Assessing AI-Based Forecasting for Bank Valuation: The Banco CTT Case

The valuation of banks is complex due to several factors: special regulatory capital requirements, balance sheet dynamics where deposits serve as both liabilities and funding sources and interest rate risk exposures that render traditional cash flow metrics unreliable. Given the complexity, this study aims to answer whether machine learning can generate credible stand-alone bank valuations using public financial and macroeconomic data, while identifying influential drivers. A neural network was trained on 103 features from 415 European banks (1999–2024) to forecast P/B ratios, employing chronological splits to prevent look-ahead bias. The results show that the artificial neural network was unable to generate reliable valuations: test-set R² (–2.87), MSE (4.09) and MAE (0.53) exceeded 50% of the mean P/B ratio (≈1), performing worse than a naive mean predictor. However, integrated gradients identified economically relevant drivers: ROE increased P/B, while loan-loss provisions and surplus deposits decreased it, aligning with traditional valuation theory. Based on these findings, it can be concluded that while machine learning is unable to replace analyst-led valuation for novel cases, it is effective in quantifying universal drivers. The model’s inability to fully price instrument-specific aspects leads to the conclusion that human input is still required. This information is critical for analysts, financiers and regulators, as it highlights that the assessment of intrinsic value cannot yet be fully delegated to algorithms. In summary, although AI has the potential to accelerate valuation workflows, expert judgement is still necessary for accuracy in cases where firm-specific nuances matter.