In the existing literature, many definitions of currency crises have been adopted. Generally, exchange rate, interest rate and international reserve are the components in these definitions.  The authors tend to determine a crisis period in two ways. The first group bases on the combination of the components. Some notable examples are the definitions of Eichengreen, Rose and Wypslosz (1996) and Kaminsky, Lizondo and Reinhart (1998). The advantage of these indexes is that they consider not only successful speculative attacks (large devaluation of exchange rates) but also unsuccessful ones (loss of reserves and high increment of interest rates). This feature comes as a result of a weighted combination of exchange rates, interest rates and international reserves. The second group considers the components separately. For instance, the definitions of Frankel and Rose (1996) and Zhang (2001). The positive feature for applying this approach is that it fits well in the case when the impacts of three components cancel out each other. However, it considers only successful attacks and ignores unsuccessful ones. As a result, it is inconsistent when we aim to compare these forecasting models with different crisis definitions. Toward a unified framework for predicting currency crises and following the mainstream of applying simulation methods to determine the probability of rare events including the models of Anderson (2007), we make a new approach by using vector autoregressive model (VAR) for three variables that are exchange rates, interest rates and international reserves. By doing so, we can apply any existing definition to obtain a comprehensive crisis forecast given the forecasted results of the three components.

Most of the conventional models in forecasting currency crises apply discrete dependent variables coding the tranquil regime as zero and the crisis regime as unity. The crisis regime is defined at the time when exchange market pressure index signals a crisis some months (6, 12 or 24 months for example) prior for the purpose of prediction Bussiere (2006). This approach shows itself to be illogical because two different periods which are the time before a crisis and crisis time are tied together and given the same value of unity. Compounding the problem, the pre-crisis period is chosen arbitrarily which does not rely on any evidence from the data. In this study, we can avoid that pitfall by predicting three components in advance from 1-month lead to 6-month lead.

This study also promotes a method of simulation for the model that is likely to provide the probabilistic forecasting of crisis occurrence. By bootstrapping the data input, we can obtain distributions of exchange rates, interest rates, international reserves and as the result, the exchange market pressure index. The ability to forecast the probability of crisis in a non-parametric way given various forecasting horizons is a good starting point for the rest of the analysis.

It is worth noting that we enrich the EWS literature by using vector autoregressive models with three components which are exchange rates, interest rates, international reserves. These time series are available to be accessed at monthly frequency which means that the model does not need to convert from low-frequency data (quarterly and annually) to higher-frequency data.  This feature is quite appealing when we note that the conventional EWSs often use fundamentals as the main inputs. Clearly, such models lacks precision on two counts; they distort the data by adopting different methods to transform data to higher frequencies because the fundamental variables are usually available at either quarter and annual frequencies, and they do not cover a long research time period and include many countries in the sample because a large number of fundamental variables are just available recently.

In the same vein, using only three well-updated variables as the data inputs for the model helps us come close to the real-time forecasting which are impossible in the case of the EWSs with many dependent variables. In the real world, macroeconomic data often has lag time to be published officially. To put this another way, the conventional EWSs which rely too much on a wide range of data that can not provide an in-time prediction for crises. As a result, the governments are unlikely to be warned before the crises occur and can not give out defensive policy in a timely manner.

In this research, we have developed a new approach for the Early Warning System of currency crises by first predicting three components of a crisis definition and then establishing the crisis signal given these predicted components. The conventional Early Warning Systems use the discrete crisis variable as the dependent variable. We also predict each separate time point (month in our study) as crisis or non-crisis while the existing models bundle all time points in a certain period of time as a crisis. According to this approach, either the policy makers or investors can have predictions of crisis of every month rather than only a probability of crisis for a specific period. 

This study achieves important findings given the analysis on both in-sample and out-of-sample. Firstly, through 5 main performance measures, the ROC curves and AUC values, both model 1 and model 2 have powerful performance especially when applied to the KLR and KLR-revised crisis definitions. For the case of Zhang's crisis definition, although the relative amounts of false alarms is rather high, the other remaining measures such as the AUC values show good results. Secondly, the out-of-sample performance of our models is even substantially better than the in-sample performance which is a plus point when we aim at real-time forecasting.