یادگیری عمیق با شبکه های حافظه کوتاه مدت
Abstract
1- Introduction
2- Data, software, hardware
3- Methodology
4- Results
5- Conclusion
References
Abstract
Long short-term memory (LSTM) networks are a state-of-the-art technique for sequence learning. They are less commonly applied to financial time series predictions, yet inherently suitable for this domain. We deploy LSTM networks for predicting out-of-sample directional movements for the constituent stocks of the S&P 500 from 1992 until 2015. With daily returns of 0.46 percent and a Sharpe ratio of 5.8 prior to transaction costs, we find LSTM networks to outperform memory-free classification methods, i.e., a random forest (RAF), a deep neural net (DNN), and a logistic regression classifier (LOG). The outperformance relative to the general market is very clear from 1992 to 2009, but as of 2010, excess returns seem to have been arbitraged away with LSTM profitability fluctuating around zero after transaction costs. We further unveil sources of profitability, thereby shedding light into the black box of artificial neural networks. Specifically, we find one common pattern among the stocks selected for trading – they exhibit high volatility and a short-term reversal return profile. Leveraging these findings, we are able to formalize a rules-based short-term reversal strategy that yields 0.23 percent prior to transaction costs. Further regression analysis unveils low exposure of the LSTM returns to common sources of systematic risk – also compared to the three benchmark models.
Introduction
Prediction tasks on financial time series are notoriously difficult, primarily driven by the high degree of noise and the generally accepted, semi-strong form of market efficiency (Fama, 1970). Yet, there is a plethora of well-known capital market anomalies that are in stark contrast with the notion of market efficiency. For example, Jacobs (2015) or Green, Hand, and Zhang (2013) provide surveys comprising more than 100 of such capital market anomalies, which effectively rely on return predictive signals to outperform the market. However, the financial models used to establish a relationship between these return predictive signals, (the features) and future returns (the targets), are usually transparent in nature and not able to capture complex non-linear dependencies. In the last years, initial evidence has been established that machine learning techniques are capable of identifying (nonlinear) structures in financial market data, see Huck (2009, 2010), Takeuchi and Lee (2013), Moritz and Zimmermann (2014), Dixon, Klabjan, and Bang (2015), and further references in Atsalakis and Valavanis (2009) as well as Sermpinis, Theofilatos, Karathanasopoulos, Georgopoulos, and Dunis (2013).