16.5 Summary
Deep neural networks have successfully demonstrated an outstanding performance in many domain-specific areas, such as image recognition, natural language processing, board games, self-driving cars, etc. The million-dollar question is whether this revolution will extend to financial systems.
Some problems like sentiment analysis of news for trading purposes have clearly benefited from the advances in natural language processing. However, other problems related to financial time series forecasting and portfolio design remain unclear. Among the many challenges that set these problems apart from other successful applications, we can list the following:
Lack of data: The amount of financial data is generally limited (e.g., two years of daily stock prices amount to just 504 observations). Perhaps high-frequency data provides a more promising perspective.
Low signal-to-noise ratio: The signal in financial data (capable of generating alpha) is extremely weak and totally submerged in noise.
Nonstationarity of data: Financial time series are clearly nonstationary, which makes learning the statistics complicated.
Feedback adaptiveness of data: Patterns discovered in financial data and exploited for trading tend to disappear immediately due the feedback loop mechanism.
Lack of prior human evidence: There seems to be no human capable of forecasting the future performance of companies to design a portfolio with a significant alpha. Recall the provocative statement (Malkiel, 1973): “a blindfolded chimpanzee throwing darts at the stock listings can select a portfolio that performs as well as those managed by the experts.”
Despite these challenges, the jury is still out on whether the deep learning revolution will fully extend to financial systems. It is still too early to adventure any future prediction.