Using convolutional neural networks to model the relationship between institutional investors' horizons and bank risk-taking

Introduction: Bank risk-taking is a fundamental component of financial system stability and is shaped by a variety of factors, including ownership structure, shareholder composition, and investor behavior. Among these, the investment horizon of institutional investors—whether short-term or long-term—plays a critical role in determining how banks respond to different types of market risk. Institutional investors with varying horizons adopt distinct approaches to oversight, resource allocation, and responsiveness to market fluctuations, which in turn can have direct implications for banks’ risk exposure. Due to the multidimensional and nonlinear nature of the relationship between investment horizon and risk-taking, traditional financial modeling techniques such as linear regression or feedforward neural networks face significant limitations. These models often fail to capture hidden patterns, nonlinear dependencies, and localized features within financial data. Therefore, the primary objective of this study is to utilize Convolutional Neural Networks (CNNs) to more accurately model the relationship between institutional investment horizon and bank risk-taking, using structured financial data from banks listed on the Tehran Stock Exchange.

Method: This research is applied in nature and follows a descriptive-correlational methodology. The statistical population consists of all active banks listed on the Tehran Stock Exchange between 2014 and 2020. Financial data related to institutional investment horizons (short-term and long-term) and key indicators of bank risk-taking were extracted and transformed into 4×4 matrices to enable convolutional processing. Three CNN architectures were designed and trained, each with a different number of filters: 64, 128, and 256. These models were developed to identify complex patterns and improve predictive accuracy in modeling the target relationship. Model performance was evaluated using standard classification metrics, including Accuracy, F1 Score, and Area Under the Receiver Operating Characteristic Curve (AUC).

Results and Discussion: The results revealed that the correlation between short-term institutional investment horizon and bank risk-taking was 0.86, significantly stronger than the correlation observed for long-term horizons, which was 0.72. This indicates that institutional investors with shorter horizons, due to their heightened responsiveness to market dynamics, play a more effective role in mitigating bank risk. Among the three CNN models, the 64-filter architecture (CNN64) demonstrated the highest predictive performance, achieving an accuracy of 0.829, an F1 score of 0.815, and an AUC of 0.842. This model successfully extracted localized features and identified complex patterns within the financial data, outperforming traditional approaches in both precision and generalizability.

Conclusion: The findings of this study confirm that Convolutional Neural Networks—particularly lightweight configurations such as CNN64—are powerful tools for analyzing complex and nonlinear relationships in financial datasets. Applying CNNs to assess the impact of institutional investment horizons on bank risk-taking not only enhances predictive accuracy but also enables deeper interpretation of market behavior and more informed investment decision-making. The emphasis on short-term horizons as a key factor in reducing bank risk offers practical implications for regulatory policy, risk governance, and strategic asset allocation in the banking sector. Future research is encouraged to explore a broader range of machine learning models, such as decision trees, ensemble methods, and deeper neural networks, and to incorporate macroeconomic variables including interest rates, inflation, and monetary policy into the analytical framework. Moreover, extending the scope of analysis to include international financial markets and longer time horizons will enhance the external validity of the findings and contribute to theoretical advancement in bank risk management.