پیش بینی بحران بدهی های مستقل با استفاده از شبکه های عصبی مصنوعی: یک رویکرد مقایسه ای

Recent episodes of financial crisis have revived interest in developing models able to signal their occurrence in timely manner. The literature has developed both parametric and non-parametric models, the so-called Early Warning Systems, to predict these crises. Using data related to sovereign debt crises which occurred in developing countries from 1980 to 2004, this paper shows that further progress can be achieved by applying a less developed non-parametric method based on artificial neural networks (ANN). Thanks to the high flexibility of neural networks and their ability to approximate non-linear relationship, an ANN-based early warning system can, under certain conditions, outperform more consolidated methods.

The aim of this paper is to further develop the Early Warning System (EWS) literature on financial crisis. In particular, it presents a less explored non-parametric method, i.e. the artificial neural network (ANN), and tests its capacity to predict crises imminence. The paper shows, with an empirical application to sovereign debt default in developing countries, that a well-developed ANN can outperform both parametric and non-parametric traditional methods in emitting timely signals of crisis episodes. Financial crises that occurred in emerging countries in the 1990s have revived theoretical and empirical interest in understanding their causes and consequences, as well as in developing statistical and econometric models able to signal their occurrence in timely manner.1 According to Krugman, 1999 and Krugman, 2001 and Kaminsky (2003), economic theory has developed three generations of models explaining financial crises: the “first” and “second generation” models focus on currency crises, while the “third generation” ones cover a wider variety of crises and are better able to explain episodes that occurred in the late 1990s. In the “first generation” models, poor economic policies conflict with the goal of a fixed exchange rate and produce a continuous loss in foreign exchange reserves. Once reserves have fallen below a critical level, the authorities are forced to abandon the exchange rate peg. The building blocks of “second generation” models are the existence of multiple exchange rate equilibria and self-fulfilling speculative attacks. Even in the presence of sound economic policy, a government may consider the costs of maintaining a fixed exchange rate to be excessively high once the currency is subject to speculative attack. If investors doubt the authorities’ commitment to maintaining the peg and start to sell the currency, the government is induced to abandon the peg, which would otherwise have been sustainable. In this sense, speculative attacks are self-fulfilling. “Third generation” models were developed after the Asian crises, and they shifted the focus from public to private imbalances because public finances in those countries were quite sound, whilst those of the corporate and banking sectors displayed excesses. The literature analyses not only currency crises but also bank and “twin crises” (currency and banking crises), balance of payment crises, and sovereign debt crises. The theoretical underpinnings of third generation models vary: the moral hazard problem due to an implicit government guarantee which, together with poor regulation, induces over-landing and over-investment; the balance sheet effect due to a mismatch between assets and liabilities; a self-fulfilling liquidity run when (government, bank or corporate) debt has short-term maturity; a sudden stop to capital inflow due to external shocks. In all these cases the currency crisis is “more a symptom than a fundamental aspect of these crises”,2 and government, bank, corporate and currency crises are often related each other. In the past decade, many empirical studies have sought to develop models able to emit timely signals of the occurrence of a financial crisis, the so-called Early Warning Systems (EWSs). Using statistical and econometric techniques, these models are applied to predict the likelihood of financial crises, using for the purpose a large number of indicators related to internal and external factors, as well as social and political conditions. According to the type of approach adopted, models can be classified between parametric and non-parametric.3Frankel and Rose (1996) and Kaminsky et al. (1998) (KLR) are the seminal papers in the two classes of approaches applied to currency crisis prediction. In parallel with the previous literature on currency crises, the empirical literature on debt crises, which is comparatively rather small, can also be classified between parametric and non-parametric. Detragiache and Spilimbergo (2001) and Ciarlone and Trebeschi (2005) have used parametric models,4 while Manasse et al. (2003) have used both parametric and non-parametric approaches to develop an Early Warning System for debt crises.5 In the latter paper the authors show that a combination of the two approaches improves the performance of the logit in predicting the entry into a crisis. The current research tries to develop further the application of non-parametric methods in predicting sovereign debt crises in developing countries. In particular, it compares the performances of artificial neural networks with those of a traditional parametric method: the random effect probit estimator (REP). In economics, neural networks have been principally used in two classes of applications: classification of economic agents and time series prediction.6 In regard to classification, which is the subject of this paper, ANNs are widely employed for bankruptcy prediction,7 while very few applications focus on financial crises. Nag and Mitra (1999) use a dynamic ANN, one for each country, to test its performance in predicting Malaysian, Thai, and Indonesian currency crises, and compare the results with those of the signal approach. They find that the ANN model performs better than the KLR model, in particular when comparing out-of-sample predictions. Using data from East Asian countries, Franck and Schmied (2003) show that a multi-layer perceptron8 outperforms logit in predicting currency crises and in particular is able to signal the currency crises that hit Russia and Brazil in the late 1990s. This paper develops a multi-layer perceptron for debt crises prediction and compares its performance with those of a probit regression. Hence its approach is similar to that of the second of the last two cited papers. It differs in its data analysis because Franck and Schmied (2003) develop an ANN for each country, while in what follows a single ANN is used to analyse the entire sample. Furthermore, it differs in the regularization technique employed: Franck and Schmied (2003) adopt the Early Stopping procedure, while here a simple recursive algorithm is developed in order to improve the ANN's performance. The universal approximation theorem guarantees, under particular conditions, that it is possible to approximate, with arbitrary accuracy, the function that relates default probability and explanatory variables, so that a well-developed ANN should outperform other statistical techniques. The organization of the paper is the following: Section 2 describes the dataset. Preliminary analyses and benchmarks selection (the random effect probit estimator) are developed in Section 3. Section 4 contains a brief description of ANNs’ functioning, the performance of various models and a comparison in performances. Section 5 concludes.

This paper has shown that, thanks to the universal approximation theorem, an ANN with only two layers can outperform a traditional Early Warning System in predicting a sovereign debt crisis, if one chooses the right number of hidden units, training epochs, and an efficient training algorithm. This is possible because, if the relation that links the debt crisis indicator and the explanatory variables is highly non-linear, the flexibility of ANNs should yield, at least ideally, results that are at least as good as those of the parametric traditional methods. Further refinements can be made to improve the ANNs’ performance, for example using network committees. The characteristic of a committee is that it performs at least as well as the best single network. The drawback of ANNs is that they do not offer immediate intuitions in regard to policy implications. In fact, one cannot straightforwardly interpret, for example, the marginal effect, in terms of change in crisis probability, of an increase in an independent variable. This does not mean that it is always impossible, but the high non-linearity does not guarantee an explicit solution. On the other hand, ANNs do not claim to be policy models, but only forecasting models. Neither public nor private institutions entrust their predictions to a single automatic procedure. For example, the IMF uses a dual-core Early Warning System which comprises both a parametric (probit) and a non-parametric (signal) model.33 The flexible functional form of ANNs is able to improve the performance of an Early Warning System because of its strong data processing ability. But, at the same time, it should be used together with traditional methods and, last but not least, together with a “real” human brain.