Nicolas Kuhaupt - Probabilistic Forecasting with DeepAR and AWS SageMaker

Discover how to use DeepAR and AWS SageMaker for probabilistic forecasting, leveraging scalability and advanced features like dynamic inputs, transfer learning, and optional categories, to improve accuracy and reliability.

Key takeaways

DeepAR is a prepared algorithm for forecasting
SageMaker is a machine learning service by AWS *ufenability to access S3 storage is crucial
Hyperparameters, such as instance name, need to be set correctly
Dynamic features can be included, like weather or temperature
LSTMs can only do point forecasts, not probabilistic ones
DeepAR is unique in learning from multiple time series
Neural networks are time and resource intensive to train
There are optional features, like category and dynamic feed
Autoregressive comes from the autoregressive nature of time series
The algorithm also includes transfer learning
The probability function has two parameters: mean and standard deviation
Time series can be handled by scaling inputs before putting them into the network

Nicolas Kuhaupt - Probabilistic Forecasting with DeepAR and AWS SageMaker

More talks