A Qwen earnings-call score looks strongest when the return starts before the call. For after-close calls, 206 of 552, a previous-close label absorbs the whole call and the overnight session. The Qwen3-4B answer-valence coefficient changes by +1.355 pp per signal SD between that label and the completion-aligned one. Of the shift, +1.256 pp occurs before the first regular-session price after the documented start. The paper puts the finding plainly: "Most of the association shift is therefore localized before the documentary source-start anchor." It calls all representation-level inference "retrospective and exploratory."

Ma, Koa, Liu and Huang are examining a particular timing problem. Some benchmarks they cite, including VolTAGE, pair a complete-call representation, full Q&A text or audio, with a return that starts before the representation could exist. A chronological train/test split leaves that within-event overlap untouched. Their "Availability Contract" asks for a nine-field declaration of the model input, evidence of its completion, the assumed processing allowance, the label's price anchors and the claim the evidence supports. Its four-rung ladder runs from event-window association to proxy-aligned association, verified post-input and deployability.

The study covers 552 calls at 24 large U.S. firms from 2018 to 2024, drawn from a parent panel of 2,094 calls at 30 firms. Across 8,610 Q&A pairs, the authors compute four call-level probes: a Loughran-McDonald-style tone count, FinBERT, Qwen valence and an E5 "narrative revision" score measuring distance from the same firm's prior answers.

When is the full call available?

Each return ends at the following-session close. The starting price changes:

Because these are log returns, the removed intervals add exactly. The MSFT call on 27 July 2021 shows how the clocks work: its header gives 17:00 ET, audio ends at 18:00:20 and stipulated readiness falls at 18:15:20. Both S and A therefore use the next morning's 09:35 bar. The L return, -0.984 pp, divides into -0.562 removed before the call anchor and -0.422 after it.

The authors are candid about what those clocks establish. Most start times come from documentary headers; there are only six manual corrections and three verified no-change checks. The 15 minutes is stipulated, while transcript availability and model inference completion remain, in their words, unobserved. They place the audit on the proxy-aligned rung and describe the bar closes as "label prices, not execution prices."

The move between clocks

From S to A, return ranks scarcely change: Spearman is 0.900. The absolute difference is 0.179 pp at the median and 0.623 pp at the mean. That mean is 3.5 times the median, a tail large enough to flip the return sign for 9.4% of calls (firm-cluster interval 5.6% to 13.6%). Fixed A-clock quintile cutpoints put 28.3% of calls in a different quintile; recomputed cutpoints put 30.1% in a different one.

All sign flips and fixed-quintile moves occur among the 311 calls (56.3%) for which S and A select different bars. In that group, 16.7% flip sign and 50.2% move across fixed quintiles. The 206 after-close calls provide a useful control: S and A select the same bar, giving identical returns. Yet 6.3% change relative quintile as the surrounding cross-sectional distribution moves. With the allowance varied from 0 to 120 minutes, the sign-flip rate ranges from 6.5% to 15.0%.

The L to S step is larger: 32.1% change sign, and the mean absolute change is 3.427 pp.

It also accounts for nearly all of the Qwen coefficient shift. The coefficient is +1.122 pp/SD under L, -0.135 under S and -0.233 under A; only +0.098 of the +1.355 total lies between S and A. The authors' checks retain the L to A shift: retrospective Holm p = 0.0024, with delete-one-firm shifts from 1.149 to 1.456. The post-call estimate gives less reason to trade. Keep the A start and set the horizon to 390 trading minutes, and the coefficient is -0.226. Across the exact-390-minute family, BH-adjusted q values range from 0.754 to 0.964. The apparent +1.1 pp per SD positive-tone result chiefly reflects the move between the previous close and the first post-start bar.

The score has its own qualifications. Qwen was inspected before the test family was frozen. The authors used a 2025-vintage checkpoint with no established knowledge cutoff, in June 2026, on calls from 2018 onward. Answers were 48.62% positive, 48.19% neutral and 2.54% negative, so the score mainly distinguishes positive answers from neutral ones.

Volatility rankings hold, losses move

The paper also revisits the public Dense Earnings Call volatility benchmark of Yu et al. It rescores four frozen prediction streams, TSMixer and three TMLP configurations, on 1,710 events. For the 1,191 before-market events, the authors move targets forward one session without changing predictions. Three-day MSE increases by 18.1% to 21.4% across the streams; at 15 days, the increase is 3.8% to 6.2%. Point rankings remain unchanged under MSE, MAE and Spearman. Kendall's τ = 1 at every horizon.

The stable order offers limited reassurance. Under issuer resampling, the best-MSE model at three days stays best in only 67.1% of draws, and at least one pair reverses in 43.1%. These models are near-ties, so their unchanged order says little about whether the target window was appropriate. This one-session counterfactual uses four streams without retraining. It measures sensitivity of the losses to that change; it cannot establish which model would win after retraining on shifted targets.

Where the evidence stops

We could not rerun the audit. Its completion-aligned window requires documented call-start times and archived audio durations or sentence-alignment endpoints; our data contains neither. We also lack point-in-time stamps for full-transcript availability, leaving us unable to date a transcript signal before its return window.

The available sample is narrow. Just 26.4% of the parent panel survives the joins. The source-clock join alone removes 1,339 calls and changes mean log market cap by about one pooled SD. After every join, 2 calls remain in 2018, and inference rests on 24 firm clusters.

Ma, Koa, Liu and Huang have produced a disciplined audit that is frank about its evidentiary rung. Their abstract says "labels and losses can depend materially on the declared outcome clock even when model rankings appear stable." The 9.4% sign-flip rate and the 18.1% to 21.4% MSE increase support that claim. Their nearest approximation to completion is a proxy clock: Qwen's coefficient there is -0.233 pp/SD to the following close and -0.226 over 390 minutes, with the latter nonsignificant after BH correction. A verified transcript clock remains open. A positive, significant coefficient measured on that clock across more than 24 firms would make me treat the score as a post-call signal.