Do you know of any papers which consider pairs trading (or statistical arbitrage) on foreign exchange? I couldn't find any. I asked this question on several forums and got no reply. Thus I guess this trading strategy is inapplicable due to the properties of currency markets or other fundamental reasons. However, it is not obvious to me what are these reasons.
Fatih Yilmaz, formerly of Bank of America (currently BlueGold), has a piece called "Imaginal Spreads and Pairs Trading" on exactly this topic, if you can find it (I couldn't find a copy on the public internet), originally published April 17, 2009. He writes:
Academics and industry practitioners generally concentrate on time series aspects of currency markets. This is primarily the result of limited number of traded currencies. Most markets for emerging currencies are far from ignorable frictions. Hence, if one is confined to G10 currency markets, considering that shocks to the USD typically account more than 50% of variations in G10 (see Figure 1), not much room is left for cross-sectional selection skills or market neutral strategies. Pairs trading is a market-neutral (or USD-neutral) strategy and it can capture different opportunities within G10 currency markets from a statistical point of view. Moreover, given the strategy sells winners and buys losers, it is likely to be low or negatively correlated with most traditional directional models (such as momentum strategy).
Our focus in this note is to test pairs trading strategy within G10 currency markets. Our currency data set is monthly (end of month data obtained from Reuters and DataStream) and we use short-term money market rates for carry calculations (obtained from DataStream). Our data set is from 1973-2009. We take the USD as the numéraire currency and form 36 possible pairs using all 9 USD crosses. Our pairs-matching algorithm and trading strategy is described below:
Excess returns, Sharpe ratios and directional accuracy statistics generally indicate promising results. In particular, as we increase threshold misalignment level for trading signals, all performance statistics tend to improve. Generally speaking, misalignment levels around 1.5-2.0 standard deviations tend to produce consistently good results.
The Sharpe Ratios he refers to are about 0.7-0.8 for a 3M holding period.
The presented results in this note should be taken as exploratory. Nevertheless, the first set of results appears to be encouraging for several reasons. Performance statistics are relatively attractive and robust for an active G10 strategy. Especially considering that the strategy is USD neutral and the forecast horizon is over 25y. Moreover, the strategy is contrarian and concentrates on relative value trades. Hence, likely to produce low correlated returns to traditional directional currency models. If we take the presented results in this note at face value, then we should ask an important question: what are we being paid for? Transaction costs in this study are not relevant given that we used monthly data. Bankruptcy and liquidity risks and short sales constraints can generally be ignored for G10 currency markets. It would be interesting to analyse the correlation of pairs strategy returns with macroeconomic and related asset market cycles (i.e. time-varying risk premia for cycles). In their study, Goetzmann et. al. (2006) argue that the pairs strategy might be rewarding because of (hidden or latent) common factor as a main driver of the equities they analyse. There is always the risk of being arbitraged away; however the strategy appears to produce relatively robust results even in the past decade or so (when hedge fund activity increased significantly). Other possibilities might be that the strategy can be rewarding for pushing markets towards equilibrium via arbitrage trades. Given that the strategy is market neutral and relies on relative value trades, there is also the risk of missing strong market moves with such a strategy from a model allocation point of view. In any case, in our view, understanding fundamental risk-reward characteristics of such a strategy is important and requires further analysis in this context.
Disclaimer: I know nothing about FX trading, other than that I've heard something to the effect of "The first rule of FX trading is that you do not trade FX. The second rule..." you know how it goes.
I'm not into macroeconomics, but I get the impression that the benchmark for FX models is a random walk. That is to say that the fundamentals have nothing to say about FX at anything on a short horizon, which I think is considered four years. I think what has complicated a lot of the research here is limited data in floating exchange rate regimes, small policy interventions, and rare huge policy interventions.
I think Stock and Watson have the best, recent exchange rate models. These papers won't discuss trading, but could be thought-provoking in how you look at the problem
JASA 2002, Journal of Business & Economic Statistics 2002 (sorry, couldn't find link).
HTH (someone with practical knowledge will have to chime in with how to implement :) )
You may be interested in trading using correlations between different quotes - then it is like optimal selection theory for a usual portfolio. The only difference is in the model for FX quotes (while in optimization of portfolios stock models are used) - this model I am also looking for and cannot you advise anything at the moment.