"Can Markov Switching Models Predict Excess Foreign Exchange Returns?"
PDF:
Abstract: This paper merges the literature on technical trading rules with the literature on Markov switching to develop economically useful trading rules. The Markov models' out-of-sample, excess returns modestly exceed those of standard technical rules and are profitable over the most recent subsample. A portfolio of Markov and standard technical rules outperforms either set individually, on a risk-adjusted basis. The Markov rules' high excess returns contrast with mixed performance on statistical tests of forecast accuracy. There is no clear source for the trends, but permitting the mean to depend on higher moments of the exchange rate distribution modestly increases returns.
Keywords: technical trading rules, Markov switching, exchange rates, excess returns, predictability
...
6. Conclusions
This paper has used Markov switching models to create ex ante trading rules in the foreign exchange market. Markov models generate statistically and economically significant out-of-sample returns that are 95 basis points larger, on average, than those of conventional technical trading rules, and these returns appear to be fairly stable over time. The Markov rules provide at least two marginal benefits over conventional MA rules. An equally weighted portfolio rule of the Markov and MA rules provides a better risk-return trade-off than either alone. In addition, the Markov rules are strongly superior to the MA rules on the most recent data, in which the MA rules' profitability seems to have disappeared.
The Markov switching models deliver strong out-of-sample portfolio returns, although they fail to outpredict a naive, constant-return benchmark by MSE and MAE criteria. While the mean returns have diminished after 1991, tests reject structural breaks in Markov mean returns, which are still positive in every subsample, including the period from 2002 to 2005:6. Thus, Markov rule returns have been more stable than those of the conventional MA rules.
The ability of the Markov trading rules to identify trends in exchange rates might be linked to their use of information about higher moments. The fact that in-sample LR tests always preferred linking either the distribution's dispersion (scale of the variance) or kurtosis to the mean return supports this contention. Restricting the mean of the Markov model from using higher moments reduces overall mean annual out-of-sample returns by 1.5 percentage points and Sharpe ratios by 14 basis points. This suggests, but does not prove, that higher moments belong in the expectations of the Markov trading rule. The technical trading literature has not previously exploited higher moments in constructing rules.
The use of econometric methodology, rather than technical rules, to make trading decisions has at least two potential advantages. First, one can generate the entire multi-period distribution of exchange rate returns, enabling the risk-averse investor to better assess the risk-adjusted expected returns. A second potential advantage of an econometric methodology is that the stability of the model structure--rather than the return moments--can be assessed in real time, enabling traders to change their trading rules with the structure of the data-generating process. This paper did not explore those advantages.
This paper has used Markov switching models to create ex ante trading rules in the foreign exchange market. Markov models generate statistically and economically significant out-of-sample returns that are 95 basis points larger, on average, than those of conventional technical trading rules, and these returns appear to be fairly stable over time. The Markov rules provide at least two marginal benefits over conventional MA rules. An equally weighted portfolio rule of the Markov and MA rules provides a better risk-return trade-off than either alone. In addition, the Markov rules are strongly superior to the MA rules on the most recent data, in which the MA rules' profitability seems to have disappeared.
The Markov switching models deliver strong out-of-sample portfolio returns, although they fail to outpredict a naive, constant-return benchmark by MSE and MAE criteria. While the mean returns have diminished after 1991, tests reject structural breaks in Markov mean returns, which are still positive in every subsample, including the period from 2002 to 2005:6. Thus, Markov rule returns have been more stable than those of the conventional MA rules.
The ability of the Markov trading rules to identify trends in exchange rates might be linked to their use of information about higher moments. The fact that in-sample LR tests always preferred linking either the distribution's dispersion (scale of the variance) or kurtosis to the mean return supports this contention. Restricting the mean of the Markov model from using higher moments reduces overall mean annual out-of-sample returns by 1.5 percentage points and Sharpe ratios by 14 basis points. This suggests, but does not prove, that higher moments belong in the expectations of the Markov trading rule. The technical trading literature has not previously exploited higher moments in constructing rules.
The use of econometric methodology, rather than technical rules, to make trading decisions has at least two potential advantages. First, one can generate the entire multi-period distribution of exchange rate returns, enabling the risk-averse investor to better assess the risk-adjusted expected returns. A second potential advantage of an econometric methodology is that the stability of the model structure--rather than the return moments--can be assessed in real time, enabling traders to change their trading rules with the structure of the data-generating process. This paper did not explore those advantages.
My Excerpts:
...
Christoffersen and Diebold (2003) demonstrate that serial dependence in higher moments, such as the variance and kurtosis, affects the expected sign of returns in the presence of a non-zero unconditional mean return. This sign dependence creates predictability in the direction of returns. Our model does not directly exploit this effect; instead, it exploits dependence between conditional moments to better estimate the conditional mean.
For example, a rise in volatility can generate a change in conditional mean return through safe-haven effects. Higher volatility causes investors to seek safe-haven currencies, like the dollar. Decomposing volatility into both time-varying kurtosis and dispersion might improve our model's ability to detect the type of risk response associated with safe-haven effects.
Scott-> This effect is probably exactly what we are seeing now. Rises in volatility triggering the seeking of safe-haven currencies (ie. USD and JPY).
(( Conversely, as volatility starts tapering off, money will start flowing out of safe havens, and dollar and yen will begin to fall...))
Leave a comment