High frequency exchange rate prediction using dynamic bayesian networks over the limit order book information

Abstract

Abstract. This work presents a special case of a Dynamic Bayesian Networks (DBN) to capture the USD/COP market sentiment dynamics choosing from uptrend or downtrend latent regimes based on observed feature vector realizations calcu- lated from transaction prices and wavelet-transformed order book volume dy- namics. The DBN learned a natural switching buy/uptrend, sell/downtrend trading strategy using a training-validation framework over one month of market data. The model was tested in the following two months, and its performance was reported and compared to results obtained from randomly classified market states and a feed-forward Neural Network. It is separately assessed the contribution to the model’s performance of the order book in- formation and the wavelet transformation. This work also constructs key trading strategy estimators based on the Ran- dom Entry Protocol over the USD/COP data. This technique eliminates unwanted dependencies on returns and order flow while keeps the natural autocorrelation structure of the Limit Order Book (LOB). It is still con- cluded that the DBN-based model results are competitive with a positive, statistically significant P/L and a well-understood risk profile. Buy-and-Hold results calculated over the testing period are provided for comparison reasons. A general characterization of the USD/COP Limit Order Books and theory behind the Dynamic Bayesian Networks are included as part of the main document.