ResearchBib Share Your Research, Maximize Your Social Impacts
Sign for Notice Everyday Sign up >> Login

Exploring the Impact of Magnitude- and Direction-based Loss Function on the Profitability using Predicted Prices from Deep Learning

Journal: International Journal of Engineering and Management Research (IJEMR) (Vol.10, No. 1)

Publication Date:

Authors : ; ;

Page : 116-125

Keywords : Deep Learning; Price Prediction; Program Trading; Time Sequence; LSTM; TCN;

Source : Downloadexternal Find it from : Google Scholarexternal

Abstract

Researches on predicting prices (as time series) from deep learning models usually use a magnitude-based error measurement (such as ). However, in trading, the error in the predicted direction could affect trading results much more than the magnitude error. Few works consider the impact of ill-predicted trading direction as part of the error measurement. In this work, we first find parameter sets of LSTM and TCN models with low magnitude-based error measurement, and then calculate the profitability using program trading. Relationships between profitability and error measurements are analyzed. We also propose a new loss function considering both directional and magnitude error for previous models for re-evaluation. Three commodities are tested: gold, soybean, and crude oil (from GLOBEX). Our findings are: with given parameter sets, if merchandise (gold and soybean) is of low averaged magnitude error, then its profitability is more stable. The proposed loss function can further improve profitability. If it is of larger magnitude error (crude oil), then its profitability is unstable, and the proposed loss function cannot improve nor stabilize the profitability. Furthermore, the relationship between profitability and error measurement for models of LSTM and TCN with or without customized loss function is not, as commonly believed, highly positively correlated (i.e., the more precise the predicted value, the more trading profit) since the correlation coefficients are rarely higher than 0.5 in all our experiments. However, the customized loss functions perform better in TCN than in LSTM.

Last modified: 2020-03-30 16:58:59