نوع مقاله : مقاله علمی پژوهشی
نویسندگان
1 دانشجوی دکترا رشته اقتصاد مالی، دانشکده اقتصاد و مدیریت، دانشگاه ارومیه، ارومیه، ایران.
2 دانشیار گروه اقتصاد، دانشکده اقتصاد و مدیریت، دانشگاه ارومیه، ارومیه، ایران.
3 دانشیار گروه اقتصاد، دانشکده اقتصاد و مدیریت، دانشگاه ارومیه، ارومیه، ایران
چکیده
کلیدواژه‌ها
موضوعات
عنوان مقاله [English]
نویسندگان [English]
Objective: This study aims to evaluate the sensitivity of machine learning models to input variables and identify the most significant factors influencing the prediction of the Tehran Stock Exchange index. Additionally, it compares the performance of different models in forecasting the stock index, focusing on the impact of input parameters, and offers strategies for optimizing input data and reducing model complexity. The key input variables considered include open, high, low prices, and trading volume.
Method: In this research, due to the presence of a target variable (closing price), a supervised learning approach is employed. Furthermore, because the target variable is continuous, the problem is defined as a regression task. Predictions from four powerful machine learning algorithms—Linear Model (LM), Support Vector Regression (SVR), Artificial Neural Network (ANN), and Random Forest (RF)—were compared using performance evaluation metrics such as Mean Absolute Error (MAE), Mean Squared Error (MSE), and the coefficient of determination (R²). MAE represents the average magnitude of errors, while MSE indicates the differences between the predicted and actual values. Lower values for both metrics suggest greater accuracy of predictions. The R² statistic represents the percentage of variance in the data that is explained by the model, with values close to one indicating a higher level of accuracy. The dataset used in this study consists of six selected indices from various sectors, including pharmaceutical, automotive, financial, food industries, basic metals, and petroleum products, covering the period from 2020 to 2024 on a daily basis. Sensitivity analysis was conducted to evaluate the importance of input variables such as open, high, low prices, and trading volume on the predictions of each model.
Findings: The sensitivity analysis revealed that Support Vector Regression (SVR) exhibits lower sensitivity to input parameters. This can be attributed to the use of nonlinear kernels in the SVR model, which allow it to model complex relationships between the variables in a transformed feature space, thus reducing the direct impact of input parameters. In contrast, models such as Random Forest, Artificial Neural Networks (ANN), and Linear Regression demonstrated relatively higher sensitivity to input variables. These models are more directly dependent on the input data and provide a better representation of the relative importance of the variables in predicting the output. ANN, in particular, showed superior performance in forecasting stock indices due to its ability to capture complex, nonlinear relationships effectively.
Conclusion: This study contributes to identifying the most influential variables in stock index fluctuations and demonstrates that more complex models, such as Random Forest and ANN, can be more effective in providing accurate predictions due to their higher sensitivity to input data. These findings also highlight the value of sensitivity analysis in identifying significant variables and eliminating unnecessary features, which helps improve the speed and reduce the complexity of models. Ultimately, the results can assist investors and policymakers in making more informed decisions and refining prediction models for better capital market strategies.
کلیدواژه‌ها [English]
)