Everyone today talks about AI, big data and machine learning, yet most do not delve into the fundamental properties of how they will operate and how they might be an actual threat to asset managers.
Some view technological methods as tools to assist them instead of being such a threat, and it would help provide both perspectives of the argument.
Deep Learning: How It Works
Deep learning is a branch of machine learning that uses particular architectures of neural networks. These are artificial networks that attempt to actually replicate how the neural structures in human brains operate. Such methods have successfully been applied to areas such as computer vision – i.e. image processing and classification – as well as speech recognition.
The techniques are readily available to any undergraduate student willing to learn the process. However, in the field of asset management, one have not seen the process of such methods, since they are likely kept as trade secrets.
An application of neural networks has been used for a video game that everybody knows: Mario. By connecting different inputs, such as moving left or jumping, with different outputs, such as dying or killing something, the computer program attempts to play and complete levels in the game.
Studying The Neurons
The beautiful way this is done is that the program “learns” as it iteratively attempts to pass the level. After trying again and again, the program studies how to beat the level, just as one would learn as children playing the game. This iterative nature allows it to build a library of knowledge which develops its neural network.
To relate the concept to asset management, it is necessary to explain the very basics of how a neural network operates. Neurones are modelled as nodes that can receive input from other neurones and provide output to others. Every neurone has a weight, and this determines the effect of each input line connecting it to another neurone.
A Parallel With The Economy
A very simple example from economics would be to think of neurones as inflation and consumer confidence (CC) in different countries. One first assumes that CC in the US will affect inflation in the US more than it would affect inflation in Japan.
Thus the input line weight of CC in the US to inflation in the US will be greater than the input line weight of CC in the US to inflation in Japan. However, after looking at the output – inflation in Japan and the US – one might see that CC in the US has a large impact on Japan’s inflation, hence the weight would be adjusted.
The advantage of this system is that the weights can adapt to changes, where the whole network “learns” how to assign weights. Also, the more information one has, the better one can assign controls for the regression and pinpoint the cause of any event.
Now the scary part begins. Big Data is no different from normal data, it is simply much larger. Anyone who knows how to analyse data knows can analyse big data. The problem is that data is growing too fast and too large for humans to analyse in a given time frame.
This is exacerbated by the fact that more and more transactions are moving toward the electronic field, where new structures such as Blockchain are introduced. At this point, deep learning comes in. With unmanageable data at hand, utilisation of deep learning is inevitable.
Asset managers decide how to allocate their assets in accordance with the risk-return profile of their clients. The decisions are specific and tailored according to the demands of the client. Therefore, it is not possible to write a code that will simply take an input, pass it through a pre-written set of models and obtain an output that is suitable for every client.
Moreover, most of these models are static: they do not have the attributes of neural networks. It is no coincidence that a recent McKinsey study found that:
“A dozen European banks are replacing statistical modelling techniques with machine learning.”
Also, asset managers make their decisions based on the information available to them, where they have to explain their reasoning for choosing a distinct asset. When you look at this from a big data perspective, there is a very meagre doubt that a trained deep learning program would outperform the asset manager in financial and economic data collection as well as in neural input line weight distribution.
The Human Touch
An important input that the asset manager provides is the assumptions made regarding the inputs of the valuation model. A deep learning algorithm cannot intuitively make the correct assumption, but that is exactly the point. The asset manager accumulates knowledge and experience over the years to properly make those assumptions. They observe their surrounding financial environment and are guided by their peers.
This is called the training process for the deep learning algorithm. Using past information of thousands, if not millions, of financial valuations, it can learn how the different model inputs, such as terminal growth rate and revenue growth rate, affect the final value of the firm. It should be noted that with the power of modern supercomputers, the program will be able to conduct weeks’ worth of work in a few minutes.
This all seems very naïve given that the investment process is more complex than this. However, the idea is to provide a simple example as a framework of what deep learning AI can do.
An Electric Future For Investing?
At the end of the day, will deep learning to remove the job title ‘asset manager’? Most likely not. However, the need to have so many within a firm will diminish. With deep learning AI on the table, getting into an asset management firm would require the candidates to have an understanding of its principles – if not vast knowledge and experience in the area.
Why will deep learning not eradicate this job title? Even though the program provides an output, it will be up to the manager to use it or not. And in order to construct a sound argument, the reasoning process of the program will be used. As Burak Kakdaş from a prominent asset management firm in London said:
“The process of identifying an opportunity in the market, coming up with an investment thesis and executing that idea is as creative as it is technical. Algorithms may assist the asset manager in the technical part of the process through automation or more in-depth analysis. Yet, even the preliminary questions of ‘what to analyse’ and ‘where to look at’ require a creative mind.”
The adaptation of asset managers into the world of machine learning will improve their efficiency, as well as accuracy, and possibly offer higher returns. This shift will no doubt occur over time, not overnight. But then again, this is the fundamental principle of deep learning -training using new and old data, updating the position and neural network structure, and analysing the next set of data.
However, this shift will not take decades. Even if not implemented at 100% reliance to their output, as more and more use it for a longer period of time, the algorithms will improve beyond human recognition throughout the next decade.
As a comparison, examine how Google, Facebook and Tesla used this technology to implement breakthroughs in their areas of expertise. The idea is not restricted to a fantasy science fiction novel. It is here today.