AI in Investment Management

 

AI in investment management is approaching? What can a fundamental investment manager learn from machine learning? On the contrary, what can a data scientist learn from the fundamental investment? For both, plenty, it turns out. For the past years, BALANSTONE has conducted deep-dive technology research on machine learning and deep neural network. This research activity has extended and resulted in the design and development of the new AI of fundamental investment. In this article, we would like to share what we have learned through our journey: the keys to developing AI for fundamental investment research.

Evolution of AI (ML and DNN)

AI has made considerable progress in the last several years. Likewise, machine learning (ML) has made progress, and text analysis, including recommender systems, has got advanced significantly in the last decade.

In particular, deep neural network (DNN) has got a major advancement. Many successful cases of DNN were reported for the selected applications such as natural language processing, image recognition, and text analysis. Deep reinforcement learning’s current success has been remarkable, and it beat the world’s top-ranked professional. That event has become a milestone in the birth of super-human level AI.

There have been several catalysts that have enabled the success. They include the elements and designs of the model and the size of data available for training and the improvement in the computing capability. All of them led to the development of the theme: “AI replaces humans.” https://goo.gl/FBX5j3

Fundamental Active Investment

The story of evolutional AI is nothing new. Still, most professionals in the fundamental investment community should have asked the question in mind: does AI technology innovate the fundamental active investment, and if so, how? Everyone in the fundamental investment industry is aware of its potential influence, as indicated by the equity market. At the same time, the fundamental investment industry is in a structural shift. To search for further return potential and accuracy (i.e., low risk), the bottom-up information gathering activities have got deep and intensified. Focus and engagement became popular strategies. Some even find the limit of early acquisition of legitimate information, which ended up as several insider information cases. On the other hand, new rule-based & low-cost funds are gaining share. The elevated fee for HF and PE is getting scrutinized. All in all, the application of the fundamental approach got diversified. Still, the core approach of fundamental research remains largely unchanged, and the same textbooks have been used for many years. Innovations in fundamental investment research do not seem to be progressing at the same velocity as other industries in the world while seeking new opportunities in ESG and other initiatives.

We had the same question before and after the work of elucidating the current feasibility of AI for fundamental investment, which is structurally and conceptionally different from what we traditionally call ‘quantitative investment.’ We have learned several clues uniquely from the journey of our development process.

Artificial Fundamental Investment Research Intelligence (AFIRI) 

In contrast, algorithmic program trading has been progressing for years, and the application of AI to such algorithms has already had good enough results. It evaluates the positions of trading rapidly and selects the best next moves constantly without a human. A human cannot execute such ultra-high-speed trading. There are even performance comparison sites of AI trading for the platform to provide the development environment, make the participants compete, reward them, and gain information from their activities. The continued innovations will make the trading more and more automated, and human traders will be finding new roles as the developers and the managers of the trading machine.

However, AI for fundamental investment research is “fundamentally” different from automated AI trading systems. The objective that we focused on was to develop the process of the intrinsic value-oriented investment and not to learn from successful AI trading strategy and to apply it. Intrinsic-value-oriented fundamental investment is not made of consecutive successful trades. The following description would be useful.

“Intrinsic value dominates the process of fundamental investment management, and the price is just an attached tag to the business as a system to deliver an excess return. According to the speculative activities and the market-making function, the price fluctuates by its nature to balance demand and supply continuously. Thus, price movement must be understood as a crucial factor of action, but it is not the dominant and primary factor that governs the fundamental investing decision.”

To further clarify it, try to give thoughts on what would happen if the objectives of a portfolio were set to maximize the short-term risk-adjusted return on the following rationale.

“Because the long-term return is simply a sum (compounded) of short-term returns, to be successful in the long-term is the same as to be successful in the short term. As such, we have to control short-term returns and measure our results as the most important performance indicator. That makes investment process successful in the long term.”

What is wrong? It looks as if there were no gap between short-term trading and long-term investment.

Most of you must have discerned the trick of this rhetoric immediately. However, it is completely ignorant of the truth that the process and approach of short-term trading are totally different from those of the fundamental investment. Time horizon forms the distinct border between the two.

The best investment is not the same as continuous and successfully executed trading ideas. A trading strategy needs to be oriented toward predicting price, and it is essentially a price-oriented decision generator. Intrinsic value does not play a key role as the determinant of the decision. AI is typically deployed to analyze and predict the relative price behaviors across many assets.

As a result, it turns out that the great AI framework developed and optimized for best trading is not compatible at all with fundamental investment research that has a different objective. However, AI for trading may (depends on the execution process) help traders of an investment team execute the orders.

A distinct name tells what specific task AI does. Automated AI algorithmic trading can, for instance, be called ATI, Artificial Trading Intelligence. In fundamental investment management and research, it should be called AFIRI, Artificial Fundamental Investment Research Intelligence.

This makes it clear that AI belongs to the specific process of the targeted process that gets intellectualized.  The process must be explicit to get modeled accurately.

Framing the General Process of AI

Understanding the process flow of AI at a high level clarifies the ideas of what and how data should be created and processed in each element of the pipeline. The typical five steps are:

1. Environment
2. Representation of Environment
3. ML/DL Process
4. Reasoning
5. Prediction

The environment selects the targeted activity and sets how the whole task of AI is defined against it.  Representation is a process of converting the targeted activity into the format that can be processed through the training of ML frameworks or deep neural networks. Hyperparameters are set and tuned for ML/DL process, and the dataset is ingested in the training process to process reasoning and develop knowledge while minimizing the penalty for error. The completed network is tested and then used to provide predictions on a new input dataset.

As the start step defines the following pipeline, selecting the right set of the environment in the most representative manner is crucial. Specifically, for AFIRI, fundamental investment AI, what must be done first is to explicitly frame the fundamental investment process and identify the distinct characteristics of the process.

In our experience, this process should be very carefully and thoughtfully done. Contrary to the general notion of the advantage of large data size (i.e., the larger the dataset becomes, the better the results become), the right configuration at this step significantly reduces the time for the fine-tuning process of the hyperparameters that are requisite for developing a robust neural network. Data should not be expanded just because of the probable betting that there may be a potential and hidden causality. Those presumably ‘the larger, the better’ approach ends up with uncontrollable data lake.

As touched, the challenge of artificial intellectualization is a composition of the two challenges: the cognitive functions of artificial intelligence and the human intelligence as the corresponding role model. Thus, the road toward AFIRI from the legacy fundamental investment research accompanies a journey to rediscover the details of the current process of value-oriented investment decisions.

As such, the targeted decision-making model needs to get understood and analyzed carefully and to give the intellectualization approach of AI a solid and consistent framework. The next step is to understand the properties of the gap of the decision process between fundamental investment research and ML/DNN task. Finally, we will compare and contrast the two steps of the decision-making process that are similar but distinct from each other.

 

Difficult Judgement and Complex Decision

The two steps which must be understood are difficult judgment and complex decision. You can also apply this framework to evaluate the feasibility of self-driving functions, for which some have been claiming the perfect world is coming soon. I am not sure if this will be the case.

Apparently, there does not seem to be an obvious difference between these two. Therefore, as to AFIRI, the question to be answered becomes:

“Is fundamental investment research a difficult judgment or a complex decision?”

Describing the cases of AI is the best guide to get the ideas of what difficult judgment and complex decision are, respectively.

First, investigate the task of image recognition and AlphaGo.

Image Recognition

The difficulty comes from building the pipeline that handles the parameters that represent the original data effectively enough to train the network for the image recognition task. The two key developments were 1) convolutional neural network architecture and 2) other components of network structure, including activation function, regularization, and residual learning. Two-dimensional image data is expressed digitally in the format of the vector of pixel data.

Why was a convolutional neural network required instead of a plain vanilla deep neural network? It is just because the network would get too big to get handled.  1920 × 1080 HD image has just over 2m pixels. So the first layer of a plain vanilla DNN would need to have over 100k neurons (assuming 1/20 reduction from 2m to 100k in the first layer), which is followed by many layers of neurons (although a declining number of neurons in general). If you train such a network and use millions of pictures as a training dataset, it becomes practically impossible to complete this job without an unrealistically large computing capacity.

AlphaGo

For a board game such as Go, the difficulty lies in the challenge that there are exponentially large numbers of legal moves of stones, equal to b^d possible sequences of moves, where b is the game’s breadth (number of legal moves per position), and d is its depth (game length). For example, GO has 250¹⁵⁰ possible moves. Even the largest and the most advanced computer cannot deal with such a huge tree of moves.

The solution is to train the model without predicting every possible position of stones and determining the best possible move. When the first article was published in 2016, AlphaGO used the unique hybrid approach of Monte Carlo Tree Search and trained neural network as an intuitive predictive engine for future legal positions of stones to overcome the difficulty. It also used a convolutional layer in which the board position was passed as a 19 × 19 image to construct a representation of the position, reducing the search tree’s effective depth and breadth.

In 2017 AlphaGo evolved into an unsupervised version, AlphaGo Zero, which was further developed to become a generalized version for the board games: AlphaZero. The network is trained with random play and reinforcement learning with the knowledge of the game rules without domain knowledge.

When training the network, larger training data would result in better results of low variance and bias. For image recognition, the volume of the training dataset is not always generated by the training algorithm, and the variety and volume of data could become a constraint. However, one of the special features of a board game is the simple and static rule in a closed and controlled environment. Therefore, it is the perfect task for the training algorithm to execute unlimited numbers of self-play simulations and self-generate data to execute reinforcement learning. In other words, unsupervised reinforcement learning has proven its great potential through AZ. However, it was not accomplished without the special and unique conditions that the board games have. The real world is so inconvenient and so complex that it does not allow simulations with true results for reinforcement learning.

What definition should be given to those tasks of difficult judgment?

As seen in the two tasks, the successful solution was achieved by cracking the key bottleneck of extremely large processing volume in addition to the benefits from a rise in the computing processing capability. The AlphaGo used the hybrid network architecture as well as a convolutional layer. The image processing used several new architectural approaches, including ReLU activation functions, dropout, short-cut connections, and convolutional layer as the most important ones. They were not general techniques of handling a large volume of data, however. Instead, new approaches were uniquely developed to suit the type of datasets in the pipeline. The difficult judgment was required when the process needed a massive volume of processing, and it was solved by an original approach in designing the pipeline.

Another key and common feature of these two cases is that their process ingested all the variables with possible causality. The process was successfully got isolated and became complete without any other data. Their success required neither finding and representing missing relevant variables nor adding them to the dataset.

In other words, the difficult judgment has a vector of the complete relevant variables that have been identified in a highly accurate manner. Thus, the execution of the judgment was difficult, but the ingestion of the complete variables was not.

The third feature is that the dataset of difficult judgment was homogeneous. GO was represented by the set of locations of black and white stones. Image recognition used pixel data as its representation. There are 3Vs in big data, volume, velocity, and variety, and for the difficult judgment, the dataset had two Vs except for variety. Thus, each of the two cases did not need to develop a new technique to deal with the heterogeneous composition of the dataset.

Difficult Judgment

What definition should the difficult judgment for artificial intelligence be? We think that it is the judgment with the above-mentioned three attributes of 1) extremely large processing volume, 2) complete representation, and 3) homogeneous dataset.

While 2) and 3) are given 1) becomes the main challenge of the difficult decision. The challenge of large processing volume has been conquered by the impressive development of new pipelines and techniques on a new hardware/semiconductor.

Those development has been made not only by the advances in the purely mathematical approach but by the engineered pipeline and the new functions that were inspired initially by the cognitive process and the logical reasoning process of a human being. We will touch on this later. Furthermore, taking advantage of the unique closed and controlled condition was another driver. As such, understanding the approaches as well as the conditions that solved the difficult judgment successfully would guide what types and directions of new ideas can be promising in the future.

Complex Decision

Complex decisions are different from difficult judgments. First of all, judgment is followed by a decision. That is a fundamental sequence of the whole decision process. It is also important to be cognizant that the judgment is a subset of all the input required to complete the decision process.

There are several essential characteristics of a complex decision in ML/DNN, which include the following elements:

• Incomplete Representation

There are two types of incompleteness for complex decisions, and type 1 relates to the lack of explicit breadth of input data. Remember that the 2^nd feature of difficult judgment is a complete representation and is antithetical to this incompleteness. As complex decision relies on the aggregated dots of fragmented information and views, the definition of explicitness is generally unwarranted. In contrast, the complete input data covers the entire information and views that have any causality with the final decision.

However, for the tasks that require complex decisions in the real world, it is impossible to gather data explicit and complete enough to cover everything and process the dataset ideally to train ML/DNN model.

Type 2 relates to the unstructured and descriptive information that is difficult to get converted to a dataset with limited consistency and availability. Think about the impression of the interview with a person. If an interviewer is responsible for reporting and a template of the scores of stereotypical categories is used, how accurately and effectively can the stylized format capture all the relevant nuance of the impression for the best decision-making? Is there a proven system to measure the effectiveness and to improve it continuously for better results?

Some might have a question if the unstructured text data mining system of sentiment analysis such as LDA could be applied to type 2 to process the conversion. It does the conversion from analog and descriptive information, and they have already established successful cases of AI’s unstructured data processing. However, those technologies have a limited scope of applications. As they are intended to classify large unstructured data or find some commonalities from it, significant information with infrequent appearances is hard to be identified as such. It is especially good at finding the popularity among big data. Human professionals with domain insight can identify a small clue of logic with significance from a small piece of unstructured data in the big data lake very well through presumably connecting multiple sets of independent cognitive networks.

• The requirement for Body of Highly Intellectual Knowledge

The complex decision deals with uncertainty. To lessen the unfavorable uncertainty of the decision, a human has done the activity to improve the quality of decision, and it is called an education. We build the system of intellectual knowledge from learning multiple subjects. For instance, in investment management, the certified professional must equip the body of knowledge (BoK) of investment management and research, covering economics, accounting, financial management, derivatives, and corporate governance. The owner of BoK must integrate them and apply them to form predictive views.

BoK is also shaped by the professional experience, and they are not static but a continuous system. Although there are more and more questions raised from academics about the actual cognitive ability of a human to learn from the experience and gain expert insight, the complex decision has the nature of advancing its network through intellectual reinforcement learning as professional work experience.

What does it mean from the perspective of a neural network structure? There are a couple of possible impacts on the network. The first one is the existence of multiple neural networks, combined and working together. Each subject is processed by a specific network. Because the result requires integrating multiple subjects with a different and independent network, the neural network should be in the form of an ensemble learning structure. Unfortunately, it results in a new challenge, and the processing pipeline gets complicated and difficult to set up and compute.

The network also needs to have a capability of self-improvement to learn from the experience, dynamic nature of BoK. There are growing interests in the new architectural approach such as progressive and sequential machine learning to improve neural network training.

• Uncertainty

Uncertainty is what difficult decisions must handle. In the investment industry, investors know that risk is not the same as uncertainty. The risk is measurable. Price volatility is called a risk. The major market crashes in the past have got triggered by the rise in fear of uncertainty.  Nassim Nicholas Taleb developed the theory of black swan events, but we propose ‘Haunted House Theory’ to separate uncertainty from risk. In the haunted house, you are blind and unable to measure and predict what would happen even using all the available information and knowledge. Uncertainty is the status where one faces the events with the unmeasurable outcome, which is an antithesis to the deterministic viewpoint.

The complex decision is forced to select what to do when it faces an uncertain event with an unmeasurable nature without a logical causality between incident and result. For the prediction using artificial intelligence, this is a fundamental challenge. Using an analogy, it is just like a situation when someone walks in total darkness and must decide which way to go where the floor may disappear. As a result, there is very little data volume and reliability as input and a large margin of error. In the end, that is why there is a large market crash.

Back to the Original Question

“Is fundamental investment research a difficult judgment or a complex decision?”

It is time to review the actual fundamental research activity until the investment decision. It is a much easier task if you are an experienced bottom-up analyst or portfolio manager with a time horizon of investment, not trading.

While the final decision depends on the valuation and the reliability of the forecast, the investment decision is neither science nor art. Just to make financial forecasts alone depends on many fragmented elements, including qualitative/subjective information and quantitative data and framework of the financial model. The qualitative information is descriptive and formed by combined insights from various activities, including meeting with company executives, evaluations of the competitiveness and strategic risks of individual products/services with reference to their financial results. Therefore, each investment case requires a different mix of information and data.

As such, for investment judgment, it is virtually impossible to define a complete data representation, and even if someone could set a complete dataset sometime in the future, it would need to get cleansed significantly, or it may need to get updated to add the new information/data to the dataset. I believe it would not happen because the cost outweighs the benefits. Thus, investment research does not meet the conditions of type 1.

For type 2, think of a fundamental investment case presentation to the team. The objective of the presentation is to articulate and share the recommended action (or rating), valuation, and risk assessment and to get approved by the team or the portfolio manager. However, the real input data ingested into the investment decision process is not only the figures but also (more importantly) the structured logic supporting the certainty of the figures.

The information is descriptive, unstructured, and combined to form the logic. They have lots of subtle nuances of ideas. Using the scores of stereotypical and standardized categories instead is an approximation that may force us to nullify the uncommon distinctiveness of fundamental investment research and lose the invaluable essence of the descriptive information that substantiates intrinsic value. There is a conflict between being explicit and being nuance-filled, and the data representation is unreliable and inaccurate for the latter. As such, the incomplete representation of data is an inevitable and big challenge for advanced fundamental investors to think and act explicitly when utilizing new cutting-edge information technology.

Note there is an application of LDA toward investment articles and market news unstructured data analysis. In my view, this is a new application to enhance ATI, not AFIRI. It is ATI because the market herding activity gets analyzed by machine learning, and then the trained algorithm tells what is influencing the market trend and predicts the move. This is nothing to do with the analysis of the intrinsic value. The data from herds is designed through ML to develop insight into the activities of beauty voting in the market, not to build the AI perspective into the fundamental value.

BoK of investment research consists of many categories. It requires having both the knowledge and working experience of handling tools and analytical frameworks of investment research, in addition to the in-depth and latest insight into specific industries and business & financial models of companies. Continuous learning is required to understand the dynamic landscape of industry and company as well as the analytical techniques.

For uncertainty, investment decisions are made not only under a wide variety of uncertainty. More importantly, investors evaluate and investigate uncertainty (through a certain brain network) and use the results as the input vector of another decision network. It is not a condition of the decision but a target of analysis of the decision. For the substantial uncertainty in a case such as a major market crash, evaluation gets very difficult due to a lack of relevant information and data.

All in all, it is obvious that fundamental investment research is a complex decision. This has a profound implication for developing DNN/ML as AFIRI. While difficult judgment can be and has been answered by new AI, which adopted a new network architecture and functions, the complex decision cannot be made easier by the same approach. Different approaches and expertise for development are required.

Is it an impossible target to develop AFIRI? If not, what needs to be done to ease or overcome the challenge?

To search for the clues, we can garner invaluable inspiration from the history of AI.

 

Catalysts for Success: Cognitive and Logical Process

David Hubel and Torsten Wiesel, who received the 1981 Nobel Prize in Physiology or Medicine for their discoveries concerning information processing in the visual system, found the key discoveries that led to the development to solve the difficult image recognition task of AI.

The first finding was many small local fields in the visual cortex, which independently work to get the information and recognize the whole image as the aggregate of small fields. The second finding was that visual neurons have a specific line orientation. Some react to one particular line orientation, and others do so to another line orientation. In addition, visual neurons have different functions, and they work together to recognize complex patterns.

Those discoveries together led to the neocognitron, a hierarchical, multilayered artificial neural network in the 1980s. It was used for handwritten character recognition. The model needed a time of further time for development. In 1998, LeNet-5 architecture, a much more polished model of convolutional neural networks, was introduced. Specifically, the use of convolutional layers, as well as feature maps from vertical and horizontal filters in LeNet-5 architecture, was influenced clearly by the findings of the visual neuron mechanism.

The current CNN has its origin in LeNet-5. It is not exaggerated to argue that the image recognition AI has developed initially from the analysis inspired by a human cognitive process starting from the visual cortex as sensing devices.

In the case of the original Alpha GO, which uses CNN, its development was built upon the discoveries from human logical processes to a lesser extent. It used two independent networks to build the pipeline. The first network is the policy network which selects the next positions intelligently. The second network is the value network, evaluating the whole positions precisely. The dual-process architecture of the original Alpha GO has a similar logical architecture that human players use to play GO.