In this paper, we compare the performance of two non-parametric methods of classification and regression trees (CART) and the newly multivariate adaptive regression splines (MARS) models, in forecasting bankruptcy. Models are tested on a large universe of US banks over a complete market cycle and run under a K-fold cross validation. Then, a hybrid model which combines K-means clustering and MARS is tested as well. Our findings highlight that (i) Either in training or testing sample, MARS provides, in average, better correct classification rate than CART model (ii) Hybrid approach significantly increases the classification accuracy rate in the training sample (iii) MARS prediction underperforms when the misclassification of the bankrupt banks rate is adopted as a criteria (iv) Finally, results prove that non-parametric models are more suitable for bank failure prediction than the corresponding Logit model.

Introduction

The financial crisis, which started in 2007, has dramatically affected banks sector throughout the world. The shock wave epicenter was in the US and it took long time for regulators to stop the default chain and save big banks. Therefore, the prevention against the systemic risk – failure of the banking system- becomes an ineluctable concern and the need of new forecasting tools is of major importance to not only regulators but also academics. In this sense, federal banking supervisors [the Federal Reserve, the Federal Deposit Insurance Corporation (FDIC), and the Office of the Comptroller of the Currency (OCC)] and other financial supervisory agencies provide a supervisory rating (convenient summary of bank conditions at the time of an exam). This helps investors to detect banks that have a great default probability ratio. A key outcome of such an on-site exam is a CAMELS rating. The acronym “CAMEL” refers to the five components of a bank’s condition that are evaluated: Capital adequacy, Asset quality, Management, Earnings, and Liquidity. A sixth component, a bank’s Sensitivity to market risk, was added in 1997; hence the acronym was changed to CAMELS. The FDIC developed also a Statistical CAMELS Off-site Rating system (SCOR) to perform the bank’s stability evaluation. Collier et al. (2003) examine the performance of this model over the period 1986–۲۰۰۲ and point out the limitations of this model despite the usefulness of SCOR which is based only on financial ratios. Cole and Gunther (1995) prove the same results and report that the CAMELS ratings decay rapidly. Predicting bank bankruptcy has reached a specific interest in financial literature. Thus, numerous models have been developed since the early 70s. All proposed models are based on classification methods in a multidimensional space defined by a set of specific variables. The literature is rich of non-parametric and parametric models. With regard to the later, Beaver (1966) was one of the first researchers who focused on univariate analysis to study bankruptcy prediction. He tested the power of financial ratios to classify and predict bankrupt firms. Cash flow and debt ratios appeared to be the important predictors of bankruptcy. Altman (1968) used Multivariate Discriminant Analysis (MDA) to develop a five-factor model to calculate the well-known “Z-score” and predict bankruptcy of manufacturing firms. As an examples of well-known statistical methods to predict failure, the logistic regression (logit) (Ohlson 1980; Demirgüç-Kunt and Detragiache 1997; Affes and Hentati-Kaffel 2017), Probit analysis (Zmijewski 1984; Hanweck et al. 1977) and factor analysis. West et al. (1985) demonstrated that that the combination of factor analysis and Logit estimation was promising in evaluating banks’ condition. The factors identified by the Logit model as important descriptive variables for the banks’ operations are similar to those used for CAMELS ratings.