Thematic investing

Using AI for early discovery of novel investment themes from company filings

Introduction

Do you regret not getting on the AI investment theme sooner? Are you never quite sure if something really is a theme?

You're not alone, it's a recurring theme when talking to wealth managers.

Quantitative methods hold the potential to enhance thematic investing significantly.

For one, they can aid in formulating thematic hypotheses. Humans often face constraints in mental and physical resources, which can hinder the identification of investment themes. However, quantitative algorithms are not subject to these limitations. They can swiftly analyze extensive data sets without fatigue, enabling them to detect novel and evolving themes —even those that might elude human investors or be noticed too late.

Moreover, unlike humans, algorithms are immune to emotional ties with investment themes or specific stocks within those themes. Emotional investments can cloud human judgment, leading to a loss of impartiality. In contrast, algorithms operate solely based on their programming, allowing for an objective assessment of investments. This objectivity could also facilitate timely exits from investment themes when the supporting data no longer warrants the investment, potentially optimizing investment outcomes.

methodology

01. filings

We passed 10-K filings to an LLM, asking what the main company activities are.

As an example, for Tesla it returns "Electric vehicle manufacturing", "Autonomous vehicles" & "EV battery technology".

02. Processing

The returned activities are processed to weed out filler words such as "global", which can interfere with the clustering step.

03. clustering

The activities are embedded and then clustered. This will group terms which are semantically close to each other, such as "electric vehicle manufacturing" and "electric vehicle manufacturer".

04. dynamic analysis

We check for large changes in clusters to extract new themes. Sensibility checks are performed on the size (both in number and market cap), novelty of the terms etc.

05. Constituents

From the clusters we already have the main constituents of a theme. One more step is performed: searching for highly similar embeddings of constituents within the universe. This will ensure that companies with business descriptions that largely coincide are added to the theme as well.

results

Using the above methodology, around 8 themes are discovered every year. Of note is the AI theme, which is observed already mid-2021, well before ChatGPT burst onto the scene. But there are also smaller, lesser known themes that come up. For instance, this year fiber optics bubbles up as an interesting theme.

The timeline below shows a selection of discovered themes since 2019.

We also performed a very simple markout analysis on the themes as an illustration their potential. For each theme, the constituents are market cap weighted, and the returns against the Russell 3000 are shown in the graph below. A stop-loss is set at -25%. The resulting returns and their average are shown in the graph below.

As can be seen, themes return on average 8% above the broad market in the year after formation, with a steady growth rate. In fact, on average the first months return very little; an indication that these themes are picked up ahead of the broader market.

The lists are incredibly useful to flag novel themes early, and provide inspiration on which companies belong to it. It's perfect for wealth managers wishing to anticipate new trends. The possibilities are endless for inclusion into your workflow.

Do reach out to us if you would like to learn more through the button below.

Further work

We're currently working on the following:

Want to keep up to date on current investment themes? Let us help you.

Get in Touch