FinQA

Goal 

The FinQA dataset is designed to advance numerical reasoning over financial documents. Its primary goals are:

To enable multi-step numerical reasoning over financial reports
To support program synthesis for answering complex financial questions
To benchmark quantitative reasoning in financial AI

Description 

FinQA is a large-scale dataset for numerical reasoning over financial documents, requiring models to synthesize reasoning programs to answer questions from earnings reports.

Dataset Composition 

Source: Earnings reports from S&P 500 companies
Time Period: Reports from 2016-2019
Sample Size: 8,281 question-answer pairs
Document Coverage: Mix of tables and text from quarterly and annual reports

Example 

Context: Financial table showing revenue and expenses

Question: “What is the change in operating margin from 2018 to 2019?”

Answer: “5 percentage points” (requires multi-step calculation)

Task Description 

Financial Numerical Reasoning 

Input: Question, financial document (tables and text)
Output: Numerical answer with reasoning program
Challenge: Requires extracting relevant facts, understanding financial concepts, and performing multi-step numerical operations

Key Use Cases 

Automated Financial Analysis: Extracting quantitative insights from earnings reports
Investment Research: Answering analyst questions about company performance
Financial Education: Interactive learning systems for financial statement analysis

Evaluation Metrics 

Execution Accuracy: Percentage of questions where the predicted program generates the correct answer
Program Accuracy: Percentage of questions where the predicted program exactly matches the gold program

Zhiyu Chen, Wenhu Chen, Charese Smiley, Sameena Shah, Iana Borova, Dylan Langdon, Reema Moussa, Matt Beane, Ting-Hao Huang, Bryan Routledge, and William Yang Wang. “FinQA: A Dataset of Numerical Reasoning over Financial Data.” In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2021.

For dataset access, visit the official repository or HuggingFace.