APPLICATION OF MACHINE LEARNING IN FORECASTING CARRY TRADE RETURNS USING MULTIVARIATE DETERMINANTS

Abstract

Carry trade strategies have grown increasingly popular over the past few decades due to their potential to allow investors to earn high arbitrage profits from both interest returns and the exchange of currencies. However, a key obstacle to investing using carry trades is mitigating the negative skewness and positive kurtosis of returns, which results in high crash risk and has yet to be proven diversifiable. Timing the purchase and sale of currencies for carry trade portfolios has also been a significant challenge, due to the difficulty in determining the risk factors that accurately predict sudden currency movements that correct exchange rates towards fundamental values. Following the 2008 financial crisis, recent data has introduced further complexities and raised questions on the underlying theoretical mechanisms behind carry trades, including the well-known forward premium puzzle. This paper investigates whether trends observed by researchers for past data still hold for recent data, taking into consideration the increasing prevalence of negative interest rates. Given the limitations of traditional linear models, machine learning approaches, including CNN-LSTM, would also utilise several variables of uncertainties to predict, backtest and compare performances of carry trade portfolio returns.