A short time ago I wrote a brief article on the impact of blockchain and deep learning on securities trading regulation. This article goes a little deeper (pun intended) into deep learning and how I anticipate it affecting securities trading regulation. Deep learning is another technological advance that has important implications for securities regulation. In simple terms, deep learning is a form of machine learning that involves learning data representations and patterns using simulated neural networks. Deep learning requires access to a very large amount of data and immense computing power. I expect deep learning to be used by certain sophisticated traders, such as hedge funds and proprietary trading firms.
To understand the potential impact of deep learning on securities trading, it is important to have a grasp of the underpinnings of this technology. Basically, deep learning is a subset of machine learning. The latter is a computer technology that employs algorithms to “learn” from data and patterns of data to produce outcomes and decisions without the need to reprogram the algorithms. In essence, the algorithms automatically adjust their “weights” based on data inputs to perform better a specific task. For example, when you perform a search on Google but input a typo into the search term, Google might correct the typo and offer a search for the term you intended. That occurs because the Google search algorithm has learned from the data of previous searches of many users that had the same typo.
Deep learning uses neural networks to produce a more complex form of machine learning. Neural networks employ algorithms that attempt to replicate the brain’s learning capability. They do so by analyzing data to perform more refined calculations. While it is beyond the scope of this article to describe this process in detail, a major element is the addition of extra layers of analysis between the data input and the final result. Each extra layer sends an output to the next layer for performance, which then sends the output to a further layer for analysis. The process, known as forward propagation, appears in a diagram like a spider web of data result transfers. The key is that the extra layers of analysis of forward propagation attempt to produce a more accurate outcome by using a large number of increasingly sophisticated analyses.
Currently, deep learning is very expensive to employ because it requires far more computing power and resources than regular machine learning. Some giant technology companies use it because the cost involved is worth the marginal increase in the refinement of their applications. This cost, however, typically is not worth the expense for the incremental benefit it provides to most users of machine learning. This is the case for most securities traders. There are some securities traders that likely already are using deep learning and others that might begin to use deep learning as this technology becomes more refined. Specifically, large hedge funds and proprietary trading firms that seek an edge over their competitors might find that the large cost of deep learning is worthwhile if it provides an advantage in analyzing quote and trade data that enables them to obtain profitable trading results.
For example, in the 1980s, some large broker-dealers used index arbitrage trading strategies to buy or sell stocks when the leading stock index futures began to trade at a premium or discount to fair value. The initial use of this strategy was profitable because it enabled a small group of firms to capture a pricing disparity before others could react to it. Over time the strategy became less profitable as more marketplace participants engaged in the practice. Similarly, some high frequency traders 8-10 years ago began to profit from differences in the speed of access to quotation and transaction data through co-location, computing power, and multiple layers of quote and trade dissemination. Like the stock index arbs of the 1980s, the high frequency traders were able to obtain a means to profit from technological and computational advances. Also like index arbitrage, the high frequency trading field has become more congested.
It is highly likely that some large trading entities will decide that the huge cost of using deep learning for their trading strategies is acceptable to achieve a small edge over their competitors. The edge in this case is the ability to improve the accuracy of the prediction of marketplace behavior via analysis of quotation and transaction data and other variables. Although the improvement may be tiny, it could provide sufficient returns to justify the large computing costs of the technique.
When this occurs, the implication for regulation of securities trading will be profound for several reasons. First, trading based on deep learning will be very difficult for regulators to surveil and comprehend. While regulators will be able to discern the initial input of data and the final output resulting in a quote or trade, they will find it incredibly difficult to understand why the final output occurred. The additional layers of analysis incorporated into the deep learning structure will make it exponentially more difficult for a regulator to determine why a firm placed a quote or trade.
Second, trading incorporating deep learning complicates the notion of scienter in a trading strategy. If the trading produces a result that regulators believe might be abusive or manipulative, that result might be the unintended consequence of multiple layers of reaction to algorithms automatically adjusting weights based upon inputs from the previous layer of algorithms in the forward propagation process. There may have been no intent on the part of the trading firm to produce a result that regulators find abusive or manipulative. Consequently, the mere impact of a trading strategy on the marketplace could be insufficient to prove manipulation or scienter-based fraud.
Third, a trading firm’s compliance department will face similar obstacles in constructing written supervisory procedures and compliance reports and alerts for deep learning strategies. While procedures and alerts will be based upon the final results of a deep learning-produced quote or trade, the complexity of the process that produced the results will be difficult for compliance staffs to analyze. As a result, compliance staffs may need to revisit how to conduct oversight over this type of trading.
While at present the prevalence of neural network based trading may be limited in scope, it is not fanciful that in a relatively brief period such trading may expand greatly due to technological advances in artificial intelligence and reductions in the costs of the computing power needed to employ neural networks. At that point it is likely that the impact of deep learning trading on “fair and orderly markets” will cause regulatory consternation in the same manner that index arbitrage did in the 1980s and high frequency trading has over the past 8 – 10 years. It is not too early for securities regulators to begin to contemplate how they will oversee such trading and provide guidance on the legal ramifications of such trading. Similarly, trading firms in adopting deep learning based trading should involve Compliance at the outset of designing their neural networks. This might involve adding computer science expertise to Compliance staff. One thing is for certain: our securities markets will be changed profoundly when – not if – deep learning becomes more widely used in trading strategies.