Hu, Yu-Chiang

2018-03-29T12:17:19Z

2018-03-29T12:17:19Z

2007

The current research develops a theoretical framework based on the ResourceAdvantage Theory of Competition (Hunt, 2000) for the selection of appropriate variables. Using a review of the literature as well as to interviews and a survey, 170 potential retail performance variables were identified as possible for inclusion in the model. To produce a relative simple model with the aim of avoiding over-fitting, a limited number of key variables or principal components were selected to predict default. Five credit-scoring techniques: Naive Bayes, Logistic Regression, Recursive Partitioning, Artificial Neural Network, and Sequential Minimal Optimization (SMO) were employed on a sample of 195 healthy and 51 distressed businesses from the USA market over five time periods: 1994-1998, 1995-1999, 1996-2000, 1997-2001 and 1998-2002.

Analyses provide sufficient evidence that the five credit scoring methodologies have sound classification ability in the year before financial distress. Moreover, they still remained sound even five years prior to financial distress. However, it is difficult to conclude which modelling technique has the highest classification ability uniformly, since model performance varied in terms of different time scales. The analysis also showed that external environment influences do impact on default assessment for all five credit-scoring techniques, but these influences are weak. These findings indicate that the developed models are theoretically sound. There is however a need to compare their performance to other approaches.

To explore the issue of the model's performance two approaches are taken. First, rankings from the study were compared with those from a standard rating system—in this case the well-established Moody's Credit Rating. It is assumed that the higher the degree of similarity between the two sets of rankings, the greater the credibility of the prediction model. The results indicated that the logistic regression model and the SMO model were most comparable with Moody's. Secondly, the model's performance was assessed by applying it to different geographical areas. The original USA model was therefore applied to a new US data set as well as the European and Japanese markets. Results indicated that all market models displayed similar discriminating ability one year prior to financial distress. However, the USA model performed relatively better than European and Japanese models five years before financial distress. This implied that a financial distress model has potentially better prediction ability when based on a single market.

Following this result it was decided to explore the performance of a generic global model, since model construction is time-consuming and costly. A composite model was constructed by combining data from USA, European and Japanese markets. This composite model had sound prediction performance, even up to five years before financial distress, as the accuracy rate was above 85.15% and AUROC value was above 0.7202. Comparing with the original USA model, the composite model has similar prediction performance in terms of the accuracy rate. However, the composite model presented a worse prediction utility based on the AUROC value. A future research direction might be to include more world retailing markets in order to ensure the model's prediction utility and practical applicability.

http://hdl.handle.net/1842/29166

The University of Edinburgh

Annexe Thesis Digitisation Project 2018 Block 17

Already catalogued

Developing retail performance measurement and financial distress prediction systems by using credit scoring techniques

Thesis or Dissertation

Doctoral

PhD Doctor of Philosophy