Comparative Analysis of LSTM, FB Prophet, and Moving Average Methods for Fuel Sales Prediction: A Time Series Forecasting Approach

Fadhilah, Ahmad Rizky
Nasution, Arbi Haza
Monika, Winda

Building of Informatics, Technology and Science
Vol 6 No 3 (2024): December 2024

Fuel is an important part of vehicles and machinery where sales demand is very high and has various fluctuations. The uncertainty in these fuel sales patterns poses serious problems in inventory management and fuel distribution planning in Indonesia, which can result in excess stock or fuel scarcity in various regions. Additionally, changing trends in vehicle usage and the impact of the COVID-19 pandemic have made accurate sales predictions increasingly difficult. Therefore, this research aims to understand the current and future sales patterns and trends of fuel sales in Indonesia. Careful analysis of prices and other factors such as data processing and other variables is required. This study uses time series analysis methods and compares four models, namely Long Short-Term Memory (LSTM), FB Prophet, Simple Moving Average (SMA), and Exponential Moving Average (EMA). By comparing the results using statistics such as Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE) over various prediction time frames, we assessed the patterns of each model. The results of the analysis show that the LSTM model outperformed all other methods with the lowest MAPE for the prediction of gasoline in the next 31 days, which is 17.11%, while the FB Prophet outperformed all other methods with the lowest MAPE for the prediction of diesel in the next 31 days, which is 18.32%. Although the LSTM model generally outperformed all other algorithms, the FB Prophet model can be used to predict future trends, such as increased use of diesel and decreased use of gasoline which are expected to last within one year. This analysis also provides insights for choosing the right model for a time series problem, including the characteristics of the data to be predicted and analyzed, as well as the assumptions of stationarity and normality of the data. The results of this study indicate that machine learning algorithms can improve the accuracy of time series predictions significantly compared to traditional statistical methods.