Improving Forecast Accuracy with Machine Learning

Forecasting is an essential business function that allows organizations to develop informed strategies based on past data points. Predicting demand can be critical for running an efficient business because acurate demand forecasting helps minimize overprovisioning and underprovisioning, optimize profitability, and increase customer satisfaction. Forecasting can be applied to predict retail inventory demand, supply-chain planning, workforce status, web traffic forecasting, and more.

The difference between simply generating a good forecast versus using the most optimal forecast configuration can represent millions of dollars of loss in certain scenarios. For some businesses, the cost of overstocking can be significant when there are excessive carrying costs, a high cost of capital, and/or perishable goods. For other businesses, understocking and missing customer demand can result in significant revenue loss and poor customer experiences. For example, in retail forecasting, forecasts are used to estimate future sales, predict when and how many units must be reordered, and reduce inventory holding costs. Key metrics such as revenue, sales, and cash flow must also be predicted.    

This solution provides an automated pipeline for generating, testing, and comparing Amazon Forecast predictors and forecasts—allowing developers and data scientists to bring models to production by generating, testing, comparing, and iterating on Amazon Forecast forecasts.  

Amazon Forecast outputs probabilistic predictions at three default quantiles to address each business’s sensitivity to over and understocking. A sample forecasting graph is shown on the right. Businesses avoiding overstocking can use the p10 forecast, where the true future demand value is expected to be lower than the predicted value only 10% of the time. Businesses more sensitive to missing customer demand can use the p90 forecast, where the true value is expected to be lower than the predicted value 90% of the time. And, businesses aiming to retain compatibility with their legacy tools, or equal sensitivity to over and understocking, can use the p50 forecast.

Businesses can configure this solution, drag-and-drop formatted demand data into Amazon Simple Storage Service (Amazon S3) to generate forecasts using combinations of related data, and then visualize results in the included Amazon SageMaker Jupyter Notebook.