Predicting bankruptcy in Czech Republic: The role of data transformation

conference paper

English

The traditional bankruptcy models and their predictors cannot be used to predict bankruptcy in the Czech Republic as they have been intended for different business environments reflecting their specific features. Moreover, there are studies (Grice, Dugan, 2001; Wu, Gaunt, Gray, 2010; Niemann et al. 2008) showing that the precision of a bankruptcy model is significantly degraded if used in a field, period, and/or business environment different from that in which the learning data were observed. Building a new model is associated with pitfalls resulting from the character of bankruptcy data, e. g. the non-normality. In general, the fulfilment of method assumptions (e.g. normality) is one of the factors determining the quality of the rating model (Niemann et al, 2008). The discriminant analysis, as the most frequently used classification method used for bankruptcy prediction purposes, is based on the assumption of normality (Aziz, Dar, 2006). In praxis, financial data in the form of financial ratios are very often not normally distributed. This deviation from normal distribution may be caused by the lack of proportionality between the numerator and denominator of financial ratios (Barnes, 1982, 1987; Nikkinen, Sahlstrőm, 2004). In the case studied, the proportionality means, that: “the relationship between the two variables is linear and the constant is zero“(Whittington, 1980). Nikkinen and Sahlstrőm (2004) investigated the usefulness of Box-Cox transformation (Box, Cox, 1964) in normalizing the financial ratios in different kinds of accounting environments. They found, that Box-Cox transformation would lead to a substantial reduce of skewness, but only to partial reduce of kurtosis. The purpose of this paper is to analyse the possible deviation from normal distribution of the bankruptcy data. The research sample consists of 207 Czech industrial companies, the period of interest is a time series 2007-2010. Each analysed company is represented by 44 financial ratios. These financial ratios were used in previous bankruptcy prediction studies published from 1966 to 2010. A Shapiro-Wilks procedure was used to test normality (Shapiro, Wilks, 1965). In the event that non-normality is proved, the Box-Cox transformation will be used (Box, Cox, 1964) for achieving approximately normal distribution of the data. By the mean of skewness and kurtosis reduction, the use of Box-Cox transformation, lead to an expected result. The average kurtosis was reduced more than the average skewness. The effectiveness of Box-Cox transformation measured by the Shapiro-Wilks test was as follows. Before the transformation none of the analysed financial ratios met the condition of one-dimensional normality, not even on the 1% level. After the transformation, the condition of one-dimensional normality was met, at the 1% level, by 34% of the analysed financial ratios. The same condition, but at the 5 or 10% level was met by 27% of the analysed financial ratios. The condition of normality, for untransformed Czech bankruptcy data, seams nearly as impossible to fulfil. This conclusion implies the use of nonparametric methods, such as artificial neural networks. However, the comparison of the parametric methods performance using the untransformed or transformed data is a subject of further research.

Bankruptcy, normality, financial ratios, Box-Cox transformation, Shapiro-Wilks test

KARAS, M.; REŽŇÁKOVÁ, M.

2012

20. 9. 2012

Brno

978-80-214-4581-9

International Conference "Trends in Economics and Management for the 21st Century"

1

9

8