From Earnings Calls to Earnings Events: An LLM-Guided Separation of Discretionary and Systematic Volatility Trading Signals

Abstract

This paper studies whether earnings-call trading signals can be separated into discretionary components that are firm specific and systematic components that proxy market-wide state. The empirical design is rebuilt on public data releases dated 2022 or earlier in the underlying academic record and 2023 or earlier in the accompanying market-data releases. The call-level text comes from ECTSum, an earnings-call transcript benchmark released in 2022, and the market variables come from S&P 500 news/sentiment and options-implied-volatility panels. The matched sample contains 563 earnings-call events for 254 firms from 2020-10-22 to 2023-01-03. A deterministic ontology assigns every engineered variable to either a systematic or discretionary family before estimation. The outcome is the direction of abnormal at-the-money implied-volatility adjustment around the call, measured relative to the cross-sectional market option-implied-volatility benchmark. Rolling chronological evaluation compares logistic regression, random forests, and gradient boosting. For event-day abnormal IV direction, the best model is a combined pre-event logistic specification with an out-of-sample AUC of 0.579, compared with 0.565 for the discretionary-pre block and 0.499 for the systematic block. For next-day continuation, the best specification is a discretionary-plus-reaction logistic model with an AUC of 0.556. Sorting out-of-sample next-day probabilities into quintiles produces a top-minus-bottom abnormal-IV spread of 0.552 points with a Welch t-statistic of 2.47. The evidence is statistically modest but economically ordered. It supports a layered interpretation: earnings-call language and issuer-specific option state carry most of the signal, while systematic market conditions mainly condition the transmission of that signal.