After identifying the two sets of leading indicators for A and BEWS-B, we can proceed to build the early warning systems accordingly for each country. Take ARGENTINA for example, By weighing with the Spearman correlation coefficients, we create an index variable for BEWS-A with the leading indicators extracted in factor analysis. As Figure 5.11 shows, we cre-ate an index variable to compare between the posterior probability and the forward crisis variable. Give an probability threshold of 0.2 for the index vari-able, as long as the posterior probability of the idex variable is greater than 0.2, then the system BEWS-A will send an alarm signal that financial crises may occur within four months in the future. On the other hand, BEWS-B will be built with the other set of leading indicators in the process of dis-criminant analysis. Except of BRAZIL, CHINA, POLAND and SLOVENIA
Figure 5.11: Given an probability threshold of 20% for the index variable to identify crisis signals for BEWS-A and compare with the forward crisis variable
which didn’t suffer any crises during 1994 to 2003, all the goodness-of-fit anal-yses of 24 countries for BEWS-A and BEWS-B are summarized in Table 5.7, where St= 0 represents that the system didn’t send an alarm signal at time t, St = 1 represents that the system sent an alarm signal at time t, Yt = 1 represents that a crisis truly occurred within the next four months at time t, and Yt = 0 represents that no crisis occurred within the next four months at time t. We compare the the two systems in terms of the five goodness-of-fit criteria, that are probability of observations correctly called, probability of
Table 5.7: Simulation results for BEWS-A and BEWS-B
Prob. of obs. correctly called: 95.9 Prob. of obs. correctly called: 95.7 Prob. of crises correctly called: 83.6 Prob. of crises correctly called: 76.7 Prob. of false alarms: 18.1 Prob. of false alarms: 15.2 Prob. of crisis given an alarm: 81.9 Prob. of crisis given an alarm: 84.8 Prob. of crisis given no alarm: 2.2 Prob. of crisis given no alarm: 3.0 crises correctly called, probability of false alarms of total alarms, probabil-ity of crisis given an alarm, and probabilprobabil-ity of crisis given no alarm. Table 5.8 to 5.9 provide a comparison of the goodness-od-fit of our systems with those models of previous studies on the subject. Our systems perform better in each of the five goodness-of-fit criteria than the Pooled logit BF model (on the left of Table 5.8), the IMF-DCSD model (on the right of Table 5.8), the Kaminsky-Lizondo-Reinhart model (on the left of Table 5.9) and the Goldman-Sachs model (on the right of Table 5.9).
The system BEWS-A correctly calls the highest ratio of observations (95.9%) and of crises months (83.6%), and the system BEWS-B gives the fewest false alarms of any of the model above-mentioned. Besides, the con-ditional probabilities of having a crisis if an alarm occurred are more than 80% for both BEWS-A and BEWS-B, which are much higher than any of the other models. The better performance of our systems means that the leading indicators we have been using are more reliable, and that the country sample and time period are more appropriate.
Table 5.8: Pooled logit BF model and IMF-DCSD model
Prob. of obs. correctly called: 84.1 Prob. of obs. correctly called: 76.7 Prob. of crises correctly called: 66.7 Prob. of crises correctly called: 65.1 Prob. of false alarms: 50.0 Prob. of false alarms: 62.8 Prob. of crisis given an alarm: 50.0 Prob. of crisis given an alarm: 37.2 Prob. of crisis given no alarm: 6.7 Prob. of crisis given no alarm: 7.8
Table 5.9: IMF-KLR model and Goldman-Sachs model St= 0 St = 1 Total
Prob. of obs. correctly called: 70.2 Prob. of obs. correctly called: 66.1 Prob. of crises correctly called: 59.8 Prob. of crises correctly called: 66.2 Prob. of false alarms: 70.3 Prob. of false alarms: 74.0 Prob. of crisis given an alarm: 29.7 Prob. of crisis given an alarm: 26.0 Prob. of crisis given no alarm: 9.8 Prob. of crisis given no alarm: 8.4
Chapter 6 Conclusions
In this thesis, two new Early Warning Systems (EWS) for predicting finan-cial crises had been developed. The main difference to existing EWS models and the intended contribution of this thesis focused on three areas. First, it proposed a systematic framework by combining Bayesian Theorem with four kinds of feature selection methods, including F statistics, Spearman correla-tion, factor analysis, and discriminant analysis, to identify two different sets of leading indicators. The second area is that our EWS offered each of the leading indicators a weighting to create an index variable. By monitoring the posterior probability of the index variable, the early warning signals will be determined whether to send out or not. And third, we found out different leading indicators for different countries in terms of their own characteristics.
From a policy perspective, developing EWS that help to reliably antici-pate financial crises could become a more important tool for policy-makers in the future. Many financial crises over the past few decades had caused dam-age to social security, economics and development of industries. Developing reliable EWS there can be of substantial value by allowing the policy-makers to obtain clear signals when and how to take pre-emptive measures in
or-der to mitigate or even prevent financial turmoil. It should be stressed that EWS can not replace the sound judgment of the policy-maker to guide pol-icy, but it can play an important role as a neutral and objective measure of vulnerability.
This thesis showed that detecting regime shifts of financial variables and estimating the degree of similarity between the obtained posterior probability and a forward crisis variable can indeed identify more reliable leading indica-tors and improve the performance of EWS substantially. Besides, applying factor analysis and discriminant analysis to distinguish different classes of variables avoids including the leading indicators with the same characteristic in a system. It would be beneficial to keep the predictive stability for a sys-tem. From the simulation results of section 5.4, it is fair to say that both of the systems developed in this thesis performed much better than any EWS models of previous studies on the subject in terms of predictive power. How-ever, because the collected data of financial variables and crisis variable in this thesis only have 120 samples, it is insufficient for making cross validation in the process of parameter optimization. It may cause an issue of overfitting for the simulation results. In addition, because we don’t have enough data to regard as testing samples, we only can provide simulation results, instead of experiment results.
In conclusion, the simulation results accords with our expectancy. But it should be emphasized that the EWS developed in this thesis doesn’t consti-tute the final step towards a comprehensive EWS of financial crises. Future research on EWS may focus on the following points:
1) Considering an overlap between pre and post windows or the magnitude of the regime shifts is helpful to eliminate false detection. Besides,
a further method is to combine the posterior probability calculated through different methods.
2) Regime shifts could be classified into three kinds of types, including the smooth regime shifts, the abrupt regime shifts and the discontinuous regime shifts [26]. It is possible to use different sets of window defini-tion for classificadefini-tion of different types of regime shifts so as to reduce classification delay.
3) Considering that a post-crisis bias will sometimes appear in an EWS.
If we fail to distinguish between pre-crisis and post-crisis periods, it may bring about a bias in the estimation results. So, suggesting a new method to solve this problem could be a further step towards developing an EWS that is more powerful in terms of prediction.
4) Developing a framework that allows the policy-makers to design the features of their EWS according to their preferences and degree of risk-aversion. It includes adding dynamic components to EWS, or providing choices of the timing and the length of different regimes.
References
[1] M. Bussiere and M. Fratzscher, Towards a New Early Warning Systen of Financial Crises. European Central Bank, Working Paper Series NO.
145, (May 2002).
[2] G. L. Kaminsky and C. M. Reinhart, 1998, Financial Crises in Asia and Latin America: Then and Now. American Economic Review, Papers and Proceedings; 88(2), pp. 444-48, (May 1998).
[3] A. Berg and C. Pattillo, 1999b, Predicting Currency Crises: The Indica-tors Approach and an Alternative, Journal of International Money and Finance. 18(4), pp. 561-86, (August 1999).
[4] G. L. Kaminsky, S. Lizondo and C. M. Reinhart, Leading Indicators of Currency Crises. IMF Staff Papers, 45, NO. 1, pp. 1-48 (1998).
[5] G. S. Maddala, Limited-dependent and Qualitative Variables in Econo-metrics. Econometric Society Monographs, Cambridge University Press, (1989).
[6] J. F. Ducre-Robitaille, L. A. Vincent and G. Boulet, Comparison of Techniques for Detection of Discontinuities in Temperature Series. In-ternational Journal of Climatology, 1087-1101, (2003).
[7] E. S. Epstein, Detecting Climate Change. J. Appl. Meteorol. 21, 1172, (1982).
[8] T. R. Karl and C. N. Williams, An Approach to Adjusting Climatological Time Series for Discontinuous Inhomogeneities: Journal of Climate and Applied Meteorology. 26, 1744-1763, (1987).
[9] N. J. Mantua, Methods for Detecting Regime Shifts in Large Marine Ecosystems: a Review with Approaches Applied to North Pacific Data.
Progress in Oceanography, 60, 165-182, (2004).
[10] G. L. Kaminsky and C. M. Reinhart, The Twin Crises: The Causes of Banking and Balance-of-payments Problems. International Finance Dis-cussion Papers 544, Board of Governors of the Federal Reserve System (U.S.), (1996).
[11] S. N. Rodionov, A Brief Overview of the Regime Shift Detection Meth-ods. Joint Institute for the Study of the Atmosphere and Ocean, Uni-versity of Washington, Seattle, WA 98195, (2005).
[12] J. Wang, and Q. P. He, A Pattern Matching Approach for Fast Dis-turbance Detection and Classification Using Bayesian Statistics. AMD (Sep. 2005).
[13] H. H. Yue and M. Tomoyasu, Weighted Principal Component Analysis and its Applications to Improve FDC Performance. Tokyo Electron AT Ltd, Nirasaki, Yamanashi, Japan, (2004).
[14] H. Liu and H. Motoda, Feature Selection for Knowledge Discovery and Data Mining. Kluwer Academic Publishers, London, (1998).
[15] Janet, E. A. Forrester, Ph. D. Fall, CMH110 Epidemiology and Bio-statistics. (2004/2005).
[16] K. A. Brownlee. Statistical Theory and Methodology in Science and En-gineering. JohnWiley, New York. (1960).
[17] J. F. Hair, R. E. Anderson, R. L. Tatham, and W. C. Black, Multivariate Data Analysis (3rd ed.). New York, Macmillan, (1992).
[18] J. Kim and C. W. Mueller, Introduction to Factor Analysis: What It Is and How to Do It. Newbury Park, CA: Sage Publications (1978).
[19] Bryant and Yarnold, Principal Components Analysis and Exploratory and Confirmatory Factor Analysis. In Grimm and Yarnold, Reading and understanding Multivariate Analysis. American Psychological As-sociation Books (1995).
[20] T. Hastie, R. Tibshirani and J. Friedman, The Elements of Statistical Learning. Data Mining, Inference, and Prediction. (2001).
[21] J. O. Kim and C. W. Mueller, Factor Analysis: Statistical Methods and Practical Issues. Thousand Oaks, CA: Sage Publications, (1978).
[22] L. L. Thurtstone, Multiple-Factor Analysis. Chicago: University of Chicago Press, (1947).
[23] G. H. Dunteman, Principal Components Analysis. Thousand Oaks, CA:
Sage Publications, (1989).
[24] C. E. Lance, M. M. Butts, and L. C. Michels, The sources of four com-monly reported cutoff criteria: What did they really say? Organizational
Research Methods. 9(2): 202-220. Discusses Kaiser and other criteria for selecting number of factors, (2006).
[25] P. A. Lachenbruch, Discriminant analysis. NY: Hafner, (1975).
[26] Collection of Definitions of Regime Shift, http://biocore.ecolab.
bas.bg/events/past/unesco-ws/ws-workgr/regshift.htm
Appendix A
A Summary of Collected Data
There are various types of financial crises: currency crises, banking crises, sovereign debt crises, private sector debt crises, and equity market crises.
Currency crises are believed to be the most important type among them because they often coincide or occur in succession with other types of financial crises [10]. For this reason, our EWS model in this thesis focuses mainly on currency crises as well as most EWS models in the literature. We proceed to define a currency crisis by employing a variable named exchange market pressure (EMPi,t). The EMPi,t variable for each country i and period t is defined as
EMPi,t is a weighted average of the change of the real effective exchange rate (RER), the change in the interest rate (r), and the change in foreign exchange reserves (res). ωRER, ωr, and ωres are the relative precision of each variable for all countries over the full sample period 1994-2003.
The rationale for using EMPi,t is expounded as follows. If investors con-sider there are some underlying vulnerable economic factors attacking a cur-rency of a specific country, the government essentially has two options to deal with this kind of situation. One is to abstain from defending the currency either by abandoning a fixed exchange rate regime or by avoiding intervening in foreign exchange markets, and to let the currency devalue. The other is to defend the currency regime by raising interest rates and running down foreign exchange reserves. We can give consideration to both sides with EMPi,t.
The next step is to define a currency crisis (CCi,t) which is given by
CCi,t=
1 if EMPi,t> EMPi+ 2SD(EMPi), 0 if otherwise,
(A.2)
as the event when EMPi,t is two standard deviations (SD) or more above its country average EMPi [1]. Besides, the thesis uses monthly data over period from 1994 to 2003 for a sample of 28 countries and each country has 48 different financial variables. These 48 variables concerned with external competitiveness, external exposure, domestic real and public sector, domestic financial sector, global factors and the contagion through trade channel, are expected to be able to express the situation that crises will occur in the future. And we will take them as the training data in our model.