Comparative Evaluation of Statistical, Machine Learning, Deep Learning, and Bayesian Optimization Models for Predicting Crude Oil Production in Pakistan

Authors:

Abstract
Energy planning, policymaking, and economic stability depend on accurate crude oil production (COP) forecasts, especially in developing nations like Pakistan. Traditional linear forecasting methods are ineffective for COP time series due to nonlinearity, structural breakdowns, and regime-dependent dynamics. This research compares statistical, machine learning (ML), deep learning (DL), and Bayesian optimization (BO) models for COP forecasting. The article uses Pakistani monthly COP data from July 2005 to March 2024. Through the investigation of statistical models like ARIMA, SETAR, and State Space, a comprehensive forecasting framework has been created. Machine learning models including Random Forest (RF), XGBoost, and Extreme Learning Machine (ELM) were added, along with deep learning architectures like LSTM and Gated Recurrent Unit. RMSE, MAE, and MAPE were used to evaluate model performance for in-sample and out-of-sample forecasting. Deep learning models exceed most traditional and machine learning COP forecasters. N-HiTS had the lowest out-of-sample RMSE at 0.0459, followed by XGBoost at 0.0517 and SETAR at 0.0574. This research presents a consistent comparison methodology for COP prediction across statistics, machine learning, deep learning, and Bayesian optimization-enhanced models in energy forecasting. It offers advice on choosing feasible forecasting methods for nonlinear and data-constrained energy applications.

This publication has 18 references indexed in Scilit: