DeepAR Forecasting Algorithm
The SageMaker DeepAR Forecasting Algorithm forecasts how the target time series will evolve based on past performance. AR, which stands for AutoRegression, is a statistical method that uses observations from previous time steps as input to a regression equation to predict the value at the next time step. The forecast is a one dimensional time series.
The input data can contain more than one set of time series data that are related and created by similar methods. DeepAR works best when there are at least 300 observations per dataset and there are hundreds of related datasets. Otherwise more traditional approaches may produce better results. The SageMaker DeepAR algorithm was launched in January 2018
The DeepAR Forecasting Algorithm uses Recurrent Neural Networks to outperform regular algorithms by using multiple time series, rather than a single one.
Attributes
| Problem attribute | Description |
| Data types and format | Tabular, numeric time series data |
| Learning paradigm or domain | Supervised Learning |
| Problem type | Time-series forecasting |
| Use case examples | Based on historical data for a behavior, predict future behavior |
Training
DeepAR Forecasting Algorithm training data is presented in JSON Lines format which can be compressed in gzip or parquet.
- json
- json.gz
- parquet
Model artifacts and inference
| Description | Artifacts |
| Learning paradigm | Supervised Learning |
| Request format | JSON |
| Result | JSON |
Processing environment
Training can be done on CPU or GPU. GPU should be used when the training data is large enough to benefit from the additional processing power.
Inference uses CPU instances only.
Video
Amazon SageMaker’s Built-in Algorithm Webinar Series: DeepAR Forecasting
A 53.40 minutes video from AWS.
Credits
Dow Jones graph Markus Spiske on Unsplash
