본문 바로가기
  • Home

Year Confusion in Financial RAG: Understanding and Mitigating Wrong-Year Retrieval for Numerical Queries

  • Journal of Software Forensics
  • Abbr : JSF
  • 2026, 22(3), pp.229~239
  • Publisher : Korea Software Assessment and Valuation Society
  • Research Area : Engineering > Computer Science
  • Received : September 2, 2026
  • Accepted : September 20, 2026
  • Published : September 30, 2026

Jeongheon Shon 1,  Hyokyung Chang 1

1한남대

Accredited

ABSTRACT

Retrieval errors in retrieval-augmented generation (RAG) cause generation models to rely on incorrect evidence. Financial filings repeat similar items and comparative figures for the same company across years, allowing filings from other years to outrank the target filing. We defined this as “year confusion” and analyzed target filing ranks by candidate count, company relationship, year distance, and content overlap. With equal candidate counts, BM25 selected the target filing less often under the same-company, adjacent-year condition than under the different-company, same-year condition. We proposed a Temporal reranking framework that incorporates reporting-period relationships between questions and filings into existing retrieval results. On 809 primary evaluation cases, Temporal achieved HR@1 of 0.7157 and MRR@10 of 0.8463, exceeding BM25 by 12.11 percentage points in HR@1. These results show that reporting-period information complements content relevance and improves target filing selection.

Journal Copyright Policy

No CCL information provided

Citation status

* References for papers published after 2025 are currently being built.