Development of a Machine Learning Algorithm for Mixed-Frequency Time Series Forecasting

Abstract

Machine learning and especially neural networks, has been the dominant form of forecasting for various applications. In many real world applications, the training data for these models come from the same source (e.g., product SKU codes, electricity consumption data, etc.) and therefore share similar seasonal patterns that can be detected by neual networks. However, certain applications require forecasting a set of heterogeneous time series, from different domains and with different sampling frequencies, making recognizing seasonality challenging. In addition, seasonal patterns are often found at multiple levels, such as weekly, monthly, or yearly, further complicating the learning process. To enable the training of neural networks, data preprocessing techniques are usually applied before feeding the data into the model. The purpose of this thesis is to develop a neural network architecture that automatically accounts for the seasonality of each time series and produces accurate forecasts for any set of heterogeneous time series, without requiring extensive preprocessing techniques.