Demand Forecasting - Predicting the Future from Time Series Data with Amazon Forecast
Input historical time series data and related variables to automatically build ML-based demand forecasting models. This guide covers forecast accuracy evaluation metrics and patterns for leveraging forecast results through S3 and QuickSight integration.
Demand Forecasting Challenges and Forecast's Role
Demand forecasting is a critical challenge across many industries including retail, logistics, manufacturing, and energy. It involves predicting future demand from historical sales data to optimize inventory, plan staffing, and develop production schedules. Traditional statistical methods (moving averages, exponential smoothing, ARIMA) are relatively easy to implement but have limitations in capturing complex patterns (seasonality, trends, external factor impacts). Deep learning-based methods (DeepAR, Transformer) offer high accuracy but require ML expertise and model tuning. Amazon Forecast is a managed service that automatically selects the optimal algorithm through AutoML and generates high-accuracy forecasts simply by importing time series data. You can leverage technology refined through Amazon.com's retail forecasting without ML expertise.
Data Import and Forecast Generation
Using Forecast begins with importing three datasets. The target time series (required) is the data to forecast, consisting of three columns: timestamp, item ID, and value (sales count, demand volume, etc.). Related time series (optional) are external factors that influence forecasts, including weather data, promotion information, and price changes. Item metadata (optional) are item attributes (category, brand, color, etc.) used for sharing patterns between similar items. After uploading data to S3 and importing it into a dataset group, you create a Predictor (forecast model). Selecting AutoML causes Forecast to automatically evaluate DeepAR+, Prophet, NPTS, ARIMA, and ETS algorithms and select the most accurate model. You specify the forecast horizon and forecast frequency (daily, weekly, monthly, etc.).
Probabilistic Forecasting and What-if Analysis
Forecast provides not only point forecasts (single predicted values) but also probabilistic forecasts (including prediction uncertainty). It generates forecasts at three quantiles: P10 (10% probability demand falls below this value), P50 (median), and P90 (90% probability demand falls below this value). In inventory management, you can use P50 to determine regular order quantities and P90 to set safety stock levels. Products where stockouts are unacceptable use P90-based ordering, while products where inventory cost minimization is prioritized use P50-based ordering, supporting business-requirement-driven decision making. What-if analysis generates scenario forecasts such as "demand if a promotion is run next month" or "demand if prices are reduced by 10%." By comparing forecasts with modified related time series values, you can evaluate the impact of initiatives in advance. To broaden your machine learning knowledge, you can also explore specialized books on Amazon.
Accuracy Evaluation and Pricing
When creating a Predictor, Forecast automatically performs backtesting (splitting historical data for validation) and evaluates forecast accuracy. Accuracy metrics including WAPE (Weighted Absolute Percentage Error), RMSE (Root Mean Squared Error), and MAPE (Mean Absolute Percentage Error) are provided for quantitative model quality assessment. If accuracy is insufficient, improvements can be made by adding related time series, preprocessing data (filling missing values, removing outliers), or changing forecast frequency. Pricing is $0.60 per 1,000 forecasts, $0.24 per training hour, and $0.088/month per GB of data storage. Generating monthly demand forecasts for 1,000 items costs just a few dollars per month. Compared to building custom forecast models with SageMaker, development effort and infrastructure management overhead are significantly reduced.
Summary - Forecast Usage Guidelines
Amazon Forecast is a service that generates high-accuracy demand forecasts from time series data without ML expertise. Its key strengths are AutoML-based automatic model selection, related data utilization, probabilistic forecasting, and What-if analysis. It is ideal for use cases requiring time series-based decision making, such as retail demand forecasting, inventory optimization, and staffing plans. If you have over a year of accumulated sales data, we recommend validating forecast accuracy with Forecast.