DeepAR Influenza Forecast

This repository contains code and resources for training and evaluating a DeepAR model for probabilistic influenza forecasting.

Table of Contents

• Overview
• Repository Structure
• Important Modules and Code Segments

Overview

The DeepAR Influenza Forecast project aims to produce probabilistic forecasts for influenza activity in German districts based on historical data and covariates. The project uses the DeepAR algorithm, which is a popular deep learning model for time series forecasting introduced by Salinas et al. (2020). In contrast to the Amazon Sagemaker implementation, we incorporate the DeepAR model through the GluonTS library. Lastly, we compare the DeepAR model to the Simple Feedforward Neural Network from GluonTS as well as to the hhh4 model from the surveillance R-package.

Repository Structure

This repository is made up of the following directories, which contain:

• HyperparameterFiles: the result-files of hyperparameter runs.
• Notebooks: all jupyter notebooks and in particular notebooks that:
  • DataProcessing/: showcase and include the data.
  • EarlyNotebooks/: display the general workflow and early stage developments.
  • FigureCreation/: are used to create figures and visualizations for the thesis/model evaluation.
  • FurtherResearch/: are outside of the scope of the thesis.
  • Modeltuning/: were used to produce and evaluate the hyperparameter training.
  • OldNotebooks/: are outdated and don't showcase relevant information.
• PythonFiles/: Python files that define important functionalities, which are accessed continuosly by notebooks.
• R/: R files, used to implement the hhh4 model from the surveillance package.

Important Modules and Code Segments

A good first start to grasp our implementation is with the StepByStepGuide.ipynb notebook. However, our final implementation looks a bit different from this.

DataProcessing

EarlyNotebooks

FigureCreation

FurtherResearch

Modeltuning

OldNotebooks

PythonFiles

R