Yin, Yanbin, Xu, Chuyu, Ye, Liqin, Peng, Jing, Zhang, Chao · Oct 28, 2026
EPSBench tests LLM agents’ earnings forecasts
Markdown · JSON · Original source
Yanbin Yin, Chuyu Xu, Liqin Ye, Jing Peng and Chao Zhang introduce EPSBench, a for testing whether agents can connect financial evidence to the income-statement items behind reported earnings. Agents use filings, earnings-call transcripts and news to estimate company-specific drivers; an accounting calculator converts those estimates into GAAP diluted earnings per share (EPS). The authors compare six models’ point forecast error with date-matched analyst consensus and assess coverage. The authors report that three models show no detectable paired difference in point forecast error from consensus, while three have higher error. Nominal 80% intervals contain reported EPS in only 37–70% of runs that return an interval. Most misses fall above the upper bound because the forecasts run low. The authors trace most of the error gap between models to operating-income estimates. EPSBench records estimates feeding the final forecasts, allowing researchers to examine how model assumptions relate to forecast error. The abstract reports results across 149 firm-quarters from 86 US companies, but does not establish how they generalize beyond the evaluated data and models.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Oct 28, 2026 (2026-10-28).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-11T09:10:29.597Z. First observed: 2026-10-30T06:42:49.633Z. Retrieved: 2026-10-31T04:33:27.539Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The supplied evidence is the abstract, so it does not provide details of the design, model identities, or evaluation results beyond those summarized there. The abstract does not establish how the findings generalize beyond the evaluated firm-quarters and models.
Open question
What accounts for the forecasts running low, which the authors say puts most interval misses above the upper bound?
Source-stated
The authors introduce EPSBench to study whether agents can connect financial evidence to company-specific income-statement drivers and use them to forecast earnings.
We introduce EPSBench, a benchmark for studying this ability across 149 firm-quarters from 86 US companies.
Source-stated
EPSBench compares model point forecast error with date-matched analyst consensus and checks prediction-interval coverage.
assessing point forecast error relative to date-matched analyst consensus and checking whether their prediction intervals achieve the stated coverage.
Source-stated
The authors report that three models had no detectable paired difference in point forecast error from consensus, while three had higher error.
Three models show no detectable paired difference in point forecast error from consensus, and three have higher error.
Source-stated
Reported EPS fell within nominal 80% intervals in 37–70% of runs that returned an interval; most misses were above the upper bound.
nominal 80% intervals contain reported EPS in only 37--70% of runs that return an interval, and most misses fall above the upper bound because the forecasts run low.
Source-stated
The authors trace most of the error gap between models to operating-income estimates.
The driver records trace most of the error gap between models to operating-income estimates.
Source and contributor context
Yin, Yanbin, Xu, Chuyu, Ye, Liqin, Peng, Jing, Zhang, Chao · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Uncertainty interval
A range that a true value is expected to fall inside, with a stated chance of doing so.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 9dd17703-da44-4e0c-8b97-9baa880194fb.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-11T09:10:29.597Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: cf2fe65e109c4ca26c89e56d897ed731d0f5318ac95c523dfaa0f773b65dacf0. Source text hash: 3d4abbc211c9ee6a2a11485fd343966942ae0977808b0e0611c076e10f8b7780.
Revision and source-check history
Version 1 · 2026-11-11T09:10:29.597Z
EPSBench tests LLM agents’ earnings forecasts
Yanbin Yin, Chuyu Xu, Liqin Ye, Jing Peng and Chao Zhang introduce EPSBench, a for testing whether agents can connect financial evidence to the income-statement items behind reported earnings. Agents use filings, earnings-call transcripts and news to estimate company-specific drivers; an accounting calculator converts those estimates into GAAP diluted earnings per share (EPS). The authors compare six models’ point forecast error with date-matched analyst consensus and assess coverage. The authors report that three models show no detectable paired difference in point forecast error from consensus, while three have higher error. Nominal 80% intervals contain reported EPS in only 37–70% of runs that return an interval. Most misses fall above the upper bound because the forecasts run low. The authors trace most of the error gap between models to operating-income estimates. EPSBench records estimates feeding the final forecasts, allowing researchers to examine how model assumptions relate to forecast error. The abstract reports results across 149 firm-quarters from 86 US companies, but does not establish how they generalize beyond the evaluated data and models.
Original source
cs.AI · Article
Source details
Original source date: Oct 28, 2026 (2026-10-28).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-11T09:10:29.597Z. First observed: 2026-10-30T06:42:49.633Z. Retrieved: 2026-10-31T04:33:27.539Z.
Limits
The supplied evidence is the abstract, so it does not provide details of the design, model identities, or evaluation results beyond those summarized there. The abstract does not establish how the findings generalize beyond the evaluated firm-quarters and models.
Open question
What accounts for the forecasts running low, which the authors say puts most interval misses above the upper bound?
Source-stated
The authors introduce EPSBench to study whether agents can connect financial evidence to company-specific income-statement drivers and use them to forecast earnings.
We introduce EPSBench, a benchmark for studying this ability across 149 firm-quarters from 86 US companies.
Source-stated
EPSBench compares model point forecast error with date-matched analyst consensus and checks prediction-interval coverage.
assessing point forecast error relative to date-matched analyst consensus and checking whether their prediction intervals achieve the stated coverage.
Source-stated
The authors report that three models had no detectable paired difference in point forecast error from consensus, while three had higher error.
Three models show no detectable paired difference in point forecast error from consensus, and three have higher error.
Source-stated
Reported EPS fell within nominal 80% intervals in 37–70% of runs that returned an interval; most misses were above the upper bound.
nominal 80% intervals contain reported EPS in only 37--70% of runs that return an interval, and most misses fall above the upper bound because the forecasts run low.
Source-stated
The authors trace most of the error gap between models to operating-income estimates.
The driver records trace most of the error gap between models to operating-income estimates.
Source and contributor context
Yin, Yanbin, Xu, Chuyu, Ye, Liqin, Peng, Jing, Zhang, Chao · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Uncertainty interval
A range that a true value is expected to fall inside, with a stated chance of doing so.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 9dd17703-da44-4e0c-8b97-9baa880194fb.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-11T09:10:29.597Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: cf2fe65e109c4ca26c89e56d897ed731d0f5318ac95c523dfaa0f773b65dacf0. Source text hash: 3d4abbc211c9ee6a2a11485fd343966942ae0977808b0e0611c076e10f8b7780.
passed · 2026-11-11T09:10:29.597Z · anthropic/claude-opus-5.5
Publication: 9dd17703-da44-4e0c-8b97-9baa880194fb
attribution — pass: The reading credits the five listed authors and attributes the benchmark design and results to them. Findings stay framed as reported by the authors, not as Escape Meridian conclusions or independently verified facts.
Authors:Yanbin Yin, Chuyu Xu, Liqin Ye, Jing Peng, Chao Zhang
summary — pass: The title, summary and question match the abstract. It leads with the actor and the development, keeps technical terms exact, and preserves the key conditions: 37–70% of runs that return an interval, forecasts running low, and operating-income estimates. 'Nominal 80% intervals' and 'paired difference' could use term-library tooltips for general readers, but the shortfall against 80% is still readable. The question follows from the evidence: the abstract reports the low bias but not its cause.
most misses fall above the upper bound because the forecasts run low
date — pass: The displayed original publication date, 2026-10-28, matches the arXiv v1 submission date. The later feed date is not used.
[Submitted on 28 Oct 2026]
support — pass: Each claim carries an exact source excerpt. The significance sentence, that recorded estimates let researchers relate assumptions to error, is grounded in the abstract's 'common basis for comparing forecast assumptions' and its driver records. The generalization limit is stated without overreach. No benefits, costs or real-world reliability are invented.
providing a common basis for comparing forecast assumptions
media — pass: The selected media is the verified arXiv abstract page for this paper. It is the direct original source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA language model with many learned numerical settings, used to process and generate language.
Definition referenceA range that a true value is expected to fall inside, with a stated chance of doing so.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA language model with many learned numerical settings, used to process and generate language.
Definition referenceA range that a true value is expected to fall inside, with a stated chance of doing so.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference