Learning machines supporting bankruptcy prediction

Abstract

In many economic applications it is desirable to make future predictions about the financial status of a company. The focus of predictions is mainly if a company will default or not. A support vector machine (SVM) is one learning method which uses historical data to establish a classification rule called a score or an SVM. Companies with scores above zero belong to one group and the rest to another group.Estimation of the probability of default (PD) values can be calculated from the scores provided by an SVM. The transformation used in this paper is a combination of weighting ranks and of smoothing the results using the PAV algorithm. The conversion is then monotone.This discussion paper is based on the Creditreform database from 1997 to 2002. The indicator variables were converted to financial ratios; it transpired out that eight of the 25 were useful for the training of the SVM. The results showed that those ratios belong to activity, profitability, liquidity and leverage.Finally, we conclude that SVMs are capable of extracting the necessary information from financial balance sheets and then to predict the future solvency or insolvent of a company. Banks in particular will benefit from these results by allowing them to be more aware of their risk when lending money. * The financial support from the Deutsche Forschungsgemeinschaft via SFB 649 'Ökonomisches Risiko", Humboldt-Universität zu Berlin is gratefully acknowledged. †

Oracally Efficient Two-Step Estimation of Generalized Additive Model

“…In this study, we compute the GAM SBK score S = b {ĉ + 8 α=1m SBK,α (X α )}, b (x) = e x /(1 + e x ). For the RDC data, our analysis has the AR value 62.46%, better than the AR value 60.51% obtained in Härdle, Hoffmann, and Moro (2011). This is clearly due to the fact that our credit score function depends on each variate X α , 1 ≤ α ≤ d, non- parametrically via the GAM.…”

“…This is clearly due to the fact that our credit score function depends on each variate X α , 1 ≤ α ≤ d, non- parametrically via the GAM. The score function used in Härdle, Hoffmann, and Moro (2011), on the other hand, is a linear function of X α , 1 ≤ α ≤ d, thus lacking flexibility. Our AR value of 62.46% is also higher than the AR value 58.69% obtained using the GAM procedure in R. We can also estimate the functions m α (x α ) for X α .…”

“…The dataset contains d = 8 financial ratios shown in Table 1, such as Ebit/Total Assets and log( Total Assets), of 18,610 solvent (Y = 0) and 1000 insolvent (Y = 1) German companies, see Härdle, Hoffmann, and Moro (2011) for details.…”

Statistical inference for generalized additive partially linear models

“…Using GAM and SBK method, we clearly see via the SCCs that the shape of m 2 (x 2 ) is linear. Figure 3 For the RDC data, the in sample AR value obtained from GAPLM is 62.89%, which is very close to the AR value 63.05% obtained from GAM in [11] and higher than the AR value 60.51% obtained from SVM in [4]. To compare the prediction performance, we use the AR introduced in Example 1.…”

“…, where F nα is the empirical cdf for the data {X iα } n i=1 . See [4,11] for more details of this data set.…”