A neuro-structural framework for bankruptcy prediction

We develop a framework to simultaneously compute the unobservable parameters underlying structural-parametric models for bankruptcy prediction. More specifically, we such as, asset value and volatility, through learning by embedding in structural neural network that maps network’s input space (e.g. companies’ observable financial market data) parameter space. Within ‘neuro-structural’ framework, model work together as one unit during phase providing each other forward backward information, respectively, until weights of are optimized according merit function. Empirical results show models, like Black-Scholes-Merton Down-and-Out option with computed our approach, perform better than alternative specifications out sample, terms discriminatory power, information content economic impact. Importantly, they also standard network, suggesting co-joint dynamics between useful can improve prediction performance (and training efficiency) networks. Finally, approach provides methodological empirical) enhancements over logit Campbell et al. [In search distress risk. J Finance, 2008, 63, 2899–2939]. There, data inputs, output is probability whereas includes an intermediary step obtain subsequently bankruptcy.