Sales forecasting using WaveNet within the framework of the Kaggle competition

TL;DR

This paper presents a time series sales forecasting method using WaveNet, adapted with dilated convolutions, achieving high accuracy in a Kaggle competition and serving as a guideline for future research.

Contribution

The paper introduces an adaptation of WaveNet with dilated convolutions for efficient and accurate sales forecasting in a competitive setting.

Findings

Achieved 2nd place in Kaggle competition

Demonstrated effectiveness of dilated CNNs for time series forecasting

Provided insights and guidelines for future research in the field

Abstract

We took part in the Corporacion Favorita Grocery Sales Forecasting competition hosted on Kaggle and achieved the 2nd place. In this abstract paper, we present an overall analysis and solution to the underlying machine-learning problem based on time series data, where major challenges are identified and corresponding preliminary methods are proposed. Our approach is based on the adaptation of dilated convolutional neural network for time series forecasting. By applying this technique iteratively to batches of n examples, a big amount of time series data can be eventually processed with a decent speed and accuracy. We hope this paper could serve, to some extent, as a review and guideline of the time series forecasting benchmark, inspiring further attempts and researches.