Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FinCARDS: Card-Based Analyst Reranking for Financial Document Question Answering

Yixi Zhou, Fan Zhang, Yu Chen, Haipeng Zhang, Preslav Nakov, Zhuohan Xie

TL;DR

FinCARDS tackles the problem of financial QA over long filings by reframing evidence selection as a constraint-satisfaction task under a finance-aware schema. It introduces a Card Abstraction that converts document chunks into structured fields and a three-stage, zero-shot tournament reranking (Stage 1: BM25; Stage 2: Card-based screening; Stage 3: bootstrap listwise stabilization) to produce auditable, stable evidence rankings without model fine-tuning. Across FinAgentBench, the approach yields substantial gains in early-rank metrics (nDCG@10, MRR@10) and reduces candidate sets, while providing transparent audit traces and predictable costs. The work demonstrates that structured intermediate representations and stability-aware decision procedures can markedly improve reliability in high-stakes financial document QA and suggests broader applicability of such schemas in long-form retrieval tasks.

Abstract

Financial question answering (QA) over long corporate filings requires evidence to satisfy strict constraints on entities, financial metrics, fiscal periods, and numeric values. However, existing LLM-based rerankers primarily optimize semantic relevance, leading to unstable rankings and opaque decisions on long documents. We propose FinCards, a structured reranking framework that reframes financial evidence selection as constraint satisfaction under a finance-aware schema. FinCards represents filing chunks and questions using aligned schema fields (entities, metrics, periods, and numeric spans), enabling deterministic field-level matching. Evidence is selected via a multi-stage tournament reranking with stability-aware aggregation, producing auditable decision traces. Across two corporate filing QA benchmarks, FinCards substantially improves early-rank retrieval over both lexical and LLM-based reranking baselines, while reducing ranking variance, without requiring model fine-tuning or unpredictable inference budgets. Our code is available at https://github.com/XanderZhou2022/FINCARDS.

FinCARDS: Card-Based Analyst Reranking for Financial Document Question Answering

TL;DR

FinCARDS tackles the problem of financial QA over long filings by reframing evidence selection as a constraint-satisfaction task under a finance-aware schema. It introduces a Card Abstraction that converts document chunks into structured fields and a three-stage, zero-shot tournament reranking (Stage 1: BM25; Stage 2: Card-based screening; Stage 3: bootstrap listwise stabilization) to produce auditable, stable evidence rankings without model fine-tuning. Across FinAgentBench, the approach yields substantial gains in early-rank metrics (nDCG@10, MRR@10) and reduces candidate sets, while providing transparent audit traces and predictable costs. The work demonstrates that structured intermediate representations and stability-aware decision procedures can markedly improve reliability in high-stakes financial document QA and suggests broader applicability of such schemas in long-form retrieval tasks.

Abstract

Financial question answering (QA) over long corporate filings requires evidence to satisfy strict constraints on entities, financial metrics, fiscal periods, and numeric values. However, existing LLM-based rerankers primarily optimize semantic relevance, leading to unstable rankings and opaque decisions on long documents. We propose FinCards, a structured reranking framework that reframes financial evidence selection as constraint satisfaction under a finance-aware schema. FinCards represents filing chunks and questions using aligned schema fields (entities, metrics, periods, and numeric spans), enabling deterministic field-level matching. Evidence is selected via a multi-stage tournament reranking with stability-aware aggregation, producing auditable decision traces. Across two corporate filing QA benchmarks, FinCards substantially improves early-rank retrieval over both lexical and LLM-based reranking baselines, while reducing ranking variance, without requiring model fine-tuning or unpredictable inference budgets. Our code is available at https://github.com/XanderZhou2022/FINCARDS.
Paper Structure (49 sections, 9 equations, 6 figures, 5 tables)

This paper contains 49 sections, 9 equations, 6 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Financial QA benchmarks and evidence structure.
  4. Retrieval and LLM reranking.
  5. Agentic reasoning and human review processes.
  6. FinCARDS
  7. Overview
  8. Card Abstraction
  9. Why Card Abstraction.
  10. Motivation.
  11. Analyst analogy.
  12. Formal definition.
  13. Core fields.
  14. Auxiliary fields.
  15. Query Intent Mapping
  16. ...and 34 more sections

Figures (6)

  • Figure 1: Key challenge in financial QA. Reranking must satisfy the correct metric and fiscal period (often numeric), not just semantic relevance. The illustration is based on U.S. SEC corporate financial filings and shows typical failure modes.
  • Figure 2: Overview of the FinCARDS pipeline. From a SEC filing and a user question, the system constructs structured Cards, a structured query intent, and a tournament reranking module that produces the final Top-$k$ evidence chunks.
  • Figure 3: Excerpt of the prompt used to instantiate a Chunk Card (full schema).
  • Figure 4: Excerpt of the prompt used to instantiate Query Intents.
  • Figure 5: Excerpt of the Stage 2 batch selection prompt that we used in order to perform group-wise evidence filtering.
  • ...and 1 more figures