A Deep Learning stacking ensemble algorithm for Stock Market classification and risk management
| Autor: | Magnani, Carlo |
|---|---|
| Přispěvatelé: | Caraça-Valente Hernández, Juan Pedro, Pérez Pérez, Aurora |
| Rok vydání: | 2022 |
| Předmět: | |
| Zdroj: | Archivo Digital UPM Universidad Politécnica de Madrid |
| Popis: | Stock Market forecasting has been historically considered one of the most challenging issues in time series due to its inherent characteristics such as low signal-to-noise, multiple forces competing for performance and basic nonexistence of arbitrage. Artificial Neural Networks, and specifically Long Short-Term Models, are powerful tools for unveiling hidden relationships in data and have become a goto technique for applications in finance and asset management. Nonetheless, there seem to be a big gap between industry and academia, where the first have been leveraging these types of models for at least a decade, while the same can’t be said for papers on the subject which are few in number and tend to be quite more recent. When looking at the portfolio of available techniques even less research has been done on the use of ensembles for regression or classification problems in finance. Trying to bridge this gap and contribute to the field we develop a Deep Learning stacking ensemble model to classify the sign of the future SP500 Index return; we do that by leveraging a heterogeneous set of classifiers and a proprietary dataset based on trend following technical analysis indicators. When back testing our model, we find evidence of outperformance for our strategies against a naïve buy-and-hold, but we note how the stacking ensemble algorithm isn’t adding any value as is relying entirely on the prediction coming from one specific classifier. When considering the LSTM classifier though, we find both higher return and higher Sharpe Ratios for the two strategies generated with the predictions; both strategies also exhibit lower Drawdown compared to the Naïve buy-and-hold for the period we studied. While the use of ensembles doesn’t seem to have increased model quality, we acknowledge more research is needed on the topic like focusing on enhancing the dataset at hand, leveraging homogeneous rather than heterogeneous classifiers or making use of actual output probabilities. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|