Modeling of Foreign Exchange Swap Distributions : A statistical evaluation of two stochastic models

Abstract: The global foreign exchange (FX) market is one of the world's largest financial markets and a significant part of this market concerns the trading of FX swaps. For banks and other financial institutions, it is of great interest to model these swaps as accurately as possible, as this could improve their risk management. Numerous methods exist for modeling FX swaps, but it is not always clear if one model is superior to another. The purpose of this thesis is therefore to analyze, evaluate and compare different models that represent the stochastic processes in the FX swap market. To accomplish this, the thesis employs the reality model evaluation methodology developed by \citet{Blom_fx_pdf}. With this methodology, likelihood values for an out-of-sample period can be determined for a model, thereby enabling a statistical comparison to ascertain which model more accurately reflects the true distribution. This thesis will compare two models for FX swap prices: an interest rate model and a PIP-model. The PIP-model is constructed by determining a multivariate distribution based on in-sample observations of pips. The likelihood values for the out-of-sample observations can therefore be determined directly. The interest rate model, on the other hand, will be implemented using Blomvall's reality model evaluation in order to determine the likelihood values. It is constructed by evaluating risk factors of the FX swaps, rather than historical pips. The risk factors evaluated in this thesis are forward curves, the spot price and spikes in the supply and demand curve at certain dates. The results show that the interest rate model better represents the true distribution of FX swaps compared to the PIP-model. The statistical test of the out-of-sample likelihood values shows that the probability of the interest rate model outperforming the PIP-model is approximately 100 \%. Additionally, the result suggests that an implementation of the interest rate model using a Student's t-distribution is more advantageous than using a normal distribution, a conclusion also supported by a statistical test. Moreover, the effectiveness of Blomvall's reality model evaluation in determining likelihood values is confirmed, thus enabling the statistical comparison of different models.