We examine both in-sample and out-of-sample predictability of South African stock return using macroeconomic
variables. We base our analysis on a predictive regression framework, using monthly data covering the
in-sample period between 1990:01 and 1996:12, and the out-of sample period commencing from 1997:01 to
2010:06. For the in-sample test, we use the t-statistic corresponding to the slope coefficient of the predictive
regression model, and for the out-of-sample tests we employ the MSE-F and the ENC-NEW test statistics.
When using multiple variables in a predictive regression model, the results become susceptible to data mining.
To guard against this, we employ a bootstrap procedure to construct critical values that account for data
mining. Further, we use a procedure that combines the in-sample general-to-specific model selection with
tests of out-of-sample forecasting ability to examine the significance of each macro variable in explaining
the stock returns behaviour. In addition, we use a diffusion index approach by extracting a principal component
from the macro variables, and test the predictive power thereof. For the in-sample tests, our results
show that different interest rate variables, world oil production growth, as well as, money supply have
some predictive power at certain short-horizons. For the out-of-sample forecasts, only interest rates
and money supply show short-horizon predictability. Further, the inflation rate shows very strong
out-of-sample predictive power from 6-month-ahead horizons. A real time analysis based on a subset of variables
that underwent revisions, resulted in deterioration of the predictive power of these variables compared
to the fully revised data available for 2010:6. The diffusion index yields statistically significant results for only
four specific months over the out-of-sample horizon. When accounting for data mining, both the in-sample
and the out-of-sample test statistics for both the individual regressions and the diffusion index become insignificant
at all horizons. The general-to-specific model confirms the importance of different interest rate variables
in explaining the behaviour of stock returns, despite their inability to predict stock returns, when
accounting for data mining.