AWS for Industries
AWS Service Spotlight: Amazon Forecast for accurate demand forecasting
Forecasting is the key to maintaining an effective balance between meeting customer demand and effectively using your resources. An undersupply can lead to disappointed shoppers while an oversupply will sit on shelves until sold or discarded (if passed its useful shelf life). Businesses use everything from spreadsheets, workflows, to complex resource planning software to generate forecasts. However, traditional forecasts struggle to incorporate the many signals that inform demand, resulting in costs from expediting (to meet demand) or excess inventory (continuing to burden constrained supply chain capacity).
Amazon Forecast is a fully managed time-series forecasting service that allows businesses to utilize deep learning and proven statistical models with little to no prior machine learning (ML) experience. Let’s discuss how Amazon Forecast separates itself from traditional tools, and key information needed to start using the service.
Making deep learning models accessible
Traditional forecasting tools use statistical models that excel in specific scenarios but don’t always make use of all the information available in a data rich ecosystem. As businesses continue to centralize and democratize their data, deep learning forecasting models are best equipped to take advantage of industry trends. There are two key differentiators for deep learning models compared to statistical models.
- Traditional forecasts are great at making short-term forecasts with limited data, but they don’t scale as well when using large datasets or when making long-term predictions. Deep learning models are able to better understand trends over long periods of time and have the ability to apply learnings from multiple time-series.
- Traditional forecasts rarely incorporate related but independent data, which can offer important context (such as price, color, events, stock-outs, marketing promotions, and such). Without the full history and the broader context, we are providing a limited view to the forecasting model. Also, statistical models cannot incorporate product information into time-series.
The two deep learning models available with Amazon Forecast were developed to solve challenges for Amazon.com and provide accurate predictions across millions of products.
Picking the right model
Deep Learning models have powerful advantages, but they are not the best option for every scenario. Maximizing their performance requires having a sufficient amount of data and that may not always be the case for different datasets. Each forecasting algorithm has its strengths and weaknesses and deciding between them is a time-consuming task.
Amazon Forecast abstracts the model selection process away with an automatic machine learning (AutoML) feature. When you provide your data and enable AutoML, the service will create an ensemble model by training all six built-in Forecast models and choose the best forming model to provide results. The ability to rapidly train and evaluate different methods of forecasting can improve accuracy up to 40% and decrease time to production by up to 50%.
Getting started
Planning Forecast Development
Amazon Forecast is a service designed to get users going quickly by taking on much of the machine learning process. As a result, it’s important to prepare data properly and plan how a forecast will be used to get the best results. Before generating forecasts, there are key questions that need to be answered to help you determine how you use generated forecasts and how to evaluate their performance.
- Document the flow of inventory: How often do shipments go out to retail locations? This should inform what granularity the forecast is at (daily, weekly, monthly) and how far to forecast out (one month, two months, and so on).
- Identify the goal of the forecast: Will the goal be forecasting how much product to manufacture, or how much will be allocated to retail locations? What are the lead times involved in manufacturing and shipping?
- Identify the business case for forecasting: Where will the benefit be and what are the relevant KPIs to measure. Minimize stock outs, inventory costs, scrapped product.
- Outline the process for evaluating forecast performance: To understand if a forecast is impactful, real world KPIs should be measured against the KPIs that would result from following the forecast.
Gathering Data
Working through these questions will help plan the creation of a forecast and identify metrics that can be used to evaluate performance. Existing processes can also be used to inform these questions.
The next step is to start gathering the necessary data. There are three datasets that influence predictions:
- Target Time-Series: Time-series data for the items you want to make predictions for centered around the target variable you want to forecast. In a retail scenario the target value is typically units sold or total sales. This dataset is required.
- Related Time-Series: Time-series data not included in the target time-series that can help inform predictions. This dataset is optional.
- Item Metadata: Static information about the items in your target time-series dataset. This dataset is optional.
You can learn more about the datasets and the specific schemas in the documentation about importing datasets. Once we have our target time-series data ready, we can start generating forecasts.
End-to-end Forecast flow
Conclusion
We discussed how Amazon Forecast separates itself from traditional forecasting tools by making deep learning accessible, and using AutoML to pick the best model in a given situation. We also discussed what questions should be answered to inform how a forecast is made, and what data is needed to get started.
Amazon Forecast is designed to get teams results as quickly as possible. Anyone can get started with the service and have machine learning make an impact on their business.
Contact an AWS Representative to know how we can help accelerate your business.