Demand Forecasting with Amazon Forecast - Ingesting Time Series Data and Improving Prediction Accuracy
Learn how to build time series forecasting models with Forecast, leverage related data, and export prediction results.
Demand Forecasting Challenges and Forecast Positioning
Demand forecasting is a critical challenge across retail, logistics, manufacturing, and energy industries. Traditional statistical methods (moving average, exponential smoothing, ARIMA) are relatively easy to implement but have limitations capturing complex patterns such as seasonality, trends, and external factor effects. Deep learning-based approaches (DeepAR, Transformer) achieve high accuracy but require ML expertise and model tuning. Amazon Forecast is a managed service that automatically selects the optimal algorithm via AutoML and generates high-accuracy forecasts simply by importing time series data. It leverages the same ML algorithms used for demand forecasting at Amazon.com, enabling high-accuracy demand forecasting without ML expertise. It provides probabilistic forecasts with P10, P50, and P90 prediction intervals, supporting risk-aware decision-making through uncertainty quantification.
Data Preparation and Model Building
Forecast usage begins with importing data into Dataset Groups. The target time series (required) is the forecast target data, consisting of three columns: timestamp, item ID, and value (sales count, demand volume). Related time series (optional) includes external factors such as weather data, promotion information, and price changes. Item metadata (optional) includes item attributes such as category, brand, and color, used for pattern sharing between similar items. Data is placed in S3 in CSV format, and a schema definition (JSON) specifies the attribute type and data type for each column. Timestamp granularity (1 min/5 min/15 min/1 hour/1 day/1 week/1 month/1 year) is set according to business requirements and must match the forecast horizon unit. AutoML builds models with 6 algorithms - DeepAR+, Prophet, ETS, NPTS, ARIMA, and CNN-QR - then automatically selects the most accurate model through backtesting.
Probabilistic Forecasts and What-If Analysis
Forecast provides not just point forecasts (single prediction 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). For inventory management, you can use P50 for standard order quantities and P90 for safety stock levels. Products where stockout risk is unacceptable use P90-based ordering, while products where inventory cost minimization is prioritized use P50-based ordering, supporting business-specific decision-making. What-If analysis generates scenario forecasts such as "demand if promotions run next month" or "demand if prices are reduced by 10%." Multiple scenarios with modified related time series values can be evaluated in parallel to quantitatively assess the impact of initiatives in advance. For practical techniques related to Forecast, you can also check out related books on Amazon.
Improving Prediction Accuracy
Forecast prediction accuracy depends heavily on input data quality and the richness of related data. Adding related time series data (weather, holidays, promotions) allows the model to learn demand fluctuations caused by external factors, improving accuracy. Forecast includes built-in holiday information (Holidays featurization) that automatically accounts for major holidays simply by specifying a country code. Item metadata (category, brand, price range) is effective for mitigating the cold start problem (predicting demand for new products), estimating demand patterns for new items from the performance of similar items. Predictor backtest results provide accuracy metrics including WAPE (Weighted Absolute Percentage Error), RMSE (Root Mean Squared Error), and MAPE (Mean Absolute Percentage Error) for quantitative model quality assessment. If accuracy is insufficient, improvements can be pursued through adding related time series, data preprocessing (filling missing values, removing outliers), or changing prediction frequency.
Use Cases and Industry Applications
Forecast is used across a wide range of industries including retail, manufacturing, logistics, and energy. In retail, SKU-level demand forecasting optimizes inventory placement, reducing both overstock and stockouts. In manufacturing, raw material order quantity forecasting creates procurement plans that account for procurement lead times, contributing to cost optimization across the entire supply chain. In logistics, worker shift demand forecasting optimizes staffing, preventing understaffing during busy periods and overstaffing during slow periods. In the energy industry, power consumption forecasting supports generation planning and grid management that accommodates renewable energy output variability. For web services, it can be applied to server capacity planning, optimizing auto-scaling configurations based on traffic forecasts.
Pricing and Quota Considerations
Pricing consists of $0.60 per 1,000 forecasts, $0.24 per training hour, and $0.088 per GB per month for data storage. For generating monthly demand forecasts for 1,000 items, costs start at just a few dollars per month. AutoML trains multiple models so training costs are higher than manually specifying an algorithm. If prediction accuracy is sufficient, reduce costs by changing training frequency from weekly to monthly. Compared to building custom forecasting models with SageMaker, Forecast significantly reduces development effort and infrastructure management burden. For quotas, the number of dataset groups per account is limited to 500, and predictors are limited to 500 per region. Dataset import size is capped at 10 GB, and the forecast horizon depends on data granularity with a maximum of 500 time steps. Training time for large datasets (millions of rows) may require several hours, so build slack into operational schedules.
Summary
Forecast is a managed service that automatically builds ML models from time series data for demand forecasting. It overcomes limitations of traditional statistical methods through ML, with AutoML automatically selecting the optimal model from 6 algorithms. Adding related time series data and built-in holiday information improves prediction accuracy, while probabilistic forecasts provide P10/P50/P90 prediction intervals. What-If analysis supports scenario-based decision-making, applicable across industries from retail inventory optimization to manufacturing procurement planning and energy grid management. If at least one year of historical sales data has been accumulated, evaluating prediction accuracy with Forecast is recommended.