Forecasting Natural Gas Futures Price Movements Using Machine Learning and Deep Learning Models: A Comparative Study

초록

Purpose: This study aims to evaluate the predictive performance of traditional statistical, machine learning (ML), and deep learning (DL) models in forecasting natural gas futures price movements. In particular, it investigates the role of hybrid modeling using both structured time-series data and unstructured sentiment data to address volatility and non-linearity in energy markets. Design/methodology/approach: The research integrates SARIMA models with ML techniques such as Logistic Regression, Decision Trees, and Random Forests, along with DL architectures including Neural Networks, Recurrent Neural Networks (RNNs), and Convolutional Neural Networks (CNNs). The dataset includes daily records of natural gas futures from January 2023 to June 2024, alongside sentiment data extracted from online forums. Textual sentiment data were preprocessed through tokenization, stop-word removal, and lemmatization. For CNN models, visual inputs such as word clouds and trend graphs were employed to reflect sentiment and price dynamics. Findings: Neural Networks achieved the highest average accuracy (86.52%), followed by SARIMA (81.54%), confirming the strength of traditional models in capturing seasonal trends. Random Forests exhibited robust performance (78.26%), while sentiment-enhanced CNNs reached 70.00% (word clouds) and 59.38% (price graphs). Logistic Regression and RNN models achieved 69.39% and 58.12%, respectively. Wilcoxon signed-rank tests showed that the Neural Network significantly outperformed Decision Tree, Logistic Regression, both CNN variants, and RNN (p<.05), while differences versus SARIMA (p=0.492) and Random Forest (p=0.051) were not statistically significant. Research limitations/implications: Although this study demonstrates the utility of combining structured and unstructured data, it is limited to a specific commodity and a defined period. Future studies should explore transformerbased language models and incorporate macroeconomic indicators to improve robustness and generalizability. Originality/value: This paper contributes to the growing literature on hybrid forecasting methodologies in commodity markets. It offers a novel approach by combining time-series analysis, sentiment modeling, and image-based representations. The findings provide actionable insights for practitioners and researchers seeking to enhance forecasting accuracy in highly volatile energy sectors.

키워드