{"title":"Customized Stock Return Prediction with Deep Learning","authors":"Christopher Felder, Stefan Mayer","doi":"10.1109/CIFEr52523.2022.9776177","DOIUrl":null,"url":null,"abstract":"In finance, researchers so far use standard loss functions such as mean squared error when training artificial neural networks for return prediction. However, from an investor’s perspective, prediction errors are ambiguous: in practice, the investor prefers to see underprediction of portfolio returns rather than overprediction, as the former implies higher realized returns and thus financial benefits. We present a loss function customized to this behavior and test it based on Long Short-Term Memory (LSTM) models, the state-of-the-art tools in time series analysis. Our model learns unique signals, predicts returns more cautiously, and improves profit chances over the standard LSTM and reversal signals. Daily and weekly revised portfolios achieve on average five percentage points higher annualized returns. We show that our loss function is robust to market sentiment and beneficial in nonlinear optimization.","PeriodicalId":234473,"journal":{"name":"2022 IEEE Symposium on Computational Intelligence for Financial Engineering and Economics (CIFEr)","volume":"15 1","pages":"0"},"PeriodicalIF":0.0000,"publicationDate":"2022-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"2022 IEEE Symposium on Computational Intelligence for Financial Engineering and Economics (CIFEr)","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1109/CIFEr52523.2022.9776177","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 0
Abstract
In finance, researchers so far use standard loss functions such as mean squared error when training artificial neural networks for return prediction. However, from an investor’s perspective, prediction errors are ambiguous: in practice, the investor prefers to see underprediction of portfolio returns rather than overprediction, as the former implies higher realized returns and thus financial benefits. We present a loss function customized to this behavior and test it based on Long Short-Term Memory (LSTM) models, the state-of-the-art tools in time series analysis. Our model learns unique signals, predicts returns more cautiously, and improves profit chances over the standard LSTM and reversal signals. Daily and weekly revised portfolios achieve on average five percentage points higher annualized returns. We show that our loss function is robust to market sentiment and beneficial in nonlinear optimization.