When Does "Post-Call" Begin? An Availability Contract for Earnings-Call NLP Benchmarks

Xiaoyu Ma; Kelvin J. L. Koa; Yuncong Liu; Ke-Wei Huang

q-fin.CPOct 15, 2026

When Does "Post-Call" Begin? An Availability Contract for Earnings-Call NLP Benchmarks

Authors: Xiaoyu Ma, Kelvin J. L. Koa, Yuncong Liu, Ke-Wei Huang

Organizations: National University of Singapore Asian Institute of Digital Finance Singapore

Abstract

Corporate financial disclosures often unfold over time, with announcements, prepared remarks, and discussion spanning multiple high-frequency return intervals. Financial NLP benchmarks can therefore suffer from a within-event timing mismatch when the prediction target begins before the textual or audio input used by a model could be complete. Chronological train--test splits do not prevent this mismatch, because it occurs inside individual events rather than across samples. We introduce an Availability Contract, which requires a benchmark to state the input boundary, evidence for input availability, the outcome start, and the claims supported by that evidence. We apply this framework to 552 earnings calls from 24 firms. For each call, we hold the return endpoint fixed and compare the previous close, a documentary call-start proxy, and a completion proxy based on reconstructed audio end plus a stipulated processing allowance. At this baseline, moving from the call-start proxy to the completion proxy flips return signs for 9.4% of calls. Holding completion-clock cutpoints fixed, 28.3% change quintile. In an independent release-session analysis of a public earnings-call volatility benchmark, shifting before-market target windows forward by one trading session raises three-day mean squared error by 18.1-21.4% across four frozen prediction streams, while their point rankings remain unchanged. Together, the results show that labels and losses can depend materially on the declared outcome clock even when model rankings appear stable. The Availability Contract makes this assumption explicit and ties post-input claims to the available timing evidence.

Figures & tables

Claim	Permitted interpretation
Event-window association	The return may span the event; interpretation is an event-window association, not post-input attribution.
Proxy-aligned association	A stated completion proxy precedes label start; interpretation is conditional on the proxy.
Verified post-input	Observed input completion, or a defensible upper bound, precedes label start.
Deployability	Point-in-time input, model, order, and execution clocks are observed or conservatively bounded, and costs are modeled.

Table 1. Claim ladder for availability evidence.

Figure 1. Availability contract and paired three-clock audit. Panel A maps source and stipulated completion proxies to non-executable price anchors; transcript and model readiness remain unobserved. Panel B constructs three common-endpoint returns and their removed intervals. Panel C crosses four retrospective representations with three targets in 12 common-sample OLS fits. Panel D reports eight adjacent-clock coefficient contrasts. These contrasts are associational, $r_{i}^{390}$ is a separate duration check, and the timeline is not to scale. Four panels show the proxy timeline, return decomposition, common-sample OLS audit, and adjacent-clock coefficient contrasts. The representations are computed retrospectively, and the price anchors are labels rather than execution prices.

Figure 2. Target-clock sensitivity. (A) On 552 paired calls, $S\!\rightarrow\!A$ changes 9.4% of raw-return signs, 28.3% of fixed- $A$ quintiles, and 30.1% of re-estimated relative quintiles. (B) A separate one-session perturbation of DEC before-market targets changes targets and four-stream MSE but not point rankings ( $\tau=1$ ). (C) Exploratory $LS$ and $SA$ coefficient components; $\dagger$ marks the previously selected Qwen construct. Panel A compares sign and quintile changes for the legacy to source-start, source-start to completion-proxy, and total clock adjustments. Panel B shows a separate one-session perturbation of DEC before-market targets: grouped relative and fixed-cutpoint quintile reassignment bars at three, seven, and fifteen days, with MSE changes below. Panel C shows the exploratory coefficient components removed over the first and second clock transitions for four representations.

Stage	Calls	Retained from prior	Retained from parent
Parent event panel	2,094	100.0%	100.0%
Usable source clock	755	36.1%	36.1%
Audio-duration proxy	669	88.6%	31.9%
Complete common sample	552	82.5%	26.4%

Table 2. Complete-case sample construction. Percentages are relative to the previous and parent panels.

Metric	$L\!\to\!S$	$S\!\to\!A$	$L\!\to\!A$	After-close
Calls	552	552	552	206
Spearman $\rho_{s}$	0.506	0.900	0.465	1.000
Mean $\|\Delta r\|$ (pp)	3.427	0.623	3.520	0.000
Median $\|\Delta r\|$ (pp)	2.661	0.179	2.587	0.000
Sign flips (%)	32.1	9.4	34.1	0.0
Relative moves (%)	64.3	30.1	68.5	6.3

Table 3. Distinct target-clock sensitivities.

S\!\rightarrow\!A

reassigns raw-return targets; separately, perturbing DEC before-market targets changes losses under frozen-prediction rescoring but not point rankings.

Dimension	Variation	Focal estimate or range	Design details
Allowance $\ell$	0–120 min	Sign: 6.5%–15.0%; relative: 23.9%–39.9%; fixed- $A_{15}$ : 21.0%–35.9%	Stipulated allowance scenarios.
Audio-duration $G$	MP3 / alignment	8 / 3 anchors move	Maximum coefficient change $<0.004$ pp/SD.
Endpoint $E$	Following close / 390 min	$-0.233$ / $-0.226$	Secondary BH $q=0.754$ – $0.964$ .
Dependence	Two-way CGM	$\Delta\beta=+1.355\;[+0.505,+2.204]$	24 firms, 175 weeks; asymptotic sensitivity.
Concentration	Delete one group	Firm: 1.149–1.456; year: 1.119–1.540	Delete-one-firm and delete-one-year refits.

Table 4. Descriptive design sensitivities.

Estimate	Lex. tone	FinBERT	Qwen val. †	Narr. rev.
$\beta_{L}$	+0.100	+0.196	+1.122	-0.579
$\beta_{S}$	-0.009	-0.084	-0.135	-0.017
$\beta_{A}$	-0.073	-0.082	-0.233	+0.134
$\beta_{L}-\beta_{S}$	+0.109	+0.281	+1.256	-0.561
$\beta_{S}-\beta_{A}$	+0.064	-0.003	+0.098	-0.151
$\beta_{L}-\beta_{A}$	+0.173	+0.278	+1.355	-0.712

Table 5. Exploratory association localization.

Field	This audit
Claim	Proxy-aligned three-clock target and association sensitivity.
Input boundary	Complete spoken Q&A.
Readiness $(G,\ell,M)$	Reconstructed audio-duration proxy plus 15 minutes; 0–120 minutes is the sensitivity grid. Transcript and model-inference completion are unobserved.
Model clock	Retrospective checkpoints; Qwen inference ran in June 2026 with a 2025-vintage checkpoint.
Label start	$L$ : previous close; $S$ : close of the first bar starting after the documentary source-start proxy; $A$ : the corresponding rule after the readiness proxy. These are label prices, not execution prices.
Endpoint	Following-session close; secondary endpoint after exactly 390 trading minutes.

Table 6. Availability Card for the earnings-call audit.

Look-Ahead Bias in Language Models

See all 8 papers

Jul 26, 2026

Earnings25: A Comprehensive 500-Hour Speech Benchmark for Finance

We introduce Earnings25, a finance-domain benchmark for evaluating automatic speech recognition (ASR) on English-language earnings calls under realistic conditions. Earnings25 comprises two complementary test sets: (i) testset-full, 498 hours of full English-language S&P 500 earnings calls from Q4 2025, and (ii) testset-segmented, a 46-hour industry-balanced set of 290 segments sampled from English-language U.S. earnings calls in 2025. The benchmark provides aligned transcripts and structured metadata, including speaker roles, industry labels, and call structure, enabling speaker- and industry-aware evaluation beyond aggregate word error rate (WER). We report reproducible baselines for Whisper and Parakeet-TDT using standardized scoring.

Denglin Jiang, Haoran Zhou, Anshul Wadhawan +7

Bloomberg, United States

ASR Evaluation Speech Recognition Benchmarks Automatic Speech Recognition Finance

Ask PDF

May 4, 2026

Effective Performance Measurement: Challenges and Opportunities in KPI Extraction from Earnings Calls

Earnings calls are a key source of financial information about public companies. However, extracting information from these calls is difficult. Unlike the templatic filings required by the U.S. Securities and Exchange Commission (SEC) to report a company's financial situation, earnings conference calls have no built-in labels, are unstructured, and feature conversational language. We explore this challenging domain by assessing the information captured by models trained on SEC filings and in-context learning methods. To establish a baseline, we first evaluate the generalization capabilities of SEC-trained models across established SEC datasets. To support our investigation, we introduce three novel benchmarks: (1) SEC Filings Benchmark (SECB), (2) Earnings Calls Benchmark (ECB), and ECB-A, a subset with 2,460 expert annotation groups to support our qualitative analysis. We find that encoder-based models struggle with the domain shift. Finally, we propose a system utilizing LLMs to perform open-ended extraction from unstructured call transcripts, verified by human evaluation (79.7% precision), providing a baseline for this valuable domain through the consistent tracking of emergent KPIs.

Rasmus T. Aavang, Rasmus Tjalk-Bøggild, Alexandre Iolov +3

Department of Computer Science, Aalborg University, Denmark · ALIPES ApS, Denmark · University of Copenhagen

Document Information Extraction LLM Information Extraction Finance

Ask PDF

Jun 29, 2026

Fast Numbers, Slow Language: Bridging Quantitative and Qualitative Earnings Signals

Earnings announcements release two types of information sequentially: quantitative surprise (numeric earnings-per-share (EPS)/revenue versus analyst estimate) arrives first in press releases and financial news, processed by algorithmic traders within minutes; qualitative language (management tone, guidance, question-and-answer (Q&A) credibility) arrives 30-90 min later in the earnings conference call transcript (ECT), requiring human interpretation overnight. Financial economists have studied quantitative surprise for 50 years; natural language processing (NLP) researchers have studied qualitative ECT signals for a decade. Despite studying the same event, the two communities used incompatible frameworks: different targets (return vs. volatility), trading setups (long top-decile and short bottom-decile vs. trade-all), and metrics (return spread between top and bottom 20% (Q5-Q1) vs. mean squared error (MSE)), making direct comparison and connection challenging. We bridge these communities with EarningsInOne, the first corpus aligning earnings news, ECTs, and intraday and next-day prices across SP 1500 (broad U.S. equity universe, 2022-2025). Applying unified trading and evaluation tools to both signal types, we confirm a clean speed separation, fast numbers, slow language: quantitative surprise peaks at announcement and is largely eliminated by the next market open; qualitative ECT sentiment peaks on the next trading day, real and tradeable, but hidden under prior transcript-based evaluation that optimised sign-agnostic volatility with pointwise MSE.

Ding Yu, Zhuo Liu, Hao Zhang +1

University of Rochester · Rochester Institute of Technology

Quantitative Finance Financial Sentiment Analysis Financial Time Series Finance

Ask PDF

Abstract

Look-Ahead Bias in Language Models

Explore similar work

Earnings25: A Comprehensive 500-Hour Speech Benchmark for Finance

Effective Performance Measurement: Challenges and Opportunities in KPI Extraction from Earnings Calls

Fast Numbers, Slow Language: Bridging Quantitative and Qualitative Earnings Signals