How Amazon Forecast Works
You use the following resources while developing forecasting projects in Amazon Forecast:
- Dataset Import: Datasets are collections of your supplied data. Dataset groups are collections of datasets that include complementary information. Forecast algorithms train custom forecasting models, known as predictors, using your dataset groups.
- Predictor Training: Predictors are custom models that have been trained on your data. You can train a predictor by selecting a prebuilt algorithm or by letting Amazon Forecast choose the optimal algorithm for you.
- Forecasting: You can create forecasts for your time-series data, query them with the QueryForecast API, and view them in the console.
Importing Dataset
Datasets include the information needed to train a predictor. You construct one or more Amazon Forecast datasets and populate them with your training data. A dataset group is a collection of related datasets that detail a set of changing parameters across time. You utilize a dataset group to train a predictor after you've created it.
Each dataset group may contain up to three datasets, one of each of the following types: target time series, associated time series, and item metadata.
You can use the Forecast console, AWS Command Line Interface (AWS CLI), or AWS SDK to generate and manage Forecast datasets and dataset groups.
Using Related Time-Series Dataset
A related time series dataset contains time series data that is not included in the target time series dataset and may increase the predictor's performance.
In the demand forecasting area, for example, a target time series dataset would have the dimensions timestamp and item id, whereas a complementary related time series dataset would also include the following supplementary features: the item price, promotion, and weather.
A related time series dataset can have up to ten forecast dimensions (the same as your target time series dataset) and up to thirteen related time series characteristics.
Using Item Metadata Datasets
An item metadata collection comprises categorical data that adds context to items in a target time-series dataset. Item metadata datasets, unlike related time-series datasets, give static information. That is, the data values, like the color or brand of an item, remain consistent over time. Item metadata datasets can be added to your dataset groups as an optional. You can use an item metadata only if every item in your target time-series dataset is present in the corresponding item metadata dataset.
Item metadata may comprise a specific item's brand, color, model, category, location of origin, or other supplementary attributes. An item metadata dataset, for example, could provide context for some of the demand data included in a target time-series dataset representing sales of black Amazon e-readers with 32 GB of storage. These qualities belong in an item metadata dataset since they do not change from day to day or hour to hour.
Item metadata can help you uncover and track descriptive trends in your time-series data. A forecast can train the model to produce more accurate predictions based on similarities between items if you include an item metadata dataset in your dataset group. For example, you may discover that Amazon's virtual assistant goods are more likely to sell out than those manufactured by other companies, and you can then design your supply chain appropriately.
Recommended read: Amazon Hirepro
Predefined Dataset Domains and Dataset Types
To train a predictor, you must first build one or more datasets, group them together, and then supply the dataset group for training.
You assign a dataset domain and a dataset type to each dataset you create. A dataset domain defines a pre-defined dataset schema for a typical use case but has no effect on model algorithms or hyperparameters.
Amazon Forecast supports the dataset domains listed below:
- RETAIL Domain: For anticipating retail demand.
- Domain INVENTORY_PLANNING: For supply chain and inventory planning.
- EC2 CAPACITY Domain: This domain is used to forecast Amazon Elastic Compute Cloud (Amazon EC2) capacity.
- WORK_FORCE Domain: For workforce planning.
- WEB_TRAFFIC Domain: Used to forecast future web traffic.
- METRICS Domain: This domain is used for forecasting metrics such as revenue and cash flow.
- CUSTOM Domain: For any other time-series forecasting applications.
Each domain may have between one to three dataset kinds. The dataset types you construct for a domain are determined by the type of data you have and the features you wish to use in training.
Each domain necessitates a target time series dataset and may also support related time series and item information dataset types.
Types of dataset
- Target time-series: This type specifies the target field for which forecasts are to be generated. For example, if you want to anticipate sales for a group of products, you must first construct a dataset of historical time-series data for each product. Similarly, you can establish a target time series dataset for indicators that you want to forecast, such as revenue, cash flow, and sales.
- Related time-series: Time-series data that is related to the target time series data. For example, price is tied to product sales data, therefore you may supply it as a time series.
- Item metadata: Metadata relevant to the target time-series data. For example, if you are estimating sales for a certain product, attributes such as brand, color, and genre will be included in item metadata. When forecasting EC2 capacity for EC2 instances, metadata such as CPU and memory of the instance types may be used.
Also check out - Amazon Forecast Part 2.
Frequently asked questions
What method of forecasting Does Amazon use?
Amazon Forecast is a time-series forecasting service based on machine learning (ML) and built for business metrics analysis.
How does Amazon forecast its sales?
Amazon Forecast forecasts future business outcomes for FBA sellers, such as resource requirements, product demand, and financial success, using machine learning technologies. This fully-managed service's built-in technologies analyze your Amazon FBA historical data to find patterns and forecast future sales.
What is the P90 forecast on Amazon?
There's a 90% probability that weekly consumer demand will be at the level indicated and a 10% chance Amazon will buy more.
Conclusion
In this article, we have discussed the following topics:
- Features of Amazon Forecast
- Working of Amazon Forecast
- Predefined Dataset Domains and Dataset Types
- Using item metadata datasets
Want to learn more about Data Analysis? Here is an excellent course that can guide you in learning.
Check out the Amazon Interview Experience to learn about Amazon’s hiring process.
Happy Coding!