PaperlayerView exact revision on arXiv

Source-linked AI summary

AnalysisBank: An Expert Analysis Pattern Library for Financial Report Generation

Yajing Yang, Yunshan Ma, Kelvin J. L. Koa, Min-Yen Kan

arXiv:2609.00818v1cs.AI

TL;DR

Financial report generation often leaves the choice of analysis to model defaults, despite the need for insights driven by specific data signals. ANALYSISBANK distills expert reports into reusable signal–move–span Analyses, retrieves and applies them at inference, and increases novel, data-grounded insights across benchmarks and backbones. The paper concludes that analytical-level generation may generalize beyond finance, while its library and effectiveness remain bounded by source-report quality, language coverage, and domain fit.

  • Problem

    Existing report-generation methods prescribe structure but leave the data-specific analytical reasoning to model defaults.

  • Method

    ANALYSISBANK distills expert reports into reusable Analyses pairing data signals, analytical moves, and expert spans, then retrieves and applies them at inference.

  • Results

    1.7–3.7× more novel, data-grounded insights are produced across two financial benchmarks and four LLM backbones than with structural-level baselines.

  • Takeaways & Limitations

    Transfer to scientific writing suggests that operating at the analytical level may generalize beyond finance to other forms of expert, data-grounded reasoning.

  • Takeaways & Limitations

    The library cannot produce analytical patterns richer than its source corpus, and experiments use exclusively English corpora; benefits are limited where fixed analyses suffice.

Abstract

from arXiv · show

We argue that financial report generation should operate at the analytical rather than structural level, composing content from data-derived insights rather than high-level topics or sections. To this end, we propose AnalysisBank, which distills expert reports into a reusable library of Analyses, each pairing a data signal, an analytical move, and the expert span it was derived from. At inference time, AnalysisBank matches input signals to library entries and applies the retrieved moves to compose the report. A study of Analyses distilled from 550 expert reports reveals a heavy-tailed distribution of 47-52 signal types spanning 13 move types. On two financial benchmarks across four LLM backbones, AnalysisBank increases the proportion of novel, data-grounded insights by 1.7-3.7x over structural-level baselines. Transfer to scientific writing suggests that the distinction generalizes beyond finance. Code and the distilled Analysis library are available at https://github.com/yajingyang/AnalysisBank.

1 Introduction

Financial report generation should determine which analyses the data warrants, rather than merely prescribe report structure. ANALYSISBANK addresses this by retrieving expert-distilled, signal-specific analytical moves, increasing novel, data-grounded insights across benchmarks and model backbones.

  • Analytical reports produce claims derived from source data, unlike summarization and data-to-text generation, which paraphrase or restate the input.
  • Existing structural-level methods prescribe topics or sections but leave the reasoning within each section to language-model defaults.
  • Signal-specific conditions call for distinct analytical moves, such as inferring relative outperformance from a smaller-than-peer revenue decline.
  • 47–52 signal types span 13 analytical move types, with over a third appearing in three or fewer instances and pairing with only one to two moves.
  • ANALYSISBANK stores each expert-derived Analysis as a reusable pairing of a data signal, an analytical move, and its source expert span.
  • 1.7–3.7× more novel, data-grounded insights are produced across two benchmarks and four LLM backbones than with structural-level baselines.

2 Related Work

Prior work improves report organization, reusable reasoning, or financial-language modeling, but generally does not explicitly connect specific data signals to the analyses they warrant. ANALYSISBANK is distinguished by making that signal-to-move connection central.

  • Long-form and report generation: Data-to-text systems generate fluent prose from tables and records but generally restate inputs rather than analyze them.
  • Long-form and report generation: Report-generation methods add organization through predefined roles, section planning, or retrieval, while leaving analytical content largely structurally specified.
  • Reusable reasoning patterns: Reusable reasoning patterns provide abstract strategies or complete procedures, rather than individual analytical moves grounded in specific data patterns.
  • Financial NLP: Financial NLP models adapt language models to financial text, but signal-mining and reasoning systems do not explicitly model what analysis a signal calls for.
  • Reusable reasoning patterns: The example Analysis links three co-occurring comparisons—prior period, expectations, and peers—to an inference of outperformance and share gains.

3 ANALYSISBANK

ANALYSISBANK represents expert reasoning as reusable Analyses and constructs a library through extraction, abstraction, deduplication, and quality filtering. The resulting libraries capture broad but heavy-tailed signal coverage across financial-report corpora.

  • 3.1 Analysis Representation: Each Analysis is a tuple of data_signal, analytical_move, and reference_text.
  • 3.1 Analysis Representation: The reference_text preserves a verbatim expert span as a generation anchor and faithfulness check, while specific entities, figures, and times limit transfer.
  • 3.1 Analysis Representation: The analytical_move abstracts a source span into a reusable instruction, while the separate data_signal expresses entity-free, number-free triggering conditions for retrieval.
  • 3.1 Analysis Representation: The three-field decomposition is a design hypothesis whose alternatives show degraded retrieval precision or generation quality when fields are removed or merged.
  • 3.2 Extraction Pipeline: The four-pass pipeline identifies candidate spans, generalizes signals and moves, deduplicates similar entries, and quality-filters transferability, actionability, and grounding.
  • 3.3 Coverage and Distribution: 1,422 DataTales Analyses and 3,889 Earnings Analyses are derived from 550 reports in each corpus, with 13 move types and 47 or 52 populated signal types.
  • 3.3 Coverage and Distribution: 90% coverage requires 19 signal types in DataTales and 22 in Earnings, while tail signals remain sparse and concentrate on one to two analytical moves.

4 Narrative Generation with ANALYSISBANK

The generation pipeline converts source inputs into typed signals, retrieves relevant Analyses, executes each analytical move independently, and composes the resulting findings into a report. Its structure is driven by the findings rather than a fixed section template.

  • The library supplies coverage and field structure for selecting and applying relevant Analyses to a given input.
  • Stage 1: Signal extraction: Stage 1 reduces raw financial input to typed signals with descriptions and supporting source spans, using LLM extraction for text and deterministic computation for structured data.
  • Stage 2: Per-type retrieval: Stage 2 retrieves one best-matching Analysis per signal type by cosine similarity, then fills remaining slots with global top-k entries.
  • Stage 3: Per-Analysis execution: Stage 3 independently applies each retrieved analytical_move to its triggering signals and source spans to produce distinct analytical findings.
  • Stage 4: Composition: Stage 4 composes findings into a report organized around investment-relevant themes inferred from the findings rather than a fixed section template.

5 Experiments

The experiments evaluate AnalysisBank on two analytical report-generation benchmarks using four language-model backbones and multiple comparison families. Evaluation covers analytical quality, factual correctness, and preference-based report quality.

  • Benchmarks: DataTales pairs structured market data with expert daily reports across 460 instances, while Earnings2Insights contains 132 earnings-call-transcript instances.
  • Models: Four backbones span small open-source models, a reasoning model, and a proprietary model, using identical configurations across conditions.
  • Baselines: The baselines include Direct, CoT, and RAG prompting, plus structural-level libraries based on Buffer of Thoughts and Agent Workflow Memory.
  • Metrics: Insight rate is the headline metric, while analysis rate measures overall expert-level analytical content and reasoning depth measures analytical hops per claim.
  • Metrics: Evaluation also measures factual precision against source numerical values and win rate against expert references using analyst and investor LLM judges.

6 Results and Discussions

AnalysisBank consistently improves novel, data-grounded insight over structural-level baselines across financial benchmarks and backbones, with gains linked to retrieved and executed analytical patterns. Ablations show that abstract signal-based retrieval and intermediate findings are central, while transfer to SciGen is positive but less decisive for insight rate.

  • Main Results: 1.7–3.7× higher insight rate is achieved by AnalysisBank than by all baselines across the two financial benchmarks.On DataTales, insight rate reaches 21.1–26.1% versus 9.2–14.0%; on Earnings2Insights, 18.0–23.8% versus 6.4–7.7% for three non-GPT backbones.
  • Baseline Comparison: Structural-level baselines remain within a 4–16% insight-rate band across backbones, despite distilling from the same 550-report corpus.This result indicates a ceiling on insight for methods that prescribe report structure without selecting analytical moves at the data-signal level.
  • Insight Analysis: 80.4% of AnalysisBank’s novel claims trace to retrieved Analyses, and it covers 4.6 signal types per instance versus 0.83 for the baseline.The broader signal coverage corresponds to a 5.5× gap and indicates that novel insights span more data signals rather than concentrating on the most prominent one.
  • Ablation Study: Retrieving with reference_text lowers insight rate by 4.5 points and executes 36.4% of retrieved moves, versus 52.8% for data_signal.Data_signal selects the more applicable move on 80% of reports, because abstraction matches patterns rather than surface wording.
  • Ablation Study: Removing findings from composition halves insight rate to 11–13%, while the findings-only condition preserves 25.1%.Adding signals and transcript restores factual precision from 94.1% to 97.2% and win rate from 96.5% to 99.6%, at a 1.3-point insight cost.
  • Cross-Domain Transfer: On SciGen, AnalysisBank reaches the highest analysis rate at 65.7% and a 65.1% win rate, but BoT slightly leads insight rate at 17.9% versus 16.7%.The authors attribute the weaker insight-rate advantage to scientific writing’s fewer and more predictable analytical moves, although overall quality is highest without domain-specific adaptation.

7 Conclusion

AnalysisBank retrieves and applies distilled expert Analyses to produce data-specific reasoning rather than generic report content. Its transfer to scientific writing suggests this analytical-level approach may extend beyond finance, while visualization integration remains future work.

  • AnalysisBank addresses the gap between structural report planning and determining which analyses specific data warrants.
  • The method retrieves and applies distilled expert Analyses at inference time to increase novel, data-grounded insights across benchmarks and model backbones.
  • Transfer to scientific writing suggests analytical-level reasoning may generalize beyond finance to other expert, data-grounded tasks.
  • The current text-only pipeline leaves iterative analysis, compositional findings, and visualization-integrated reports for future work.

Limitations

AnalysisBank’s library quality depends on the depth and diversity of its source reports, while its scope is constrained by language and domain.

  • The Analysis library cannot contain analytical patterns richer than those present in the expert reports used for distillation.
  • The experiments use only English corpora, and the effectiveness of the LLM-based abstraction pipeline in other languages remains untested.
  • AnalysisBank offers limited advantage in domains where a fixed standard set of analyses suffices because long-tail coverage is unnecessary.

Ethical Considerations

The paper releases distilled Analyses and extraction code while withholding source reports, and it emphasizes evaluation and deployment safeguards for generated financial content.

  • The authors release distilled Analyses and extraction code but not the publicly available source reports, respecting their authors’ and publishers’ intellectual property.
  • Automated evaluation relies on LLM-based judgment, while human evaluation used two unpaid voluntary annotators from the authors’ professional network.
  • Generated reports may contain analytical errors or unsupported inferences, so the paper recommends human review before investment decisions and disclosure that reports are AI-generated.

C.3 Stage 3: Per-Analysis analysis

Stage 3 applies each retrieved Analysis independently to its triggering signals and supporting context, producing analysis fragments that Stage 4 later composes into a structured report.

  • Per-Analysis generation: Each Analysis receives its analytical move, triggering signals, and supporting excerpts or market-data context in an independent LLM call.Calls run in parallel, and prompts prohibit section headers because structure is imposed later.
  • DataTales adaptation: DataTales replaces earnings terminology with market-specific language and uses deterministic market-data summaries instead of transcript spans.The adaptation asks for prices, spreads, and time-window references while retaining top-2 cosine-matched signals.
  • Report composition: Stage 4 combines per-Analysis findings into an executive summary, evidence-driven theme sections, and horizon-specific investment recommendations.Recommendations use Next Day, Week, and Month horizons with Long, Short, or Neutral calls.
  • Validation: A validator checks whether each Analysis was applied and reruns only missing Analyses before recomposing the report.Each verdict records applied status, quality, and a note; validation and retries use parallel calls.
  • Execution cost: The default Earnings2Insights configuration requires roughly 30 LLM calls per report before retries, compared with 11 for DataTales.DataTales uses deterministic Stage 1 processing, reducing the default call count.

E.2 Per-metric protocols

The evaluation extracts claims, classifies their analytical status, checks factual accuracy, measures inference depth, and compares reports against expert references.

  • Claim taxonomy: Claims are classified as factual, novel, standard, or generic after boilerplate is removed from generated reports.Novel claims are data-specific inferences not directly stated in the source, while standard claims are routine analyses.
  • Claim metrics: %insight measures the share of novel claims, whereas %analysis also counts standard claims in its numerator.Both metrics include factual claims in the denominator.
  • Inference depth: Depth is the mean number of analytical hops per claim, with depth-1 claims automatically labeled factual and excluded from analytical type scoring.The judge also emits hop types such as FACTUAL, COMPARE, ATTRIBUTE, and PROJECT.
  • Factuality: Numerical factuality is audited sentence by sentence against transcripts for Earnings2Insights and structured derived-value tables for DataTales.Each numerical value is labeled CORRECT, INCORRECT, or DONT_KNOW.
  • Human validation: Human annotators evaluated 277 claims blind to system and judge labels, while report comparisons used analyst and investor personas with randomized presentation order.The pairwise protocol produced four judgments per report through two personas and two orderings.

F.1 Insight quality

The insight-quality study compares AnalysisBank with prompting and Buffer of Thoughts through blinded human rankings of reports sampled across performance-gap strata.

  • Sampling: Thirty sampled sets contain AnalysisBank, the best prompting baseline, and Buffer of Thoughts reports across two benchmarks.Fifteen Qwen3-8B instances were sampled per benchmark from top, middle, and bottom insight-rate-gap terciles.
  • Protocol: Two blinded annotators rank each report from most to least insightful without ties, using substantive data-grounded analysis as the criterion.The definition emphasizes causal explanations, comparisons, projections, and implications rather than generic commentary or restated facts.
  • Agreement: Annotators agree on the top-ranked report in 66.7% of sets, versus 33.3% chance, and agree on pairwise preferences in 90.0% of comparisons versus prompting and 70.0% versus Buffer of Thoughts.Agreement on the middle rank is 36.7%, near chance; Cohen’s κ is 0.37 versus prompting and 0.13 versus Buffer of Thoughts.

F.2 Factuality

The factuality evaluation audits generated claims for source support and reports contradiction rates across systems and model backbones.

  • Audit design: The human audit covers 564 numerical and non-numerical claims from AnalysisBank and baselines on Qwen3-8B and DeepSeek-V4-Flash.Claims are labeled supported, contradicted by the source, or unverifiable by a fixed annotator blind to system identity.
  • Results: On DeepSeek-V4-Flash, AnalysisBank’s contradiction rate is 4.7% overall and 9.4% among novel claims, below the pooled baseline rate of 6.6%.Novel-claim baseline rates were skipped because too few such claims were available for a meaningful estimate.
Loading 2609.00818v1…