Forecasting Foreign Exchange Rates, A comparison between forecasting horizons and Bayesian vs. Frequentist approaches

Abstract: Forecasting foreign exchange rates and financial asset prices in general is a hard task. The best model has often been shown to be a simple random walk, which implies that the price movements are unpredictable. In this thesis models that have been somewhat successful in the past are developed and investigated for different forecasting horizons. The aim is to find models that significantly dominate the prediction performance of a random walk, and also to suggest a trading strategy that systematically can make profits using the model predictions. After investigating the data at different sampling frequencies, some significant predictive information is found for very short horizons (10 minutes) and for relatively long horizons (one week), while no useful information is found for daily data. With a forecasting horizon of 10 minutes, it is shown that a Markov model accurately predicts positive or negative returns in more than 50% of the cases for all currencies considered, with significance at the 1% level, and that the performance seems to increase with a Bayesian model. For a horizon of one week, it is shown that a Bayesian Vector Autoregressive (VAR) model outperforms the frequentist VAR model and also the random walk (although with low significance). The performance of trading strategies highly depends on the transaction costs involved. The transaction costs seem to ruin the performance on the 10 minutes horizon, while having less influence on the weekly horizon. A strategy that would have generated good profits on a weekly horizon past 2011, out of sample, is found.