Time Series Forecasting Using a Hybrid Prophet and Long Short-Term Memory Model

Kong, Yih Hern (2023) Time Series Forecasting Using a Hybrid Prophet and Long Short-Term Memory Model. Masters thesis, Tunku Abdul Rahman University of Management and Technology.

Text 09 KONG YIH HERN (MCS).pdf Restricted to Registered users only Download (5MB)

Abstract

Forecasting analysis is a common research topic these days. Development in this area has allowed organizations to retrieve useful information and make important decisions based on forecast results. Time series in a real business environment often contains complicated relationship mixed with linear and nonlinear. Also, the holiday effects are found to have significant effect on the time series data in the economic industry. However, the holiday effects on time series in Food and Beverages (F&B) and Cosmetics industries were not investigated. The impact of holiday effect on the forecast result was also unknown. As such, Hybrid Prophet-LSTM is introduced, which combines Long Short-Term Memory (LSTM) and FBProphet (Prophet), to handle the complicated relationship and to model holiday effects in the time series. The proposed hybrid method combines the capability to model holiday effects in Prophet, and the outstanding performance of handling nonlinear relationship in LSTM. Besides that, this study aims to investigate the effectiveness of the hybrid model. The effectiveness of the forecast using multiple variables, and the influence of the holiday effect were being examined on the forecast result. This study also demonstrated the effects of outliers to the time series and forecast result, further expanding the current research to determine to remove outliers or not. In this study, data from the 7 different Facebook pages were crawled. The pages were arbitrary selected from Cosmetic, Food, and Beverage industries. Two years of daily data, which began from June 2018 to March 2021, were collected from the selected pages as the dataset. The proposed Hybrid Prophet-LSTM is found to outperform both the standalone LSTM and Prophet by having a lower forecast error. Also, the effect of removing outliers in the time series is found to be able to decrease the forecast error significantly. The holiday effects in the selected time series data, on the other hand were found to be having a small influence on the forecast result.