"Are exchange rates predictable?" is the question posed in a recent
Journal of Economic Literature article by Barbara Rossi. This has been one of the most enduring puzzles in international finance. Starting with early work in the 1980's there has not be any strong evidence which suggests that we can use fundamental data to make predictions better than a random walk. Ouch, this is a real indictment of fundamental exchange rate models and the ability of currencies to follow fundamental information.
After reviewing all of the work on this topic, Rossi gives an unequivocal maybe on the issue of predictability after an extensive analysis of tests and procedure for measuring the drivers of currencies. For predicting out of sample forecasts using monthly and quarterly data, some models work in some environments but there is nothing to say that we have cracked our understanding of the drivers of exchange rates. This does not mean that currency markets are efficient. It does mean that we have not been able to effectively model the dynamics of these rates.
I will list some of the key finding which can be used by any forecaster to help in the future.
Purchasing power parity (PPP) and monetary models do not have any success over horizons of less than 2-3 years. Do not use current inflation differences to make a currency judgment. The PPP speed of adjustment takes a longtime and there are significant deviations from PPP in the short-run. The same applies to monetary variables especially in the short-run. Adding productivity differences or portfolio stock balances do not add to predictability.
My take on this is that inflation has been tamed in the last 3 decades, so there is just little variation in inflation relative to the movement of exchange rates. Small differences in inflation can hardly explain the large variation in FX rates. Of course this begs the question of that causes the volatility in exchange rates. With respect to monetary variables, simple use of monetary aggregates are not useful because this is not chief driver of monetary policy, Switching from interest rate targeting to monetary targeting, to a Taylor Rule, or to inflation targeting means that a simple approach will be ineffective over long time periods. The policy reaction of today is different from the '70's, 80's, and 90's. Given each country has a different policy reaction, it will be hard to use any simple relative measure to explain exchange rates. We should expect the forecasting skill of monetary models to change through time and across countries.
Uncovered interest rate and covered interest rate period have poor forecasting skill. Commodity price data for those countries that have commodity exports seems to be promising area of research. My take on this is that any simple model that is driven by uncovered interest rate parity will be flawed given the behavior of exchange rates to carry behavior. The forward rate is a biased predictor, so any model that uses this assumption will have a problem with forecasting skill.
There is little for interest rates, prices, output and money as prediction drivers. However, Taylor Rule fundamentals and net foreign asset positions are more useful. You need to have a monetary reaction functions not the raw numbers to make predictions. Simply put, if countries are both using some form of the Taylor Rule, this will show up with how they run their economies and effect relative prices of money. The Taylor Rule says that currencies will be affected by inflation differences and output gaps. Measure of financial stress and liquidity are also useful; however, it may be related to the most recent crisis events. Recent models that focus on net foreign asset flows seems to be the most promising predictors at short horizons.
From a modeling perspective, keep it simple. Linear models with the right predictors do better than those which increase statistical complexity. Predictability will change with the currency being analyzed and changing the specification of the data used (such as detrended) will have an impact on predictions. Use the right data, structure it well and employ simple specifications and you will be rewarded.
The quality of predictions change with time. What works today may be ineffective tomorrow, but given the significant changes in the currency markets, this should be expected. Even the way we pose the question will have an impact on forecasting power. You have to ask the right question For example, do you want to get the direction right or do you want to have the lowest mean squared error. A model can have higher mean squared error but do better than a random walk on predicting direction.
Surprisingly, there is little work on using technical signals or market based measures from options or sentiment as tools to predict exchange rates. There has been failure with fundamental models which is the reason why there has been a focus on price based systems. You cannot really say that random walks are the best predictors without including technical or price based forecasting tools.
There is enough evidence to suggest that working at currency predictions could be rewarded, but using a single specification or approach is bankrupt. The markets constantly change and perhaps this is the conclusion we should draw from the data.