Application of Deep Learning Algorithms in the Optimization of Banking Asset–Liability Portfolios: A Case Study of Bank Melli Iran

Abstract

The present study was conducted to investigate the application of deep learning algorithms in the optimization of banking asset–liability portfolios, with the case study focusing on real-world data from Bank Melli Iran. In a context where the national banking system is confronted with challenges such as liquidity fluctuations, high inflation, and credit risk, the design of predictive and decision-support models can play a significant role in improving the efficiency of asset–liability management (ALM). In this regard, the study proposes a bidirectional artificial intelligence-based model in which, at the first stage, recurrent neural networks based on Long Short-Term Memory (LSTM) architectures are employed to forecast temporal patterns of deposits and loans, while in the second stage, the Proximal Policy Optimization (PPO) reinforcement learning algorithm is utilized to determine the optimal portfolio composition within a dynamic and uncertain environment. Empirical findings derived from data collected from selected bank branches indicate that the proposed hybrid model demonstrates more accurate and stable performance compared to conventional methods such as the ARIMA model and the Markowitz mean–variance framework. Specifically, the prediction error (RMSE) was reduced by up to 55%, while capital adequacy ratios were maintained under critical scenarios. The findings suggest that deep learning and reinforcement learning techniques can be effectively utilized in the financial decision-making processes of banks.