Forecasting Financial Crises and Contagion in Asia using Dynamic Factor Analysis

observation (in sample) for the the vector of r estimated factors. In particular, the loading
of each factor into the unstandardised EMPi index is captured by the coefficients βi . In
order to account for serial dependence of the dependent variable, the factor projections are
augmented with past values of the dependent variable. The residual νj,t+1is the idiosyncratic
country specific shock for the j-th country.

In order to estimate the coefficients in equation (??) we need, first, to determine the num-
ber of factors r and the number of lags p for the dependent variable. Fixing the maximum
order for p and r, to four and eight, respectively, we use the Bayesian information criterion
(BIC), as suggested by Stock and Watson, (2002). We obtain the factor estimates and the
estimates for the coefficient matrices D and R, following the procedure described in section
3.1. The coefficient estimates for a0, αk, k = 1, ..., p, and βi are obtained by regressing (via
OLS) the unstandardised EMP index on an intercept, its lags, and on the estimated factors.

In order to produce density forecasts we use Monte Carlo stochastic simulation. Each
stochastic simulation replication is given by a combination of the replications of the common
(which is, then, interpreted as the regional vulnerability indicator) and idiosyncratic shocks,
ut and νj,t, respectively. In particular, the s-th replication (scenario) for E M Pi,t+1 , denoted
by E M P_i^s_,t+1 is given by

where Ri is the i-th row of R and u_t^s₊₁ and ν_j^s_,t+1 denote the s-th replication for ut+1 and
νj,t+1, respectively. Both shocks are obtained from draws from standardised Gaussian random
variables. The number of replications (hence the number of different scenarios) is 10000 and
this is considered as an exhaustive number of scenarios (therefore reducing the computational
intensity of the Monte Carlo experiment) if we fix q to unity.

3.3 Density forecasts from competing models

The density forecasts associated with various competitor models are given below:

1. The first model we consider is a naive predictor given by:

EMpna+V = μ_ι + σ_j νj,t+1

<<   6   7   8   9   10   11   12   >>

More intriguing information