Language models are now a standard tool for measuring what managers say on earnings calls, but the encoders behind these measures were usually pretrained on text from the same years a backtest covers. Holding out a test period does not remove that hindsight, because it sits in the features. We adapt NoLBERT, an encoder pretrained only on text published between 1976 and 1995, to earnings calls without letting it read anything dated after 31 December 2023. The calls supply the training signal themselves: every analyst question is followed by management's answer, and the model learns to pick that answer out from the other answers of the same call. We test on every US call from January 2025 to October 2026 (20,323 calls, 254,575 questions) and make every modeling choice on calls from 2024. The adapted model, NoLBERT-Calls-2023, ranks the true answer first for 78.0% of questions (95% interval 77.8-78.2; chance 8.0%), against 26.3% before adaptation. It is 1.0 point ahead of a call-adapted ChronoBERT, 3.4 points ahead of a call-adapted FinBERT, and 18.7 and 19.8 points ahead of the best off-the-shelf retriever and BM25. A year-by-year test over 2013-2026 finds no look-ahead premium: relative to a model adapted only through 2012, the released model does no better on the years it was trained on than on years it never saw. The price of the cutoff is compute: under one shared 150-minute recipe NoLBERT adapts least well of four encoders, and its lead comes from a longer, staged adaptation of about 13 GPU-hours.
Figures & tables
Figure 1: From transcripts to a reader without hindsight. Top: the five steps. Bottom: what each step reads, by date. NoLBERT’s pretraining text ends in 1995. Stage 1 reads calls from all countries up to 2012 and stage 2 reads US calls up to 31 December 2023. Every choice is made on 300 calls from 2024, and the test covers every US call from January 2025 to October 2026. The time test scores 500 calls a year from 2013 to 2024 and then the main test.
Figure 2: US earnings calls. (a) Calls per year and their role in this paper. The line counts calls of firms outside the US, which enter stage 1 up to 2012. (b) Share of calls that mention each subject, in a random sample of 250 US calls a year; the shaded years come after the cutoff of the released model. (c) Analyst exchanges per call, in the same samples (all calls) and in the calls of the time test and the main test (evaluation calls, at least three exchanges each). 2026 runs to 31 October.
Protocol
Training calls
Validation
Test
Main (cutoff 2023)
US calls up to 31 Dec 2023: 196,569 calls, 3,052,636 exchanges, 785,924 passages
300 US calls from 2024 (3,763 questions)
Every US call from Jan 2025 to Oct 2026: 20,323 calls, 254,575 questions
Time (cutoff 2012)
US calls up to 31 Dec 2012: 59,523 calls, 1,067,124 exchanges
300 US calls from 2013–2015 (5,398 questions)
500 US calls a year, 2013–2024 (6,268 to 8,864 questions a year), then the main test
Table 1: Sample splits.
Group
Readers
Training text and its dates
Word overlap
TF-IDF, BM25
No training. Word weights are fitted on calls up to the cutoff.
NoLBERT: text of 1976–1995. ChronoBERT: text up to each year end. FinBERT: pretraining includes calls of 2004–2019. DeBERTa-v3: no date control.
Off-the-shelf retrievers
all-mpnet-base-v2, BGE-base-en-v1.5, E5-base-v2
Web-scale training data with no date control.
Same recipe, calls up to 2023
NoLBERT-, ChronoBERT-, FinBERT- and DeBERTa-Calls-2023
US calls up to 2023; one 150-minute run on one T4 each.
Same recipe, calls up to 2012
NoLBERT-Calls-2012, ChronoBERT-Calls-2012
US calls up to 2012; one 150-minute run on one T4 each.
Our staged model
NoLBERT-Calls-2012, stage 1
Calls from all countries up to 2012; 480 minutes on one RTX 3060.
Table 2: Readers compared. The released model is shaded.
Figure 3: NoLBERT and its adaptation to calls. (a) The pretrained encoder of Kakhbod and Li (2025) , a DeBERTa-v3-base model trained to predict masked words in text of 1976–1995. (b) One training step. The questions and answers of a call pass through the same encoder, and the loss rewards placing each question closest to its own answer. The other answers of the call serve as hard negatives.
Reader
Top-1 [95% interval]
+ word overlap
Top-3
MRR
Δ to released [95% interval]
Adapted on calls up to 2023
NoLBERT-Calls-2023 (released)
78.0 [77.8, 78.2]
79.0
94.0
86.4
—
ChronoBERT-Calls-2023 at 256 tokens
77.0 [76.8, 77.2]
77.9
93.5
85.7
1.0 [0.9, 1.2]
FinBERT-Calls-2023
74.7 [74.5, 74.9]
77.1
92.8
84.2
3.4 [3.2, 3.5]
ChronoBERT-Calls-2023
72.3 [72.1, 72.6]
74.6
91.2
82.4
5.7 [5.5, 5.9]
DeBERTa-Calls-2023
68.2 [68.0, 68.4]
74.4
89.6
79.7
9.8 [9.6, 10.0]
Table 3: Within-call answer retrieval on every US call from January 2025 to October 2026 (20,323 calls, 254,575 questions).
Figure 4: What drives accuracy on the 2025–2026 test. (a) Top-1 of each pretrained encoder used zero-shot (open circles) and after adaptation on US calls up to 2023 with the shared 150-minute recipe (filled). The open teal circle reads ChronoBERT-Calls-2023 with 256-token answers, without retraining, and the star is the released model. Dashed lines mark BM25 and E5, the dotted line chance. (b) Top-1 by the number of candidate answers in the call. The released model leads in every bin.
Figure 5: How the released model was built. (a) Stage 1: validation top-1 on 300 calls from 2013–2015, pooled over questions, during the 480-minute run; the circle marks the selected checkpoint. (b) Validation top-1 on the 300 calls of 2024, weighting calls equally, against training minutes on one T4. Dashed lines: each base adapted with the shared recipe. Solid lines: stage 2, starting from stage 1, with 128-token answers (run v1) and with 256-token answers (run v4, then a second run, which is the released model). (c) Validation and test top-1 of the five stage-2 runs. An ensemble of v4 and v3 scored higher on the test but lower on validation and was not chosen.
Figure 6: The time test. (a) Top-1 by call year, on 500 US calls a year for 2013–2024 and on the full main test for 2025 and 2026. Teal shading marks the years the 2023 models were trained on, light orange 2024 (validation) and orange the test years. (b) Each reader’s lead over NoLBERT-Calls-2012 (same recipe) on the same calls. Thin dotted lines show the average lead over 2013–2023 and over 2025–2026. The stage-1 model has a 2012 cutoff and has read neither period.
Lead over NoLBERT-Calls-2012
Premium [95% interval]
Reader
2013–23
2021–23
2025–26
2013–23 vs 2025–26
2021–23 vs 2025–26
Adapted on calls up to 2023
NoLBERT-Calls-2023 (released)
+12.0
+13.3
+12.9
− 0.9 [ − 1.2, − 0.5]
+0.3 [ − 0.3, +1.0]
FinBERT-Calls-2023
+10.0
+10.7
+9.6
+0.4 [+0.1, +0.8]
+1.1 [+0.5, +1.8]
ChronoBERT-Calls-2023
+7.1
+7.8
+7.2
− 0.1 [ − 0.4, +0.3]
+0.5 [ − 0.1, +1.1]
DeBERTa-Calls-2023
+3.1
+4.1
+3.1
+0.0 [ − 0.4, +0.5]
+1.0 [+0.2, +1.8]
Table 4: Look-ahead premium in the time test.
Passage matching
Masked terms
Reader
Top-1
Top-10
Role
Top-1
Top-5
SubjECTive-QA probe
NoLBERT-Calls-2023 (released)
7.6
21.9
96.2
25.6
47.1
64.8
NoLBERT-Calls-2012, stage 1
5.8
17.9
95.9
23.9
45.1
65.4
NoLBERT-Calls-2023, same recipe
6.3
15.8
95.6
21.3
40.2
65.1
FinBERT-Calls-2023
8.2
27.9
97.8
39.6
65.0
66.5
ChronoBERT-Calls-2023
6.6
19.2
95.4
25.2
47.1
65.4
Table 5: Tasks on frozen representations, calls of 2025–2026 unless noted.
SubjECTive-QA, weighted F1
GLUE, development sets
Encoder
Six properties (sd)
Four properties
CoLA
QNLI
SST-2
NoLBERT, zero-shot base
64.4 (0.9)
54.8
14.8
88.0
85.9
NoLBERT-Calls-2023, same recipe
65.7 (1.0)
56.8
26.8
88.2
86.4
NoLBERT-Calls-2023 (released)
65.8 (1.5)
56.8
31.9
88.2
86.8
FinBERT-Calls-2023
66.2 (0.9)
57.4
—
—
—
ChronoBERT-Calls-2023
65.6 (1.0)
56.5
—
—
—
Table 6: Benchmarks after fine-tuning the whole encoder.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Reader
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
TF-IDF (word weights from calls up to 2012)
49.7
49.9
50.5
51.1
52.1
53.1
53.3
53.7
55.9
55.6
56.3
56.1
56.8
57.3
NoLBERT, zero-shot
22.5
22.4
22.9
23.1
24.8
24.2
25.0
24.8
25.1
25.9
25.7
25.9
26.3
26.4
ChronoBERT-2012, zero-shot
20.7
20.5
21.6
21.6
22.9
21.8
22.1
22.4
22.3
22.5
23.2
23.7
23.8
23.6
FinBERT, zero-shot
29.5
28.8
28.8
30.2
30.1
29.8
30.3
30.8
31.0
31.3
31.6
31.6
31.9
31.8
NoLBERT-Calls-2012, stage 1
70.1
70.2
70.6
70.2
70.8
71.2
71.9
69.8
71.3
71.4
71.2
71.1
71.3
71.5
NoLBERT-Calls-2012, 150-minute stage 1
57.6
56.9
58.5
58.5
58.5
59.7
59.2
58.1
60.3
60.5
59.9
60.0
60.5
60.7
Appendix
Table A1 : The time test by call year: top-1 in percent.
Reader
Top-1
95% interval
Top-1, call-weighted
Top-3
MRR
Δ to released [95% interval]
Adapted on calls up to 2023
NoLBERT-Calls-2023 (released)
78.0
[77.8, 78.2]
80.2
94.0
86.4
—
+ word overlap
79.0
[78.8, 79.2]
81.0
94.4
87.1
− 1.0 [ − 1.0, − 0.9]
ChronoBERT-Calls-2023 at 256 tokens
77.0
[76.8, 77.2]
79.2
93.5
85.7
1.0 [0.9, 1.2]
+ word overlap
77.9
[77.7, 78.1]
80.0
93.9
86.3
0.1 [ − 0.0, 0.3]
FinBERT-Calls-2023
74.7
[74.5, 74.9]
77.0
92.8
84.2
3.4 [3.2, 3.5]
Appendix
Table A2 : Within-call answer retrieval on the 2025–2026 test: every reader.
Run
Change
Exchanges per call
QA batches
Learning rate
Answer tokens
Steps (best)
Validation top-1
Test top-1
Test, + word overlap
v1
Warm start from stage 1
8
75%
2×10−5
128
6,963 (6,270)
75.8
73.8
75.8
v2
+ same-call packing
32
75%
2×10−5
128
7,239 (7,239)
77.1
74.4
76.6
v3
v2 + QA-only batches, higher LR
32
100%
3×10−5
128
6,882 (4,821)
78.5
75.8
77.6
v4
v2 + 256-token answers
32
75%
2×10−5
256
3,397 (3,397)
80.3
77.7
78.6
v4c (released)
v4 + a second 150-minute run
32
75%
2×10−5
256
3,722 (3,722)
80.7
78.0
79.0
Ensemble
Average of v4 and v3 (not released)
80.3
78.4
79.3
Appendix
Table A3 : Stage-2 runs. Every run lasts 150 minutes on one T4 and starts from stage 1, except v4c, which continues v4.
Passage matching
Masked terms
Reader
Top-1
Top-10
Role
Top-1
Top-5
SubjECTive-QA probe
NoLBERT-Calls-2023 (released)
7.6
21.9
96.2
25.6
47.1
64.8
NoLBERT-Calls-2012, stage 1
5.8
17.9
95.9
23.9
45.1
65.4
NoLBERT-Calls-2023, same recipe
6.3
15.8
95.6
21.3
40.2
65.1
v4 (stage-2 run 1)
8.3
22.1
96.6
24.9
46.4
65.2
v3
3.9
11.5
95.6
25.9
48.1
65.0
Appendix
Table A4 : Tasks on frozen representations: every reader.
Encoder
Assertive
Cautious
Clear
Optimistic
Relevant
Specific
Mean
NoLBERT, zero-shot base
49.0 (1.4)
59.9 (2.2)
82.5 (0.8)
55.4 (0.9)
84.9 (1.1)
54.8 (1.6)
64.4 (0.9)
NoLBERT-Calls-2023, same recipe
49.9 (3.6)
61.0 (2.3)
82.5 (0.8)
59.5 (1.7)
84.8 (1.1)
56.6 (1.3)
65.7 (1.0)
v3
49.9 (4.0)
61.2 (2.7)
82.5 (0.8)
60.0 (1.2)
84.9 (1.1)
56.4 (2.5)
65.8 (1.4)
NoLBERT-Calls-2023 (released)
50.3 (4.3)
61.3 (3.0)
82.5 (0.8)
59.4 (2.8)
84.9 (1.1)
56.3 (2.8)
65.8 (1.5)
FinBERT-Calls-2023
50.0 (4.3)
62.0 (1.9)
82.4 (0.7)
58.2 (3.6)
85.2 (1.1)
59.4 (1.9)
66.2 (0.9)
ChronoBERT-Calls-2023
48.8 (4.1)
60.7 (1.1)
82.8 (1.2)
60.0 (1.4)
84.8 (1.0)
56.7 (0.3)
65.6 (1.0)
Appendix
Table A5 : SubjECTive-QA after fine-tuning, property by property: weighted F1, mean (standard deviation) over the authors’ three splits.
Setting
Stage 1
Stage 2 (released)
Starting weights
NoLBERT
Selected stage 1 checkpoint
Training calls
All countries, up to 31 Dec 2012
US, up to 31 Dec 2023
Batches
75% question–answer, 25% passage pairs
75% question–answer, 25% passage pairs
Exchanges of one call per batch
Up to 8
Up to 32
Tokens, question and answer
128 and 128
128 and 256
Batch size
32
32
Appendix
Table A6 : Settings of the two adaptation stages.
Item
Hardware
Budget
Stage 1
One RTX 3060 (12 GB)
480 minutes
Stage 2, released model
One T4 per run
2 × 150 minutes
Released model, total
About 13 GPU-hours
Stage 2 development runs (v1–v3)
One T4 each
3 × 150 minutes
Same-recipe adaptations, six models
One T4 each
6 × 150 minutes
Stage 1, 150-minute run (reference)
One RTX 3060
150 minutes
Appendix
Table A7 : Compute. Evaluation runs are not included.
Text encoders turn earnings calls into vectors for search, for language models that answer from retrieved passages, and for forecasting models. A live retriever can use an encoder trained once on all available text, as finance retrievers are, but a backtest needs an encoder that read no text written after the date of each vector. To address this look-ahead bias, from 412,248 calls we train vintages under four pretraining recipes, with yearly cutoffs from 2010 to 2024, and each vintage reads only text dated on or before its cutoff. We then fine-tune each vintage on the question-answer pairs of the calls themselves. The vintages we train on calls alone, with 33.8M parameters, retrieve answers on held-out calls better than the general retrievers we tested in every test quarter. Their vectors also raise by about half the rank correlation between next-day returns and forecasts built from past returns and firm size. When each vintage continues the previous one, a later vintage retrieves better on earlier quarters than the vintage available at the time, and a point-in-time language model that answers from the passages the later vintage retrieves inherits the gain. The 2024 vintage of our families also forecasts earnings surprises better, although the interval of this gain includes zero on calls that no vintage read. It forecasts returns no better, so point-in-time vintages matter most for retrieval and little for forecasts of returns.
Alessio Brini
Pratt School of Engineering, Duke University, Durham, NC, USA
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.
Xiaoyu Ma, Kelvin J. L. Koa, Yuncong Liu +1
National University of Singapore Asian Institute of Digital Finance Singapore
Financial restatements correct errors in previously published financial reports. Could managers' language in earnings calls warn of them in advance? Comparing 109 calls that precede restatements with 218 matched control calls, we find that the text predicts restatements but adds no detectable ranking gain beyond public firm information. A text classifier beats chance, largely because of restatements tied to special-purpose acquisition companies, but a model using only public information scores slightly higher. With refitting, the 95% interval for the gain from adding text (-0.05 to 0.11 AUROC) still leaves room for a meaningful benefit. A large language model scores higher, but we show that it names the exact ticker in 66% of masked transcripts, and it was released after the restatements occurred. Our results suggest that finding new warning signs in call text requires a public-information baseline and a check of what a language model already knows.