Abstract
This article employs second-generation random coefficient (RC) modeling to investigate the time-varying behavior and the predictability of the money demand function in Taiwan over the period from 1982Q1 to 2006Q4. The RC procedure deals with some of the limitations of previous studies, such as unknown functional forms, omitted variables, measurement errors, additive error terms, and the correlations between explanatory variables and their coefficients. Our main findings are as follows. First, the empirical results indicate that the values of the elasticities in the RC estimation are significantly different from those in other studies, because of the use of coefficient drivers. Second, by observing the time-varying behavior of the coefficients, we find some specific points in our time profile of coefficients; that is, we can make an association with real events occurring in Taiwan, such as the financial liberalization after 1989 and the Asian financial crisis of 1997–1998. Finally, we compare the predicted values via the time intervals and different specifications and find that we should adapt different specifications of the RC model to estimate each interval.
Acknowledgements
The authors appreciate the suggestions of the Editor and two anonymous referees who helped clarify some issues. We are also grateful to the National Science Council of Taiwan for financial support through grant 100-2410-H-110-027-MY2.
Notes
§Kia (Citation2006) demonstrates that it is possible for a policy regime change to affect the parameters of the relevant model. These estimated parameters are not constant if their changes are unpredictable or conflict with what the policy-makers had predicted.
†Payne (Citation2003) indicates that the relative absence of empirical money demand studies for developing economies is partly due to the relative instability of these economies in the transition process itself, as well as the concerns over the reliability and frequency of time series data.
‡Where Q1 is the quarter ending in March, and Q4 is the quarter ending in December.
§Swamy and Tavlas (Citation1995, Citation2001) and Hondroyiannis et al. (Citation2008) distinguish between first- and second-generation random coefficient models. First-generation RC regressions are, however, not free of misspecifications, because they do not take into account the correlations between the included explanatory variables and their coefficients.
¶Similar to a general regression approach, time-varying coefficients allow us to truly reflect the estimated periods. However, general regression coefficients provide us with long-term time trends at the expense of being able to easily analyse individual points.
⊥The reason why we do not consider the unit root and cointegration is that we want to observe the time-varying behaviors of the coefficients and the performance of prediction. This is different from previous papers that focus on the treatment of data.
∥Studies that highlight these arguments include Sargent and Wallace (Citation1975), Mankiw (Citation1992), and King et al. (Citation1991).
†Schmidt (Citation2007) finds that the rolling regression results highlight significant stability within the M1 demand vector and its long- and short-run parameters.
‡M2 is generalized currency in Taiwan. It includes M1B, time deposits, and postal savings re-deposits. However, M1B includes M1A (net currency, checking deposits, and current accounts) as well as savings deposits. Over time, broad money accommodates new instruments created as a result of the ensuing development of institutional and financial structures.
†This section draws heavily on the works of Hondroyiannis et al. (Citation2001a,b).
‡Because the quality of the time-varying coefficients allows the equation to pass through every data point even if the number of observations exceeds , the equation may be nonlinear.
§This means that is the effect of the true interest rate (
) on the true value of real money balance (
).
¶Swamy and Tavlas (Citation2006) provide a formal definition of coefficient drivers. Hall et al. (Citation2009) also indicate that coefficient drivers have two performances. First, they treat the correlation between the included explanatory variables and their coefficients. Second, they render us to decompose the coefficients of the RC estimation into their respective components.
⊥If and
are assumed to be unrelated, then
for each
and all
.
†The direct effect has its own straightforward and real-world interpretation.
‡Rates of return may be conveniently categorized as own rates, because the deposits begin earning interest; besides, the opportunity cost appears to be related to the own rates.