Comparison of Linear Regression, MLP, 1D CNN, and Graph Neural Networks for Financial Asset Forecasting

Xiaoting Yang

2025

Abstract

Recent advances in deep learning have led to powerful new tools for modeling complex, nonlinear patterns in financial markets. This study conducts a head-to-head comparison of four distinct forecasting approaches—linear regression, multilayer perceptron (MLP), one-dimensional convolutional neural network (1D CNN), and graph neural network (GNN)—to predict next-day adjusted closing prices for two equities (Amazon and Netflix), one consumer-goods stock (Domino’s Pizza), and one cryptocurrency (Bitcoin). The results demonstrate that while all four methods achieve similarly high accuracy on the more stable equity series (R² ≈ 0.96–0.97), the nonlinear neural models—particularly the MLP and 1D CNN—offer clear advantages for the highly volatile Bitcoin series (R² ≈ 0.92–0.93 compared to ≈ 0.86–0.88 for the linear and graph-based models). To shed light on each model’s decision process, this paper employ SHapley Additive exPlanations (SHAP) and find that the most recent price lag (the prior day’s close) consistently carries the greatest predictive weight across all methods. These findings highlight both the strengths and limitations of deep learning approaches in one-step financial forecasting and underscore the value of interpretability techniques for understanding model behavior.

Paper Citation

in Harvard Style

Yang X. (2025). Comparison of Linear Regression, MLP, 1D CNN, and Graph Neural Networks for Financial Asset Forecasting. In Proceedings of the 2nd International Conference on Engineering Management, Information Technology and Intelligence - Volume 1: EMITI; ISBN 978-989-758-792-4, SciTePress, pages 84-91. DOI: 10.5220/0014321000004718

in Bibtex Style

@conference{emiti25,
author={Xiaoting Yang},
title={Comparison of Linear Regression, MLP, 1D CNN, and Graph Neural Networks for Financial Asset Forecasting},
booktitle={Proceedings of the 2nd International Conference on Engineering Management, Information Technology and Intelligence - Volume 1: EMITI},
year={2025},
pages={84-91},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0014321000004718},
isbn={978-989-758-792-4},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 2nd International Conference on Engineering Management, Information Technology and Intelligence - Volume 1: EMITI
TI - Comparison of Linear Regression, MLP, 1D CNN, and Graph Neural Networks for Financial Asset Forecasting
SN - 978-989-758-792-4
AU - Yang X.
PY - 2025
SP - 84
EP - 91
DO - 10.5220/0014321000004718
PB - SciTePress