Scott Cunningham, Baylor University
ChatGPT in the Economist’s Toolkit: An Early Exploration — ‘Prediction’
Large language models have the promise to disrupt life through their ability to mimic intelligent speech. This mimicry is driven through its predictive capacity. But what is not known is whether that prediction capacity extends to the future. A unique feature of OpenAI’s ChatGPT 3.5 and 4 is that they were trained on data through September 2022 which allows us the ability to explore their predictive ability for benign to somewhat more complex economic events. We overcome some of the LLM’s built-in refusal to make outright predictions by having it, instead, reenact a scene in which certain events have already happened. Thus instead of predicting the future, ChatGPT believes it is the future and is recounting the past. The results are interesting. We find that it does extraordinarily well at predicting the winner of several Academy Award categories, but fails to predict the Best Picture. When asked to perform as Fed Chair Jerome Powell and restate the time path of inflation and unemployment under a variety of shocks (like Russia’s invasion of Ukraine), it performs somewhat less consistently. We suggest that this may be because LLMs are unable to incorporate endogenous policy responses by the Federal Reserve, and thus the LLM prediction may suffer from basic Lucas Critique problems.
The Tuesday Talks Series is scheduled throughout each semester with speakers invited to discuss cutting-edge topics. Occasionally talks are given on subjects relevant to students such as career management and research. CGU and Claremont Colleges Consortium faculty, students, staff, and alumni are encouraged to attend.
Unless otherwise announced, all seminars are scheduled from 12:00 pm – 1:00 pm. The format for the talks this semester is hybrid, with the option to attend in person at Burkle 12 or online via Zoom. Please use the Zoom links below to register for that talk. For those attending in person, a light lunch will be provided!
To submit an event, please use the event submission form. NOTE: Non-CGU event submissions will be deleted.