Modeling and Forecasting Performance of ARIMA and ANN Models in the Presence of High-Frequency Data. Application to the East African Currency Exchange Rates
| dc.contributor.author | ROSELINE ONDIEKI | |
| dc.date.accessioned | 2026-03-30T15:49:19Z | |
| dc.date.available | 2026-03-30T15:49:19Z | |
| dc.date.issued | 2024-09 | |
| dc.description | Thesis | |
| dc.description.abstract | More and more high frequency time series (instead of quarterly/monthly) are being encountered and have resulted to longer and complex time series data. In modeling such data, more parameters are expected hence the need to use approaches which can capture the inherent structure of the data accurately. Existing models such as ARIMA and ANN have been applied to model mostly monthly and annual time series data with no focus on how the time series data frequency affects both their modeling and predictive performance and remedies established. This study sought to establish the modeling and forecasting performance of the two methodologies in the presence of high-frequency data and justify the need for Mixture models to model and predict such data. The study also applied the methods studied to modeling and forecasting the currency exchange rates for the East African countries. Additionally, the study endeavored to enrich the existing literature on mixture modeling approaches and their effectiveness in forecasting high-frequency time series data. The Ljung Box test was used to test for the lack of fit in the ARIMA, ANN, and Mixture models while RMSE and MAPE were used to compare the prediction performance of the models. The results established that ARIMA approach became weaker in accurately modeling and forecasting time series data whose frequency was greater than 48 while ANN was able to precisely handle time series data whose frequency was up to 336. Further, the study results showed that the mixture ARIMA-ANN model provided the best forecast accuracy compared to ARIMA and ANN models for high-frequency data and produced better fitting models even for hourly time series data, a phenomenon that could not be achieved by the single models. When applied to currency exchange rates, the study established ���������� − ������ (1,1,0)(5: 1: 2) 250 , ���������� −������(0,1,1)(5: 1: 2) 250 , ���������� − ������ (0,1,0)(5: 1: 2) 250 and ���������� −������ (1,1,0)(5: 1: 2) 250 respectively for the KSH/RWF, KSH/TSHS, KSH/USHS and KSH/BIF currency exchange rates. The models had lower MAPE and RMSE values compared to the corresponding ARIMA and ANN models. | |
| dc.identifier.uri | https://repository.cuea.edu/handle/123456789/594 | |
| dc.language.iso | en_US | |
| dc.publisher | THE CATHOLIC UNIVERSITY OF EASTERN AFRICA | |
| dc.subject | ARIMA models | |
| dc.subject | Artificial Neural Networks (ANN) | |
| dc.subject | time series forecasting | |
| dc.subject | exchange rate prediction | |
| dc.subject | high-frequency data | |
| dc.title | Modeling and Forecasting Performance of ARIMA and ANN Models in the Presence of High-Frequency Data. Application to the East African Currency Exchange Rates | |
| dc.type | Thesis |