NLP can help identify linkages between equity market peers

Artificial intelligence (AI) and machine learning (ML) models have shown promise in improving forecasting accuracy across a wide range of problems, including image and speech recognition, fraud detection, medical diagnosis, traffic pattern detection and text translation. 

ML forecasting models tend to be most effective in applications where there is a high signal-to-noise ratio in the data. Caution is in order, however, when applying these ML methods to the task of forecasting returns, as financial data often exhibit notable amounts of noise. Although practitioners have shown that ML models are able to compete with traditional estimates of risk premia, the problem of noise in asset returns remains a challenge for both ML and traditional methods.

Natural language processing (NLP) is a subset of ML that learns the information contained in a document by mapping the words into a numerical form that the computer can consume. By using NLP techniques, the researcher can be confident that the data set – a corpus of documents – is likely to contain useful information and exhibit a high signal-to-noise ratio. 

Moreover, NLP methods offer access to the unstructured data in the document without resorting to an army of analysts. NLP techniques not only offer a way to gain insights from unstructured text data but also provide a way to do so at scale, thereby facilitating access to useful information across a broad set of investment opportunities.

Economically-linked peers

At LA Capital, NLP techniques have been used in factor construction since 2018, and a fruitful area of research has been the analysis of earnings call transcripts. Quarterly earnings calls contain valuable information about the drivers of a company’s revenues, its market positioning, and risks, along with comments that may be useful for developing return expectations. 

Appropriate NLP techniques can convert the transcripts to a numerical vector that can be used to analyse the information contained within the text. In particular, the distance between a pair of vectors represents the similarity of language used by different companies and provides indications of whether firms are discussing the same challenges, market themes and competitive landscapes. In other words, whether two companies can be considered peers.

The NLP analysis of transcripts can reveal information about firms that might otherwise be missed by simple notions of similarity. For example, an electric car maker’s earnings may look similar to other vehicle manufacturers because they both make cars. But an analysis of transcripts may also associate the electric car company with battery producers – a relevant link as broad concerns over battery production may impact the electric car producer more than demand for cars. 

The group of peers for a parent company can also offer advantages over traditional classifications such as industry. The business operations of many companies can be complex, and the NLP-based approach can extract information from earnings calls that span multiple categories. 

Consider Amazon, a complex business spanning multiple markets and sectors, from retail sales of discretionary and durable goods to communication services offerings (Amazon Price streaming content) and information technology (AWS Cloud Services).

Application of LA Capital’s peer methodology reveals that Amazon’s top 10 peers span all of these traditional classifications such that the peer group becomes representative of the economic exposures of the parent company. Moreover, as new information is revealed with each quarterly earnings call, the peer groups can be updated and respond to the evolving nature of a company’s business in a way that will likely not be matched by static classification schemes. 

While NLP analysis of documentation can provide novel links between companies, managers should be careful not to disregard customary measures of economic linkage. While ML models can find non-obvious relationships between companies, they can also be misled by noise in the data. Interestingly, a simple but effective way to avoid spurious relationships between companies is to combine ML models with traditional measures of similarity. 

A recent example illustrates the advantages of this approach. Peers of Vestas Wind, which designs and manufactures wind turbines for energy production, were assigned 0 to 1 scores of similarity to the parent firm inferred from three NLP analyses of earnings transcripts and from four proxies of economic similarity: analyst co-coverage, industry group classification, risk exposures and ESG characteristics. 

Different NLP approaches 

The three NLP methods take different approaches to uncovering similarity of transcripts. The first finds similar companies based on common occurrence of important words in transcripts, the second refines this approach by using a neural network to learn the semantic context of transcripts to identify similar companies. The third aims to learn common topics of discussion from the corpus of all transcripts and defines peers as companies discussing the same themes.

The company’s top peers identified by the three NLP methods are involved in the same industry and show similarity according to all three models, share analysts in common and exhibit similar risk exposures and ESG characteristics. Interestingly, however, the NLP models find other companies with interests in sustainable technology and energy markets, such as Wärtsilä Oyj, despite low analyst co-coverage and differences in industry classification. 

Benefits of multiple methods 

However, the benefits of using multiple methods are clear. The lack of analyst co-coverage and no overlap on the themes discussed in the earnings call implies, consistent with intuition, that, for instance, Airbus is a poor peer candidate for Vestas Wind. However, low-similarity peers such as Airbus score highly on an NLP method that looks for important words – it seems plausible that ‘turbine’ would be significant to both companies, for example. 

The use of traditional approaches to economic linkage as guardrails to the output of ML models offers a way of selecting ‘high conviction’ peers for a parent company. Such peer groups can be useful tools for stock selection and risk management. As the stock returns of economically similar companies should be driven by common components, the positive momentum of peers should be a good signal for the parent company. 

The relative valuation of a company to its peers may reveal instances of asset mispricing, while peer earnings revisions may also be predictive of future changes in earnings of the parent, especially if analysts are yet to update their views on the company in question. A company and its peer group should exhibit the same uncertainty in outcome such that the peer group may act as a natural hedge to the future risks of the parent company. 

Importantly, the use of NLP techniques allows for the systematic construction of unique peer groups that can be performed at scale. As a result, unique peer groups and associated insights can be inferred not just for a handful of assets but for a broad global opportunity set.

Hal Reynolds and Edward Rackham are co-CIOs at Los Angeles Capital Management