C45 - Neural Networks and Related TopicsReturn

Bankruptcy prediction has long been regarded as a critical topic within the academic and banking community. To the best of our knowledge, no previous study in the Slovak Republic has attempted to develop a bankruptcy prediction model putting together statistical and artificial intelligence approaches performed on a such an amount of data. This paper seeks to fill this gap. Our aim is to develop a hybrid bankruptcy prediction model using a genetic algorithm in the process of training a neural network (GA-NN). The research data set comprises a balanced sample of both healthy and bankrupt firms operating in Slovakia in the period from 2014 to 2017. Financial information regarding a firm's financial situation are acquired from the Finstat database, which stores annual reports. For the purpose of comparing the classification accuracy of the proposed GA-NN model, two more models are constructed, namely BP-NN (back-propagation neural network model) as well as MDA (multiple discrimination model). The results gained by utilizing these models suggest the superiority of the developed GA-NN model to both BP-NN and MDA models in terms of prediction performance.

The paper is concerned with the use of several methods that can be useful from the point of view of trend reversal in financial time series. These methods are demonstrated on PX index time series during 2002-2009. The research itself is subdivided into four parts corresponding to individual analytical methods used. The first group contains the use of moving EGARCH(1,1) model to daily relative returns of PX index. The results obtained are indicative of the importance of negative parameter values, which can be considered as precursors of the trend reversal. The second group contains different moving characteristics that are able to signalize regime changes in certain time intervals. Particularly, the information related to intraday price variations proved to be useful. Third, selected price indicators from technical analysis were employed. Among them, Simple Moving Averages, Bollinger Bands, Relative Strength Index and Stochastic led to acceptable predictions. Last, the predictive ability of Artificial Neural Networks was tested with respect to different network structure and number of delayed values of explanatory variable. The results obtained here are promising, but further research in this direction is necessary.