Rapid growth and transformation of the Chinese economy and financial markets coupled with escalating default rates, rising corporate debt, and poor regulatory oversight motivates the need for more accurate distress prediction modelling in China. Given China’s historical, social, and cultural intolerance towards corporate failure, this thesis examines Chinese distress based on the Special Treatment (ST) system introduced by the CSRC in 1998. Regulators can assign Special Treatment status to listed Chinese companies for poor financial performance, financial abnormality and other events. This study employs an advanced machine learning technique – gradient boosting (TreeNet®) to examine the predictive and explanatory performance of more than 90 predictor variables, including financial ratios, market returns, macroeconomic indicators, shareholder ownership/concentration, executive compensation measures, corporate governance proxies, valuation multiples, audit quality factors, corporate social responsibility metrics, and other variables.
In addition to conventional dichotomous distress modelling, this thesis also models Chinese financial distress in a multi-state setting. Unlike binary distress prediction models that are subject to oversimplification of the underlying economic reality of firms, multi-state models can better approximate the continuum of corporate financial health observable across Chinese listed companies. Based on out-of-sample tests, the binary TreeNet® model is 93.74 percent accurate in predicting distress (a Type I error rate of 6.26 percent) and 94.81 percent accurate in predicting active/healthy companies (a Type II error rate of 5.19 percent). The three-state TreeNet®model is 96.82 percent accurate in predicting active or healthy companies; 76.49 percent accurate in predicting state 1 distress (ST=1); and 73.28 percent accurate in predicting state 2 distress (ST > 1). The five-state TreeNet® model is 94.47 percent accurate in predicting active or healthy companies; 61.70 percent accurate in predicting state 1 distress (ST=1); 53.12 percent accurate in predicting state 2 distress (1< ST <4); 62.56 percent accurate in predicting state 3 distress (ST ≥ 4); and 51.72 percent accurate in predicting state 4 distress (delisted).
From the analysis of the RVI metrics of the binary, three-state and five-state TreeNet® models, variables with strongest predictive value include: (i) market-price variables, particularly market capitalisation and annual market returns; (ii) executive compensation measures, such as total compensation of the top three executives and total compensation to the top three directors; (iii) macroeconomic variables, notably GDP growth, GDP per capita and unemployment rates; (iv) financial variables, particularly retained earnings to total assets, net profit margin, ROA, ROE, and market capitalisation to total debt; and (v) shareholder ownership/concentration, notably percentage of shares held by insiders. The empirical results suggest a wide range of financial and non-financial ‘Western-style’ bankruptcy predictors also provide good predictive and explanatory power in the Chinese context. In addition to conventional financial and market variables, some non-conventional variables such as executive compensation measures, macroeconomic variables and shareholder ownership/concentration variables are also found to be fairly predictive in the unique context of China’s ST system. The diverse range of top performing predictor variables also reflects several different dimensions of corporate financial distress in China. The empirical findings imply that future research on Chinese distress prediction modelling could combine variables measuring different dimensions of corporate financial health in Chinese companies.