Diverse use cases and pitfalls of AI in capital investment

There is a lot of talk about generative AI right now. But what we're talking about is a broader category of techniques. Some State Street Global Advisors (SSGA) clients are considering using machine learning to support their asset allocation process. Machine learning, which makes up the bulk of AI, can be divided into supervised learning and unsupervised learning. Furthermore, machine learning includes not only reinforcement learning but also deep learning.

In the context of machine learning, supervised learning involves training a model on a dataset. This means that each training sample is paired with the correct output. Unsupervised learning, on the other hand, aims to find hidden patterns in the input data, and there is no correct output associated with each training example. When it comes to deep learning, it uses multilayer (deep) neural networks to analyze different types of data and learn from vast amounts of data at a complexity that humans cannot match. Finally, reinforcement learning trains the model to make a series of decisions by rewarding or punishing the actions it takes and learning what to do to achieve a goal.

The difference between techniques and when to use which technique depends largely on the task at hand. For example, an unsupervised algorithm can be used to group the risk profiles of multi-asset investments in order to identify whether certain types of assets share similar risk profiles with other assets. And in the field of investing, it's very useful for understanding patterns based purely on data. Another use case for machine learning is to help make better predictions. For example, neural networks, particularly long short-term memory models (LSTMs), are often used to investigate relationships that are not linear over time. The aim is to improve past modeling techniques, which have tended to focus (primarily) on linear relationships or combinations of linear relationships. Being able to accurately model non-linear relationships is one of the reasons why our clients are increasingly interested in how artificial intelligence can be used in investing.

Apart from the examples mentioned above, another use case for artificial intelligence applied to investing is so-called sentiment analysis, which uses natural language processing (NLP) techniques to study the tone of written text. is. Using Federal Reserve Board minutes as an example, the central bank's overall tone can signal an impending central bank policy decision, and investors should adjust their tactical positions accordingly. may decide. Initially, the method used to perform sentiment analysis was based on lexical analysis, where words were classified as positive or negative in a static list, and overall sentiment was aggregated based on these classifications. . Although this approach was simple, it often failed to capture linguistic nuances such as context and irony.

As NLP technology has evolved, it has become more common to utilize statistical methods to learn emotions from large datasets of labeled examples. These models were able to understand more complex language patterns, but still struggled with language context and subtleties. Recently, the introduction of deep learning and neural networks has made significant advances in the analysis of written text. These models can process sequences of text to capture long-term dependencies and nuances, greatly improving understanding of context. Analyze sentiment in relation to the surrounding text, allowing for more accurate interpretation and sentiment. Common to all the above analysis approaches, the text needs to be preprocessed before analyzing sentiment. In other words, it must be broken down or tokenized into smaller units, and words must be simplified to their basic form by lemmatization or stemming. For example, for the word “danger,” the base word would be “danger.” ”. After these steps, the text is converted into a numerical format that is easier for machine learning models to process and then fed into a sentiment analysis model for analysis and prediction.

A common problem with machine learning is statistical overfitting. Here, the model learns the details and noise of the training data. This data gives good performance, but new, unverified data gives poor performance. This essentially means that the model has “remembered” the training data, but has not learned to generalize from it, which limits its predictive ability. Therefore, it is important that machine learning models can be generalized. Otherwise, the predictions made by the model will not provide useful analysis or results.

Previously, we presented an example of how useful information can be generated from official central bank texts to understand impending central bank actions through sentiment analysis. Another use case for machine learning involves considering how to achieve better diversification within a portfolio. Some investors believe that diversification is best achieved by using as many components as possible, such as funds, i.e., US stocks, bonds, etc. Traditionally, the idea is that you can achieve a balanced portfolio by assigning equal weights to uncorrelated assets such as stocks and bonds in a 50/50 portfolio.

From a risk perspective, a 50-50 portfolio is unbalanced because equity risk is much greater than bond risk, even though the portfolio assigns equal weights to both asset classes. yeah. In addition to this, research has shown that investment assets often exhibit a hierarchical structure. This means that the price movements and returns of investment assets often exhibit patterns of similarity that can be grouped into clusters or hierarchies. This hierarchical structure reflects the underlying correlations and common economic or market factors that affect these assets. One way he uses this observation to improve diversification is to use hierarchical risk parity in his portfolio. Simply put, different assets are grouped into clusters based on how similar they are to each other, and the risks within the portfolio are spread across different clusters to achieve proper diversification of the portfolio.

Artificial intelligence and machine learning models learn from past data, so they can struggle with unexpected surprises. If a pattern is not observed in the data used to train the model, the system may not recognize it or know how to respond appropriately. This limitation is especially noticeable in rapidly changing environments and situations like COVID-19, where some clients' models did not perform properly. Another limitation of AI is that it depends on the quality and quantity of data used for training. These systems can introduce bias and inaccuracy if the data is not representative, diverse, or large enough. This can lead to biased or unfair results. In fact, artificial intelligence and machine learning systems are only as good as the data they are trained from.

First, it's important to define what generative AI is. Generative AI is a field of artificial intelligence focused on creating new content, from text and images to data and code, based on learning from vast datasets. In investing, it can be used to test scenarios by generating simulations of diverse economies or financial markets under different conditions, helping investors understand potential outcomes and make informed decisions. Masu.

Generative AI, especially through generative adversarial networks called GANs, can greatly enhance investment scenario testing by creating realistic financial market simulations. GANs work by pitting two neural networks against each other. One generates synthetic data that resembles real market conditions, and the other tries to differentiate between real and synthetic data. Over time, the generator becomes adept at producing highly realistic scenarios. This capability allows investors to test investment strategies and evaluate risk management under a wide range of market scenarios that may not have been observed historically but may occur in the future. , thereby providing a robust framework for decision-making under uncertainty.

I think people will increasingly use different forms of alternative data. Some algorithmic trading firms already use satellite imagery to see how much carbon is emitted by factories and match these numbers with what companies report. Satellite imagery is also used in other areas, such as tracking global cargo trade and analyzing global trade flows. This type of data is invaluable in estimating how much profit companies in a particular industry are likely to make and how these stocks will perform in the stock market.