Datasets:

nogabenyoash
/

SecQue

license: mit
task_categories:
  - question-answering
language:
  - en
tags:
  - finance
  - table-text
  - numerical_reasoning
size_categories:
  - < 1K

SECQUE

Paper

SECQUE is a comprehensive benchmark for evaluating large language models (LLMs) in financial analysis tasks.

SECQUE comprises 565 expert-written questions covering SEC filings analysis across four key categories:

comparison analysis
ratio calculation
risk assessment
financial insight generation.

To assess model performance, we develop SECQUE-Judge, an evaluation mechanism leveraging multiple LLM-based judges, which demonstrates strong alignment with human evaluations. Additionally, we provide an extensive analysis of various models’ performance on our benchmark.

Results

Model	Baseline	Financial	Baseline CoT	Financial CoT	Flipped	Avg Tokens by Model
GPT-4o	*0.69*/0.79	0.62/0.71	0.67/0.76	0.63/0.73	0.68/0.78	319.84
GPT-4o-mini	0.64/0.73	0.38/0.47	0.60/0.72	0.56/0.65	0.62/0.73	289.76
Llama-3.3-70B-Instruct	0.65/0.75	0.60/0.71	0.63/0.74	0.60/0.72	0.62/0.74	341.63
Qwen2.5-32B-Instruct	0.61/0.72	0.49/0.58	0.60/0.71	0.55/0.67	0.65/0.75	331.34
Phi-4	0.56/0.66	0.55/0.64	0.57/0.67	0.56/0.66	0.57/0.67	294.33
Meta-Llama-3.1-8B-Instruct	0.48/0.60	0.41/0.54	0.44/0.56	0.40/0.53	0.47/0.59	338.38
Mistral-Nemo-Instruct-2407	0.46/0.55	0.32/0.42	0.45/0.56	0.44/0.55	0.44/0.54	231.52
Avg Tokens by Prompt	283.04	151.97	437.38	334.71	317.57	304.93

Citation

@inproceedings{
    title = "SECQUE: A Benchmark for Evaluating Real-World Financial Analysis Capabilitiese",
    author = "Ben Yoash, Noga  and
      Brief, Meni and
      Ovadia, Oded  and
      Shenderovitz, Gil  and
      Mishaeli, Moshik  and
      Lemberg, Rachel  and
      Sheetrit, Eitam",
    month = apr,
    year = "2025",
    url = "https://arxiv.org/pdf/2504.04596",
}

Evaluation Benchmark notice

This benchmark is indented solely for evaluation, and must not be used for training in any way.