Not logged in.
Quick Search - Contribution
|Title||Improving the Currency Carry Trade Strategy: From Time Series Models to Machine Learning Methods|
|Institution||University of Zurich|
|Faculty||Faculty of Business, Economics and Informatics|
|Number of Pages||37|
|Zusammenfassung||Currency carry trade strategies (CCTs) have proved to be very profitable strategy for many years both in developed and emerging market countries. These strategies involve borrowing in low yielding and investing in high yielding currencies. The strategies are motivated on the basis of empirical failure of uncovered interest parity (UIP), which states that FX movements should offset any interest rate differential between respective currencies. However, even achieving slightly higher long-run Sharpe ratios than equities or fixed income, these strategies tend to perform rather poorly in bad times. In order words, the returns of investing in CCT strategies are typically negatively skewed and posses large kurtosis (i.e. the “Peso Problem”). The aim of this paper is to employ various statistical techniques with the goal to improve the performance and statistical properties of CCT strategies. That is to be done by forecasting exchange rates and imposing a different sorting scheme in the next step. All along, various practical considerations (e.g. different date conventions on FX derivatives market) were taken into account. We approach this task by finding suitable data for the analysis. Large data set is collected and consists of 49 spot and forward/NDF price series, real effective exchange rates, commodity futures prices and various investment and portfolio flows data. The longest data range in our data set is spanning from December 1996 to December 2017. Once the data is prepared we start building the baseline strategy – the naïve CCT. In line with the literature on the topic, we construct CCT strategy by calculating forward discounts of various currencies. Once they are obtained, 6 quantile portfolios are constructed based on the based on the CDF of forward discounts in a given period. Thereafter, 5 HMLfx (long-short) portfolios are build by taking a long position in a sequence of quantile portfolios (from second to sixth) and a short position in the first quantile portfolios. Interestingly, the naïve strategies themselves, tested on multiple developed, emerging and frontier market currencies, deliver very solid performances over the past 20 years. Smaller sample of world most liquid currencies incurs big losses amid financial meltdown of 2008. For the purpose of correcting such negative effects of CCT strategies, we developed a time series framework with the focus on long-run equilibrium predictions. We aim to obtain a well-behaved forecast system and define 5 different model specifications used for estimation and subsequent prediction. These are four different VAR and one VECM representations. Along the way, we employ different tests, by testing for unit root and cointegration among the series. We allow model specifications to change over time and in particular to posses a common long term relationship. After the predictions are obtained, we compare the resulting performance statistics to the naïve CCT benchmarks and test the forecast accuracy against a zero-mean random walk model. We find that the time series currency carry strategies for high carry currencies manage to partly mitigate the “Peso problem” during the financial meltdown, it fails to work ever since. We think that this is partly because linear models fail to capture non-linear relationships in emerging currencies. Much more pleasing results are found in G10 currency space, where even time series methods yield better predictions than a driftless random walk model. Thereafter, we gradually increase the level of complexity towards a popular machine algorithm – Random Forest. We firstly specify parameters of the model, which are chosen in accordance to theory. Then we impose data preprocessing to get the large amount of data ready for training purposes. Thereafter, a complex practical backtesting challenges in training time series data are presented. Our results indicate that machine learning CCT strategies can lead to significant improvement of naïve CCT strategies. In particular, the monthly-rebalanced strategy based on the large basket of currencies yields significantly better results than the naïve counterpart. Last but not least, we see very high potential for future research in this area.|