CCE Theses and Dissertations

Date of Award


Document Type


Degree Name

Doctor of Psychology (PhD)


College of Computing and Engineering


Sumitra Mukherjee

Committee Member

Michael J. Laszlo

Committee Member

Francisco J. Mitropoulos


Time series forecasting is an area of research within the discipline of machine learning. The ARIMA model is a well-known approach to this challenge. However, simple models such as ARIMA do not take into consideration complex relationships within the data and quite often fail to produce a satisfactory forecast. Neural networks have been presented in previous works as an alternative. Neural networks are able to capture non-linear relationships within the data and can deliver an improved forecast when compared to ARIMA models.

This dissertation takes neural network variations and applies them to a group of time series datasets found in the literature to look for forecasting improvements and generalizability. Metrics used to compare the effectiveness of the variations will be taken from the literature and include the Root Mean Squared Error (RMSE), Directional Accuracy (DA), and Mean Absolute Percentage Error (MAPE).

A total of 12 datasets were used for this study: 6 series each with a daily and weekly version. Analysis of the results demonstrates that it is possible to improve performance as gauged by the metrics in most instances. Neural networks with a feature detection component such as a convolutional layer or a temporal component such as RNN variations are effective when scored by the directional accuracy metric. Convolutional layers appear to be especially effective at the weekly level of granularity in this study. The Stacked Denoising Autoencoder (SDAE) performed well when judged by the RMSE and MAPE metrics.

The directional accuracy metric was further broken down into a classification problem: precision, recall, and F1 metrics were used for this evaluation. In addition, the research included evaluating the models’ ability to predict multiple steps ahead: steps t+1, t+2, and t+3 were examined. The predictive power of the models generally decreased as timesteps increased. RNN variations continued to do well at timesteps beyond t+1 for directional accuracy. The predictive power of the SDAE held up well beyond the t+1 step and dominated the MAPE and RMSE metrics at steps t+2 and t+3.