ML in Finance: Portfolio Management via Side & Size Prediction on the Bonds Market

Abstract

The rise in automation and utilization of algorithms in the last decades had been meaningful in several areas, finance, and portfolio management included. The present study combines two approaches to reach an integral optimized model. The first one is the traditional approach, which starting from forecasting the future bond yield curve, generates a decision to take: establish a long position (expecting a rise on the price) or a short one (expecting a fall on the price). Therefore, the output of this first model will be to determine the position side. The second approach is the application of the bet-sizing technique to optimize the resulting decisions from the traditional model by assigning them a probability of being correct: decisions with a low probability of generating profits will have a lower size, while decisions with a high probability of generating returns will have a bigger size. The algorithms used were the ARIMA regression for the traditional model and random forest for the bet-sizing model. Cross-validation and out-of-sample backtests were conducted to evaluate how the model would have performed and results show that employing the integrated optimized model exhibits higher Sharpe ratios than using only the traditional approach. The work demonstrates that the modern techniques used along with the traditional ones reach better efficiency on returns than when only traditional models are employed. Additionally, generalizations to other areas inside finance, both on asset management as well as on credit risk are discussed.