3 min read

What are Material Relationships and how is Noonum differentiating with them?

What are Material Relationships and how is Noonum differentiating with them?
Photo by Alina Grubnyak on Unsplash

Materiality is key: in finance and investing, materiality is a concept that is unique to each investment product, but represents the relationship between a security and information that an investor would use to perform a current or future valuation of the security.  A good historical view of materiality can be found here: https://corpgov.law.harvard.edu/2021/05/01/corporate-governance-update-materiality-in-america-and-abroad/  

The SEC uses the concept of materiality to indicate that a company that issues a security, or investment product, must disclose any information that an investor (or potential investor) would have liked to know before making or while holding the investment.  Many things are universally material to an investor: has the CEO been recently fired, is the company being sued, or does the company have operations in a locality where a natural disaster recently occurred.  Other things are material to whole industries: same store sales in retail, environmental spillages in heavy industries, or global pandemics in the travel and tourism sectors.  Companies can also have very specific or unique material concerns that an investor would need to know: has the company just been approached for a merger, is the company facing labor or union strikes or renegotiating contracts, or has a major competitor just gone bankrupt.  The SEC regularly hands down fines related to investors suing companies for failing to disclose material issues.  An example is the company C3.AI was facing a lawsuit brought by investors that suggest the company did not disclose accurate and material information before their IPO, which would have changed how investors valued the stock.

Because materiality is a broad and contextual concept, there is not a concrete set of data that one can source that covers its breadth.  Additionally, as you can see, when one talks about material issues, the language can be very diverse depending on the speaker: a company may be facing a lawsuit, a class-action suit, a legal issue, being served papers by an attorney, or facing a criminal trial or fine by the SEC.  All of those phrases essentially mean the same material concept of ‘legal issue’ is present.  If you saw ‘legal issue’ as a negative issue for a company, then using the ‘strength’ of the material relationship in an investment strategy can represent a defensible way of buying or selling stocks.

Capturing materiality by NLP methods is difficult.  In corporate filing documents, companies are required to discuss any risks they are facing.  A naïve approach would be to focus on these sections and extract the common noun phrases.  However, you would realize that companies typically describe all types of general types of risks they might face, not specific or timely risks.  For example, once the COVID-19 pandemic took hold, nearly all companies began putting a boiler plate statement referring to the fact that their business overall faced risks due to COVID-19.  Additionally, companies that have international business will usually describe the risks by conducting business with foreign currencies and exchange rate fluctuations.  If an NLP scientist was looking for a large set of data that they could process using an unsupervised ML technique to capture ‘materiality’, they would be disappointed.  Instead, hard data science work, manual labeling, innovation on sourcing texts to label and further quality control and cleaning is required.

Capturing Material Relationships with Specialization

In Noonum, we use an array of customized data preprocessing and NLP models to identify the concepts found in texts as well as the entities, like a company name, and link them to a knowledge graph of known things.  We then apply custom, machine learning trained classifiers to extract the relationships of concepts to organization based on a material impact.  Material impact is understood by labeling thousands of examples coming from a representative group of sources and language styles, with subsequent label validation and quality control.  Our algorithms then use a separate data set for validation, seeking the most accurate and robust classifier that performs over different types of information as well as over different types of securities, from small to large cap, and across industries and markets.  The material classifier returns a predicted ‘strength’ or probability of material relevance, that is aggregated over many thousands of examples to create an exposure score.  Exposure here means the organization is ‘exposed’ to a theme, based on the aggregated predictions.   We measure the accuracy of those predictions through our custom labeled data sets with verifiable examples of material relationships in texts.   Lastly, to demonstrate the efficacy of our approach, we run various back testing using our material relationships to show that they capture not only superior performance but are also defensible in the names they pick and how they are used in different allocation strategies.