An Evaluation of the Empirical Performance of GRU-Based Models in Financial Time Series of the Global Automotive Industry
Öz
This study employs a deep learning-based GRU (Gated Recurrent Unit) model to forecast the closing stock prices of BMW, Ferrari, Ford Motor Company, General Motors, Suzuki, Tesla, Toyota, and Volkswagen Group, as well as the macro-financial variables DXY, EUR/USD, USD/CNY, and USD/JPY. Given the cyclical demand structure of the global automotive sector and its sensitivity to systematic risks in financial markets, accurately modeling these variables is important for both academic research and investment decisions. The GRU architecture was systematically analyzed using 1-, 2-, 3-, and 4-layer configurations, 32, 64, and 128 neurons, and 30-, 60-, and 90-day time windows. The dataset covers the period from October 21, 2015, to April 30, 2026, and was divided into 80% training and 20% evaluation data, including validation and test sets. In all experiments, the number of epochs, batch size, and dropout rate were fixed at 50, 32, and 0.25, respectively. Model performance was evaluated using MAE, MSE, and MAPE. The findings indicate that GRU performance is highly sensitive to dataset-specific dynamics and hyperparameter configurations. Overall, 1- and 2-layer models produced lower error values, whereas 3- and 4-layer architectures generally did not improve performance across most series. Increasing the number of neurons did not provide a systematic advantage, and optimal configurations varied across variables. Time window length emerged as a critical factor, particularly for volatile series. Descriptive statistics revealed substantial scale and volatility heterogeneity, while Pearson correlation analysis indicated strong positive co-movement among automotive stocks. The inclusion of macro-financial variables produced variable-specific differences in forecasting performance. Overall, the results demonstrate that the forecasting performance of GRU-based models is determined more by data-driven hyperparameter optimization than by model depth or capacity.
Anahtar Kelimeler
Deep Learning, Exchange Rates, Financial Time Series, Automotive Sector, GRU, Layer, Neuron
Kaynakça
- Alkhatib, K., Khazaleh, H., Alkhazaleh, H. A., Alsoud, A. R., & Abualigah, L. (2022). A New Stock Price Forecasting Method Using Active Deep Learning Approach. Journal of Open Innovation: Technology, Market, and Complexity, 8(2), 96. https://doi.org/10.3390/joitmc8020096
- Chen, Q., & Kawashima, H. (2025). Modeling Inter-Firm Dependencies with Temporal Graph Neural Networks for Stock Price Prediction. This paper has been peer-reviewed, revised, and resubmitted to a journal for consideration. http://dx.doi.org/10.2139/ssrn.6605998
- Çolak, Z. (2025). Stock Price Prediction Using Deep Learning Models: A Comparative Analysis of LSTM, GRU, RNN, and MLP Models. Journal of Management Sciences, 23(56), 1250-1286. https://izlik.org/JA49FM38JR
- Dey, M. K., Dey, S., & Das, D. K. (2024). Indian Stock Price Prediction Using an Optimal Gated Recurrent Unit (GRU) Network. In 2024 International Conference on Advancement in Renewable Energy and Intelligent Systems (AREIS) (pp. 1–6). IEEE. doi: 10.1109/AREIS62559.2024.10893641
- Diqi, M., Hiswati, M. E., & Wijaya, N. (2024). Stacked Gated Recurrent Units and Indonesian Stock Predictions: A New Approach to Financial Forecasting. IKOMTI, 5(1), 11–17. https://doi.org/10.35960/ikomti.v5i1.1106
- Friday, I. K., Godslove, J. F., Nayak, D. S. K., & Prusty, S. (2022). IRGM: An integrated RNN-GRU model for stock market price prediction. In 2022 International Conference on Machine Learning, Computer Systems and Security (MLCSS) (pp. 129–132). IEEE. doi: 10.1109/MLCSS57186.2022.00031.
- Gao, Y., Wang, R., & Zhou, E. (2021). Stock prediction based on optimized LSTM and GRU models. Scientific Programming, 2021, Article 4055281. https://doi.org/10.1155/2021/4055281
- Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press. DOI: https://doi.org/10.4258/hir.2016.22.4.351
- İncekırık, A. (2026). Analysis of Gold Price Prediction Using Commodity and Foreign Exchange Market Indicators with Unidirectional and Bidirectional Deep Learning Models. Scientific Culture. 12(3.1), 182-196. DOI: 10.5281/zenodo.18908963
- James, J. Q., Hill, D. J., Lam, A. Y., Gu, J., & Li, V. O. (2017). Intelligent time-adaptive transient stability assessment system. IEEE Transactions on Power Systems, 33(1), 1049–1058. doi: 10.1109/TPWRS.2017.2707501