Efficiency and volatility on the Istanbul Stock Exchange
This thesis investigates characteristics of the prices of shares traded on the Istanbul Stock Exchange (ISE), an important and fast-growing market. We look at five issues: the shape of the distribution of daily returns the predictability of these returns the presence of day-of-the-week effects in the mean and variance of returns the behaviour of the mean and variance of returns around stock split and dividend dates and the predictability of variances, and in particular the performance of adaptive models relative to the GARCH models. Our main findings can be summarised as follows. First, the hypothesis of normality is rejected, mainly due to excess kurtosis. To explain excess kurtosis, we used an autoregressive conditional heteroskedastic (ARCH) model, and a GARCH(1,1) model is found to fit the ISE index data well. A significant further finding, based on a t-GARCH-M model is that in the early years of the exchange, mean returns were significantly influenced by the returns variance. Second, standard tests for serial correlation, and for runs of same-sign returns, show that the hypothesis of a random walk can be rejected, with index returns showing significant first and second order serial correlation. Again, these effects are stronger in the early years of the exchange. Third, using a GARCH model, we find no strong evidence of the day of the week effect in mean returns on the index or on the 20 actively traded companies. But there is evidence to suggest that the market is more volatile on Mondays and after holidays. Again, these effects are not stable over time. Taken together, these results point to the market becoming progressively more efficient and more integrated with the international capital market over the period of the study. Fourth, the results from the EV-GARCH model, a GARCH model with event dependent intercept terms, a technical novelty, show that there is no effect on mean returns from stock dividends. Surprisingly, cash dividends do cause returns to rise/fall after their payment. On the other hand, stock dividends do significantly increase the variance of returns around the event day, and for several weeks thereafter. Finally, although we have characterised the daily returns series by an autoregressive model with a GARCH process for volatility, it turns out that the GARCH model does not unambiguously dominate alternatives in forecasting and trading applications. In 5- to 20-day ahead forecasts, the GARCH model is slightly more accurate than four alternatives, including exponential smoothing models (RiskMetrics) and historic volatility. However, it is (inevitably) less accurate than a model which pools forecasts from all models. In a simulated options market - another technical innovation of the thesis - we find that traders using a GARCH model would on balance lose money to traders using other methods, in spite of the apparently greater accuracy of the GARCH forecasts. This confirms for volatility forecasts an important result which is already know to hold for mean forecasts - that in forecasting financial markets, there is little correlation between meansquare accuracy and trading profitability.