Forecasting For the West Texas Intermediate Crude Oil Prices with Extreme Values Treatment

Abstract

"Crude oil, often called ""oil,"" is a fossil fuel and a naturally occurring liquid hydrocarbon found underground in different geological formations. It has various types of grades; the West Texas Intermediate Index was used in this study. Usually, fluctuations in crude oil prices affect people, and lead governments to manage their budgets and make policies, while businesses and firms can plan for things like production and investments in either a negative or positive way. Therefore, forecasting crude oil prices for the future has become a major topic in the investment field. Hence, this study was conducted to answer those questions. The main objective of this study is to compare the predictive accuracy of the top-performing extreme value detection technique and the top-performing extreme value replacement technique by using classical time series models and deep learning architectures to forecast West Texas Intermediate crude oil prices. Therefore, five extreme value detection techniques were used: Fixed Price Threshold, Standard Deviation Filter on Prices, Moving Window Filter on Prices, Recursive Filter on Prices, and Percentage Price Filter. After identifying spikes in the dataset, those were replaced using four different replacement techniques: mean, median, damping scheme, and threshold scheme. After doing all the data preprocessing steps, 21 datasets were finalized for the modeling phase. To be more fair, statistical and deep learning architectures were used in this study. As a statistical model, Autoregressive Integrated Moving Average was selected because the data did not follow any seasonality or stationarity. After fitting the statistical model, deep learning architectures such as Recurrent Neural Networks, Long Short-Term Memory, and Gated Recurrent Units were used, and the best model was selected by referring to the Mean Squared Error and Root Mean Squared Error values. Finally, it was confirmed that the Long Short-Term Memory model with a Fixed Price Threshold with damping scheme replacement was the best out of 21 models. It obtained a Mean Squared Value of 4.23 and a Root Mean Squared Value of 2.03. A Python Flask application was finally developed by taking the best model that can be used to forecast crude oil prices up to 100 days accurately."