Jev earnings-call pre-registration

Abstract

Purpose Large language models are increasingly used to read corporate disclosures, and several studies report that an LLM's reading of earnings-call text predicts subsequent stock returns. Most of these studies date the text at the call itself, although transcripts become machine-readable only later, and they cannot rule out that a commercial model has seen the outcomes in its training data. This project asks a narrower, practical question: when an LLM reads the prepared remarks of a U.S. earnings call, is its reading associated with the next-session abnormal return that an investor could actually trade, given when the transcript becomes available? Design The reader is TypeSafe Jev (model pinned to jev-1.13.0). It returns a typed score of how favourable the news is relative to what the market already expected. The unit is the call, not the transcript version. Each call is dated by the creation time of its first transcript version, and entry is at the first market close that the transcript can reach. The outcome is the market-adjusted CRSP return on the next trading day. The signal is a point-in-time percentile of the score within size bucket. Controls are the day-0 return, momentum, and the IBES EPS and revenue surprises, with entry-date fixed effects and standard errors clustered by date and firm. The specification was frozen on a discovery sample of about 105,000 previously scored calls. It will be tested once on two samples no model had seen at registration: about 24,000 never-scored calls from 2011-2024 and about 8,600 calls from 2025. Pre-registered tests - H1 and H2: the association replicates on the never-scored and 2025 calls. - H3: net of half the quoted closing spread, the small-cap long-short spread is positive. - H4: a joint test of whether transcript latency limits implementability without selecting the effect. - H5: an open-weight reader with a documented training cutoff (Llama-3.1-8B-Instruct) predicts returns after that cutoff. - H6: a revenue-recall probe to date Jev's knowledge cutoff. - H7: the association is unchanged when firm and executive names, tickers and dates are removed from the text. Multiplicity is controlled by Holm's procedure over H1, H2, H3 and H5. Expected outcomes Calibration on the discovery sample gives about 15 bps per standard deviation of the score. The minimum detectable effect is about 8 bps for H1 and 12.5 bps for H2. On the discovery sample, the latency hypothesis H4 is not supported, and its failure is anticipated in the registration. Every test is reported whatever its outcome, using pass/fail wording written in advance, and any deviation is logged. Latency is measured for every scoring call and reported. Data The data are Capital IQ transcripts, CRSP and IBES. They are licensed and are not shared here; all frozen files are identified by their SHA-256 hashes in MANIFEST.sha256.