The asset management industry is undergoing fundamental changes, including job cuts, the rise of alternative assets and regulatory pressure for increased transparency amid battles about the significance of Environmental, Social and Governance (ESG) standards. Inextricably connected to the industry’s evolution is the rise of generative AI.
Generative AI describes the use of AI technology to generate output, including text, images, videos and more. When successfully deployed, generative AI can expedite key asset management processes – saving costs and trimming the administrative fat off the daily schedule of any asset manager.
Leading financial consultancies like EY, KPMG and the Boston Consulting Group offer a wealth of high-quality resources on the impact of generative AI in asset management. Yet, these resources lack in-depth insight into the direction the AI arms race in asset management is taking – particularly from the AI perspective.
In this article, we’ll take a technical approach to how contemporary advances in generative AI have been funnelled into sophisticated asset management technologies. Though the future role of generative AI in asset management is not fully clear, it presents several exciting possibilities.
A viral article by the MIT Technology Review describes AI as ‘A catchall term for a set of technologies that make computers do things that are thought to require intelligence when done by people.’
AI for asset management can be broken down into various technological techniques to optimise and expedite key asset management processes. Let’s explore four of them.
Cluster analysis is a statistical machine learning (ML) technique that can group objects with similar characteristics. When ML algorithms are specifically trained, they can group assets based on risk and return characteristics, such as volatility and correlation. Asset managers can diversify portfolios to reduce risk and probe the relationship between risk and return, ultimately refining their investment strategies.
A research study from Princeton University explores cluster analysis in portfolio management, noting several limitations, though. For example, the clustering algorithms may focus on singular characteristics within each cluster, leading to a lack of diversification (within individual clusters). Cluster analysis is best implemented into a broader ML analysis framework that can visualise its limitations.
Decision trees map complex and non-linear relationships between attributes, facilitating classification and helping with forecasting. For instance, a decision tree may map factors like market and (or) macroeconomic indicators, industry trends and company financials to determine the risk of a potential investment.
Evolution AI’s article about decision trees for VentureBeat notes that generative AI-powered decision trees are becoming more accurate at replicating human cognition when analysing data and making recommendations. However, a decision tree’s inherent interpretability is counterbalanced by its limitations – its discretisation of variables struggles to represent continuous variables. Though decision trees can help asset managers stay risk-focused, users must combine them in an ML model that can predict outcomes using other techniques.
As its name suggests, genetic modelling operates like Darwinian natural selection. AI-powered genetic algorithms (GAs) evaluate and recommend the highest-performing portfolios under certain criteria (such as risk tolerance and investment horizon). The main benefit of genetic modelling is that GAs can address portfolio optimisation problems with complex constraints or conditions. In other words, the types of contexts that classic optimisation algorithms (like simple decision tree algorithms) would struggle to solve satisfactorily.
Here’s an open-source project on GitHub that illustrates genetic modelling.
Heuristic optimisation techniques (like genetic modelling) hold promise for asset management technologies. While fine-tuning parameters to achieve optimal results can be time-consuming, future advancements in metaheuristics may expedite this process.
NLP is one of the most well-known subsets of AI — and one of the most versatile. NLP algorithms mimic how people comprehend and use language. Technologies like large language models (LLMs) – where NLP is foundational – are useful for asset managers looking to analyse stock data or financial news articles. You might ask a generative AI model to analyse Tweets mentioning companies within an investment portfolio and to generate a sentiment analysis report. The report would be important when assessing brand reputation and potential risks.
NLP technologies can struggle with lexical ambiguities, such as sarcasm or context, which will likely be present in financial opinion pieces or Tweets. A more reliable application of NLP for asset managers might be to upload an annual report to an AI model and then download the extracted and calculated financial data.
The following are a few prominent examples of techniques that generative AI-powered asset management technologies might leverage. But what do these technologies look like? Let’s take a closer look.
Leading analysis and modelling platforms combine real-time and historical data to model portfolios and examine risk closely. These platforms will often use all the above techniques combined with standard financial techniques like time series analysis and regression analysis to analyse the various outcomes. These technologies then generate strategies for portfolio optimisation and investment.
Companies preferring a lighter generative AI commitment may choose to automate individual processes (or use these technologies to enhance the abilities of their analysis and modelling platforms). Examples of these processes include doing the following from various sources:
Again, generative AI-powered automation will combine statistical AI techniques to automate tasks that replace manual labour.
For example, our tool, Financial Statements AI, extracts, analyses and exports data from financial statements. Evolution MDM also combines data from groups and tables and outputs comprehensive files that standardise the required data into an organised and accessible format. The validated, standardised data is designed to save asset managers time and effort when conducting portfolio analysis, risk assessment and more.
LLMs like ChatGPT may seem convenient for simple analyses due to their ease of use and cost-effectiveness. For instance, you could upload an annual report and ask the LLM to list its key financial metrics or break it down by business segment.
We’ve tested and written about ChatGPT extensively. A key finding is that ChatGPT tends to generate one error or hallucination (i.e. fictitious information) per page. For time-strapped asset managers, ChatGPT’s erratic performance could result in spending more time correcting its output than benefitting from its speed.
Excess training data can be too much of a good thing if the input is too great or if the user trains the model for too long. AI-powered asset management technologies can simply generate the training data rather than learn to understand the relationship between the data variables in the training data.
Overfitting can be difficult to identify and prove catastrophic in asset management contexts (if left uncorrected). For example, a model trained on historical stock data might learn specific patterns. However, it may blindly apply these historical patterns if tasked with predicting prices in entirely new market conditions, leading to poor investment recommendations.
Integrating new and evolving AI technologies into company workflows may present a challenge. For instance, companies with legacy systems or talent gaps may struggle with the practicalities of implementation. Alternatively, companies looking to train generative AI technologies with their own data may lack the time to create high-quality synthetic data (or clean their own data), especially the mass volumes required for successful training.
In financial contexts, explainability is key. Suppose generative AI suggests investing in a certain alternative asset at a particular time. In that case, you cannot consider it a robust suggestion until an asset manager can vouch for its calculations.
In the long term, AI’s opacity can compromise its effectiveness. AI development can even become a black box, where its operations are inaccessible to external scrutiny. Any biases in the AI’s training data can also manifest as hard-to-spot distortions in the AI’s output. Down the line, small distortions can manifest as poor-quality forecasting or incorrect investment recommendations.
Developers and asset managers must work together to maintain social responsibility – operating explainable and auditable generative AI for asset management going forward.
Asset management is a playground for generative AI. Generative AI’s ability to quickly generate insights makes it a guiding light when dealing with the current uncertainties surrounding the asset management industry. However, asset managers wishing to use generative AI tools should prepare to carefully navigate the nuances of using a new and continuously evolving technology.
Evolution AI has worked with asset managers like Unigestion to automate key processes like data extraction and aggregation. To learn more, book a call with one of our financial data team members or email hello@evolution.ai. You can also read our case studies here.
We write extensively about the applications of generative AI to financial services – follow us on LinkedIn and X to stay up-to-date.
