A multifactor model of Philippine stock returns using latent macro risk factors

Abstract

This study addresses the measurement problem inherent in using published macroeconomic data to proxy for unobservable macro risk factors. Using a structural equation modelling framework, four latent risk factors are identified in addition to the exchange rate and market timing risk factors. The latter two are proxied one-for-one by fluctuations in the nominal peso-dollar exchange rate and the excess return on the portfolio represented by the Phisix. Using transformations of published macroeconomic data as indicator variables, the measurement part of the model appears to capture the essence of the unobservable risk factors they are supposed to measure.

Variables used as such are called manifest exogenous variables (see Maruyama, 1998 or Hair et al., 1998).

The betas or factor loadings generated by the structural model indicate that fluctuations in the risk factors have significant effects on the time variation of stock returns. However, the effects are more pronounced after the onset of the Asian financial crisis than before. Tests of model fit do not reject the hypothesis that the returns are consistent with the six-factor model for the entire period covered by the study, although individual factors are priced only during the 1997–2001 sub-period.

Notes

Variables used as such are called manifest exogenous variables (see Maruyama, 1998 or Hair et al., 1998).

Interest rate and spread variables are annual figures. To minimize any scaling problem the indicator variables for business cycle and external risk are in index form with the December 1991 data taking the index value of 100. Using the Augmented Dickey-Fuller test, the unit root hypothesis for all transformed variables, including the market excess return and change in exchange rate series, is rejected at the 0.01 level of significance.

Composite reliability is measured by construct reliability and variance extracted (see Hair et al., 1998). Construct reliability is the ratio of the square of the sum of standardized loadings to the square of the sum of standardized loadings plus total measurement error for all indicators. Variance extracted is the ratio of the sum of the squares of standardized loadings to the sum of the squares of standardized loadings plus total measurement error for all indicators. Commonly used thresholds are 0.70 for construct reliability and 0.50 for variance extracted.

In addition, a Hotelling's T test of the equality of the vectors of betas during the two sub-periods is rejected at very tight levels of significance for all portfolios. The test assumes multivariate normality of the betas and equal covariance matrix for the two sub-periods.