Breaking the Machine Learning Barrier

Amazon has recently launched a revolutionary move to open a supermarket with no checkouts. This has been tested out by staff over the past year, combining the use of ceiling-mounted cameras, smartphone app and sensors on shelves to automatically bill shoppers. According to Gianna Puerini, head of Amazon Go, “this technology didn’t exist – it was really advancing the state of the art of computer vision and machine learning”, increasing efficiency and potentially the return rate of customers. At this stage, you might be wondering: How does this relate to the hedge funds industry?

To compete in the digitally driven world, businesses are constantly looking to stay ahead of innovations and develop efficient systems that drive performance. Quant funds, for example, use quantitative analysis, mathematical models and powerful algorithms among others to select optimum investments for the investment fund. On top of that, the rise of new technologies such as Natural Language Processing capitalizes on the concept of machine learning to understand and manipulate human language, raising business efficiencies. While we laud the merits of AI in earning more business profits, statistics have shown that funds driven by AI strategies, though beating the broader hedge funds industry, have yet to surpass the stock market. This is evident from a hedge fund research conducted by Eurekahedge, where thirteen AI funds gained an average of 10.6 percent annually in six years through 2016, and rose 8.5 percent through October but falls short of S&P 500 performance.

While the AI scene is gaining much momentum in recent years, its power with data still face severe limitations. “A machine would have no basis for predicting a crisis since each one is unique,” said Dhar, a professor of data science and business at NYU. Indeed, while machines can retrieve and generate financial metrics much faster than humans, they can almost never replace humans, at least in the near future, to draw comparisons and predict crisis without sufficient historical examples. The subprime mortgage crisis is one example where humans are able to put together a myriad of political and economic factors and possibly predict a financial meltdown, even with little or no historical events/data to back it.This leads me to an interesting aspect of human-AI interaction: Transcription

As an investment fund manager, one would wish to create and maintain a comprehensive database of information ranging from investor prospect lists, legal records, email correspondences and relevant videos, among others. However, the storage of videos or audio files may take up huge storage space. Furthermore, searching information or keywords in video files may prove to be a tedious task if done manually. Fortunately, there are tools available in the marketspace to facilitate this process.

Transcription software serve to transcribe audio and video files to text. However, two issues arise from the use of these software: Accuracy and Speed. When processing chunks of data in the form of audio or video, it might pose as a challenge to automate the transcription process in a systematic way. On top of that, transcribing data fast might not guarantee the reliability and utility of the product, which brings me to the next point on Accuracy.

Consider the following excerpt transcribed from a recent CNBC interview with Rune Madsen, cofounder of Runestone Capital:

“…there's also another phenomenon to bear in mind is the low volatility track. So a lot of dealers are long ball or Gamasutra speaking by hedging that government they buy on down days and sell high on the updates. So that hedging itself has sort of killed or negated the vibration of markets generally speaking. But I think what the is also there's a lot of Stop-Loss mechanism in the strategies and that's why it has been so successful in creating increasing value.”

Confused? Now consider the human transcribed version of the same excerpt:

“…there’s another phenomenon to bear in mind is the low volatility trap. So a lot of dealers are long vol or gamma so to speak. By hedging gamma they buy low on down days and sell high on up days and that hedging itself has sort of killed or negated the vibrations of markets generally speaking. But I think one development is also that stop-loss mechanism in these strategies and that’s why they have been so successful in increasing AUM”

As a reader, you would probably be confused by the first version, although there were key fund-related terms that were accurately transcribed in both versions, such as ‘low volatility’, ‘hedging’ and ‘stop-loss mechanism’. Interestingly, human transcription has further revealed nuances of the language such as ‘long vol’ and ‘AUM’ which even humans might struggle to accurately interpret sometimes, not to mention machines.

We look to combine the merits of both human and artificial intelligence which seemingly complement each other. For instance, if out of a large sample database of audio and video files, chances of the machine mistaking ‘gamma’ as ‘gama’ are statistically significant, perhaps we can alter the machine’s functions to automatically correct ‘gama’ to improve the accuracy of the transcribed text. The same applies for terms like ‘vol’ as compared to ‘volatility’. The resulting database of information becomes more comprehensive and reliable, and it includes extracts from audio and video files.

While we laud the benefits of natural language processing, critical thinking of humans is key to breaking the barriers and surpassing the limits of machine learning, achieving the best of both worlds.