quantiles you need. With What-If analysis, you can generate scenario forecasts such as "demand if a promotion is run next month" and "demand if prices are cut by 10%." You can evaluate multiple scenarios with changed related time series values in parallel and quantitatively assess the effect of measures in advance.
Improving Forecast Accuracy
The forecast accuracy of Forecast depends heavily on the quality of the input data and the richness of related data. When you add related time series data (weather, holidays, promotions), the model learns demand fluctuations caused by external factors and accuracy improves. Forecast has built-in holiday information (Holidays featurization), and just by specifying a country code, it automatically takes the influence of major holidays into account. Item metadata (category, brand, price range) is effective for mitigating the cold start problem (forecasting new products), and it estimates demand patterns for new products from the track record of similar items. The backtest results of a Predictor provide the accuracy metrics WAPE (Weighted Absolute Percentage Error), RMSE (Root Mean Squared Error), and MAPE (Mean Absolute Percentage Error), so you can judge model quality quantitatively. You can choose the metric used for optimization from WAPE / RMSE / MASE / MAPE / AverageWeightedQuantileLoss with OptimizationMetric in the official API reference, so you align the metric with whether you weigh stockout costs or inventory costs more heavily. You can set up to five backtest windows, and by validating over separate periods, you can obtain an evaluation that is not swayed by chance in a single period. If accuracy is insufficient, you improve it by adding related time series, preprocessing the data (filling in missing values, removing outliers), and changing the forecast frequency.
Use Cases and Industry Applications
Forecast has been used in a wide range of industries, such as retail, manufacturing, logistics, and energy. In retail, demand forecasting at the SKU level optimizes inventory placement and reduces both overstock and stockouts. In manufacturing, forecasting order quantities of raw materials is used to draw up purchasing plans that take procurement lead times into account, contributing to cost optimization across the entire supply chain. In logistics, forecasting shift demand for workers optimizes staffing and prevents staff shortages in busy seasons and surpluses in slow seasons. In the energy industry, forecasting electricity consumption is used to plan power generation and supports grid management that responds to output fluctuations of renewable energy. In web services, it can also be applied to server capacity planning and is used to optimize Auto Scaling settings based on traffic forecasts. All of these designs come down to the common form of "a time series per item ID + related time series of external factors." So even when you implement the same use case anew, you can design it by directly replacing it with time series forecasting in SageMaker Canvas.
Notes on Pricing and Limits
Billing consists of five axes (as of August 2026 / N. Virginia and Tokyo Regions; measurements with the AWS Price List API show the same prices in both Regions). The first is the number of forecasts generated, at $0.60 per 1,000. The second is the number of forecast data points generated, which is tiered per month. It costs $2.00 per 1,000 for less than 100,000, $0.80 for more than 100,000 and less than 1,000,000, $0.20 for more than 1,000,000 and less than 50,000,000, and $0.02 for more than 50,000,000. The third is training time at $0.24 per hour, and the fourth is the amount of data ingested at $0.088 per GB. This $0.088 is a one-time fee charged at ingestion, not a monthly storage fee incurred for as long as the data is stored. The fifth is the number of Explainability generations, tiered from $2.00 per 1,000 for less than 50,000 per month to $0.15 for more than 10,000,000. Check the official Amazon Forecast pricing page for the latest list, including the boundary values of the tiers. What drives the bill is not the number of items itself but the number of data points, determined by number of items × forecast horizon × number of specified quantiles, so you always build estimates from this multiplication. Because AutoML trains multiple models, training costs are higher than when you specify an algorithm manually. If forecast accuracy is sufficient, you reduce costs by changing the training frequency from weekly to monthly. On the limits side, the official quota list (values listed as of August 2026) gives 500 dataset groups and 500 Predictors (both can be raised by request) and 500 AutoPredictors (cannot be raised). Dataset imports are limited to a cumulative size of 30 GB for all files in the S3 bucket (can be raised by request) and 10,000 files (cannot be raised). What determines the forecast horizon is not the granularity but the series length. For a new AutoPredictor, it is the smaller of 500 time steps and 1/4 of the length of the target time series. For retraining an existing AutoPredictor and for CNN-QR / DeepAR+ / AutoML in legacy predictors, it is the smaller of 500 and 1/3, and for ETS / NPTS / Prophet / ARIMA, it is the smaller of 500 and the series length minus 1. Training can take several hours for large datasets (millions of rows), so you leave slack in your operational schedule.
Summary
As a managed service that automatically builds ML models from time series data, Forecast adopted a design that fills the limits of traditional statistical methods with AutoML and chooses the optimal model from six algorithms through backtesting. Forecast improved accuracy with related time series data and built-in holiday information, returned prediction intervals of P10/P50/P90 by default through probabilistic forecasts, and supported scenario comparison with What-If analysis. This structure has been widely applied, from inventory optimization in retail to procurement planning in manufacturing and grid management in energy. New sign-ups ended on July 29, 2024, so if you are starting demand forecasting from now on, the official path is to begin with the time series forecasting feature of SageMaker Canvas. If you operate an existing Forecast environment, the practical key points are understanding the billing structure based on the number of data points and considering a migration plan on the premise that no new features are planned. If you have accumulated sales records for one year or more, it is a good idea to start by validating forecast accuracy in SageMaker Canvas with the same data.
References (Official AWS Resources)
The primary sources for this page are the official AWS website and documentation. Check the official pages below for the latest specifications and pricing.
- Amazon Forecast official page
- Amazon S3 official page
- Amazon Quick official page
- AWS Documentation (official)
If this page and the official documentation disagree, treat the official documentation as authoritative.