Neural network designed to estimate probability of bank bankruptcies

Abstract

      The object of research is the banking system of Russia. The study purpose is to build a mathematical model to estimate probability of bank bankruptcies due to license revocation.  An instrument to build the model is neural networks to be trained on financial statements of the Central Bank of the Russian Federation. The testing error of the trained and optimized neural network has constituted 6.3%.  The studies of the modeled area – the banking system of the Russian Federation – have been carried out through virtual computer experiments.  The neural network calculations have been made by changing one of fifteen bank-related input parameters with other parameters remaining constant.
      In particular, the impact of long-term liquidity ratio, the type of business legal status, the exposure to large credit risks and bank place of registration on bank bankruptcy probability has been investigated. As a result the conclusion has been formulated that the increase of long-term liquidity ratio reduces the bank bankruptcy probability. However, starting with a certain level, depending on other parameters of a specific bank, the increase of this indicator increases the probability of its bankruptcy. Essential impact on successful bank performance is exerted by bank’s business legal status, as well as the place of its registration. However, this impact is ambiguous and may manifest itself differently in each individual case, depending on many other bank parameters and its operations. A case study involving the mathematical model application to formulate recommendations to reduce bankruptcy probability of a bank is given.