I’m excited to further study questions related to finance, technology+innovation, and behavioral economics, and to extend the scope and credibility of machine learning in empirical research . You can find the full paper at

In summary:. I transform economic language into embedding vectors, and show these vectors are informative of perceptions and beliefs. I train LLMs that address credibility issues with ML in empirical research. I study economic mechanisms that drive valuation and misvaluation

Valuations can fluctuate when perceptions change—is a customer service firm that adopts a language model really an “AI” firm, or just a “service” firm?. I find that these perception changes relate to selective attention, firm communication, and technology transformations

Main results:.1️⃣ Embeddings explain valuations + outperform traditional characteristics .2️⃣ Returns reflect changes in how businesses are valued + changes in the perceived business model itself.3️⃣ Some changes in embeddings reflect misperceptions, which generate misvaluation

Embeddings also encode similarity between firms—geometric distance relates to established measures of perceived similarity. Taken together, these results demonstrate that a firm’s embedding is informative of its perceived business model

First, I train new language models on historical data to avoid lookahead bias. I’ve released the LLMs to the research community. Second, I use contrastive representation learning to construct embeddings of firms. The geometry of these vectors relates to economic features of firms

I transform financial news language into embedding vectors. Embeddings put quantitative structure on unstructured data, and have contributed to the success of machine learning over the past decade

Economic valuations fluctuate in ways empirical research cannot fully explain. What information are we missing? Economic theories emphasize the role of hard-to-quantify beliefs and perceptions. My job market paper develops algorithms + measurement to quantify perceptions of firms

You can find the working paper at: Thank you to @keyonV for being an amazing coauthor. Comments are very welcome! (n/n)

The ideal solution is to select from a family of LLMs with rolling cutoffs, or “time subscripts”. For example, one such model family is StoriesLM: This is only a start—there is a lot of room to further research and improve these kinds of models (8/n)

Instead, we argue researchers should use language models with historical pretraining cutoffs. This way, analysis that uses these models is out of sample (7/n)

How can we address lookahead bias? Two ad hoc strategies have issues. [Prompting] We include instructions not to look ahead in our main results. There is still lookahead bias. [Masking] Censoring identifying information from prompts doesn’t make the prompts unidentifiable (6/n)

We next show that language models can be used to predict the results of “natural experiments” that are commonly considered unpredictable. We show a language model-based analysis procedure can predict the results of close U.S. House elections (5/n)

We argue a model can also leak future information indirectly. Generated risk factors for 2020 are more likely to include sequences like “disease outbreak” and “supply chain” than generated risk factors for 2019 (4/n)

We argue if a model systematically generates language sequences from the future, it directly leaks future information. We scale this exercise to 1,000 earnings calls from September 2019 – November 2019. We find 6.8% of generated risk factors for 2020 include “COVID-19” (3/n)

An example: We query a language model with Zoom’s earnings call from September 2019, with instructions to generate the firm’s risk factors. The output includes “COVID-19,” a term introduced in 2020. It also mentions “remote work,” which became much more prominent in 2020 (2/n)

Language models’ pretraining data includes information about historical events. When we use LLMs to make economic predictions, do they leak information about the future?. @keyonV and I develop tests that find evidence of lookahead bias in LLMs, and identify ways forward (1/n)

I cannot thank my collaborators enough for their ever-present energy, creativity, and insight throughout this process. Our dataset and example analyses: Our paper: Please, take a look and let us know what you think!. [n/n].

Our dataset has numerous limitations. It is restricted to English-language text, and focuses on patent applications, which are not universally accessible and do not reflect all kinds of innovation. We discuss inequities in patent access and outcomes in our paper [9/n]

The answer may lie in differences in innovation criteria across patent categories, as well as heterogeneity in technical language. Our dataset provides rich metadata on inventors and applications that can be merged with other sources to study the drivers of these gaps [8/n].