K-NN based forecast of short-term foreign exchange rates

Abstract

Recently, day trading, that is, short-term trading that sells/buys financial instruments multiple times within a single trading day is rapidly spreading. But, there are few studies about forecasting short-term foreign exchange rates. Against this background, this work proposes a method of forecasting short-term foreign exchange rates based on k-nearest neighbor (k-NN). The proposed method restricts the search range of k-NN, and avoids forecasting the exchange rate if a certain condition is satisfied. We experimentally evaluate the proposed method by comparing it with an existing k-NN based method, linear regression, and multi-layer perceptron in three metrics: the mean squared forecast error (MSFE), the mean correct forecast direction (MCFD), and the mean forecast trading return (MFTR). The results show the proposed method outperforms the other methods in terms of both MCFD and MFTR, which implies reducing the forecast error does not necessarily contribute to making profits.