Filings, Insiders & OwnershipCore researchInvalidNew

Transcript AI Exposure

Updated dailyData needs: mediumlong onlyshort onlylong short

2023

Eisfeldt, Schubert & Zhang (2023), "Generative AI and Firm Values", NBER Working Paper w31222

In plain terms

Counts AI/LLM phrases in each earnings call and trades firms whose AI talk jumps (long) or fades (short) versus their own history. Inspired by the finding that AI-exposed firms outperformed sharply after ChatGPT launched, but built from call transcripts rather than the paper's workforce data.

How it works

Eisfeldt-Schubert-Zhang (2023) NBER w31222 "Generative AI and Firm Values" builds a firm-level GENERATIVE-AI WORKFORCE-EXPOSURE characteristic (occupation-level generative-AI exposure mapped onto firm workforce composition; not a transcript measure) and documents a cross-sectional Artificial-Minus-Human long/short earning ~0.4%/day (~5% cumulative) in the ~2 weeks after ChatGPT's Nov-2022 release. THIS FAMILY IS AN INTENTIONAL PROXY, not a replication: it scores earnings-call transcripts against an AI/LLM bigram lexicon HHVT-style (Hassan, Hollander, van Lent & Tahoun 2019, QJE) and trades the per-firm own-history z of AI-bigram density (expanding, min_periods=4, prior-call-only): long when z >= +1 (firm just elevated its AI narrative), short when z <= -1 (firm de-emphasized AI; the paper instead shorts low-workforce-exposure firms), with 21/63/126-day holds rather than a cross-sectional event-window sort.

Live results

0 times picked on its own · 12 times inside a blend (12 beat the stock) · updated 2026-07-06

This strategy is a frequent ingredient in blends that combine a few strategies on one stock. It has contributed to 12 such blended picks (12 of which beat simply holding the stock). Picking it on its own is only one of the ways it shows up.

How its picks scored vs. buy & hold

Each pick is graded on a recent year it was never tuned on, against simply owning the same stock

Where its edge concentrates

Share of picks in each company-size group that beat buy & hold

How often it trades

Active vs. patient. Bars on the left mean it waits for rare setups; bars on the right mean it trades often

Return vs. buy & hold

How much each pick beat or trailed simply owning the stock over the test year (extreme microcap moves trimmed)

Loading substrate evidence…

Data dependencies

Daily prices
Adjusted-close OHLCV for every US-listed ticker; primary price feed.
Earnings call transcripts
Full earnings-call transcripts (prepared + Q&A), tokenised.

Expected edge

Reported return: ~0.4%/day (~5% cumulative) Artificial-Minus-Human long/short in the ~2 weeks after ChatGPT's Nov-2022 release
Tested over: Event window around ChatGPT release (Nov 30, 2022), ~2 weeks post-release

Paper documents an Artificial-Minus-Human cross-sectional L/S of ~+5% in the ~2 weeks after ChatGPT's release (~0.4%/day). This family is a transcript-based proxy of the same AI-exposure theme, so realized edge is expected to differ from the paper's event-window result.

Explore Transcript AI Exposure on alphactor.ai

See which tickers this family is currently firing on, with live signals and rankings.

Open universe with this filter →Browse all 642 families