Couldn't be happier to announce that I have survived the job market and will be joining
@ChicagoBooth
as a finance AP this summer!
It was amazing to meet so many awesome people on the market and many thanks to the fantastic support from everyone at
@YaleSOM
!
Thanks Arpit! What better time to post about my AI-related JMP, "The Ghost in the Machine: Generating Beliefs with Large Language Models", a 🧵:
🤖🧠🤖🧠🤖🧠🤖
This sentiment series captures the over/under enthusiasm we’d want from a measure of systematic errors: it correlates strongly with Shiller’s PE ratio and negatively predicts future returns.
How do I generate expectations? I provide an LLM with news articles and ask what it expects will happen to various economic outcomes (S&P, indpro, CPI...) based on each piece of news. I then aggregate these article-level responses to form a time series of generated beliefs.
So what can we do with generated expectations? For one thing, greatly expand the sample of expectations data – I use news text back to 1900 to build expectations for a broad set of variables and then extract a measure of the systematic errors or “sentiment” over that sample:
Behavioral theories are difficult to evaluate in asset pricing because we don’t directly observe investor beliefs. Nowhere does this matter more than for theories of bubbles. Can we use generated expectation to help? Yes!
LLMs implicitly react to information in a similar way to humans (including all our biases and errors). If I could provide an LLM with a textual representation of the state of the world at a point in time I could use that LLM to generate “human-like” expectations.
Importantly, generated expectations also exhibit many of the same deviations from full-information rational expectations that have been documented in the literature. Some summaries of those deviations for return expectations:
Expectations are central to econ, but aren't directly observable. Rational expectations (brilliantly) solves this but we have mountains of evidence against RE.
Without RE we're in a "wilderness of alternative expectations" - I propose a new method to navigate this wilderness…
@cameron_pfiffer
@shoshievass
@yizhoujin
This is sometimes done with Gibbs sampling for latent Dirichlet allocation -- basically do updates in parallel on each chunk, then recombine each step -- works quite well and can scale to very large datasets.
I find that an asset’s exposure to sentiment can be used to predict bubbles ex-ante -- heavily exposed assets that experience a run-up are much more likely to crash and experience negative future returns.
How “human-like” are generated expectations? To test this, I compare my generated expectations to various surveys. Here’s the generated expectations for the S&P 500 against two commonly used return surveys:
This strong connection with existing surveys holds for a wide range of variables. Here are correlations between generated expectations and the survey of professional forecasters for a number of macro series: