FrontierFinance

FrontierFinance benchmark (paper): 25 long-horizon financial modeling tasks requiring ~18 hours of human expert labor per task.

About the benchmark

Each task contains a detailed specification of a deliverable, often a spreadsheet. The model is expected to work autonomously to complete the task. It must make sound modeling assumptions, retrieve and analyze data, and finally create deliverables that follow strict formatting guidelines that mimic human professional standards.

Human experts took 18.3 hours per task on average, ranging from 6.5 to 35 hours.

Task types

Five tasks per type. We had financial professionals complete two of the five tasks and recorded their time and rubric scores.

Type	Description
three_statement	Income statement, balance sheet, and cash flow linked into one internally consistent framework, then forecast forward.
merger_model	Combine two companies and determine EPS accretion or dilution given deal structure, synergies, and adjustments.
lbo	Debt-financed acquisition: operating projections, debt schedule, investor returns (IRR, MoM). Plus a PowerPoint memo.
dcf	Forecast free cash flows plus a terminal value and discount to present. Revenue growth, margins, capex, cost of capital.
lender	Credit analysis: repayment capacity, covenant compliance, coverage and leverage ratios under stress scenarios.

Each task prompt is also mirrored as a markdown file at prompts/{eval_type}/{sample_id}.md for reading without loading the parquet.

The following is the baseline data from the paper.

Usage

The dataset comes with an eval.yaml so it can run from the inspect CLI. However, it is bring-your-own-harness. We provide a very simple baseline harness similar to the one demonstrated in the paper. The sample solver.py contains the following:

from pathlib import Path
from inspect_ai.log._samples import sample_active
from inspect_ai.solver import chain, generate, solver, system_message, use_tools
from inspect_ai.tool import bash, computer, tool
from inspect_ai.util import sandbox

@tool(prompt="Submit files from /workspace for evaluation.")
def submit():
    async def execute(files: list[str]):
        """Submit completed financial model files.

        Args:
            files: Absolute paths to submit, e.g. ["/workspace/model.xlsx"]
        """
        active = sample_active()
        out = Path(active.log_location).with_suffix("") / str(active.sample.id)
        out.mkdir(parents=True, exist_ok=True)
        for f in files:
            (out / Path(f).name).write_bytes(await sandbox().read_file(f, text=False))
        return f"✓ Submitted {len(files)} files"
    return execute

@solver
def computer_use():
    return chain([
        system_message("""You are an expert investment banking associate building a financial model.

You have access to a computer with helpful tools (Firefox, LibreOffice, python3).
- **computer**: Control GUI desktop
- **bash**: Run shell commands
- **submit**: Submit your completed work when done

IMPORTANT:
- NEVER ask questions. Proceed autonomously.
- Save all work to /workspace/ (include README.md with assumptions)
- When FULLY COMPLETE, call submit() to save your work
- After submit confirms, respond with [DONE].

Work systematically. Expect 8+ hours of work."""),
        use_tools([computer(), bash(timeout=180), submit()]),
        generate(),
    ])

It is meant to be used with the following docker configuration.

Dockerfile:

FROM aisiuk/inspect-computer-tool:latest

USER root

RUN apt-get update && apt-get install -y --no-install-recommends \
    libreoffice-calc \
    libreoffice-common \
    python3-uno \
    procps \
    curl \
    jq \
    poppler-utils \
    && rm -rf /var/lib/apt/lists/*

RUN rm -rf /usr/lib/python3/dist-packages/numpy* && \
    pip install --no-cache-dir \
    aiohttp \
    beautifulsoup4 \
    openpyxl \
    numpy \
    pandas \
    lxml \
    pypdf \
    html5lib \
    xlrd \
    xlsxwriter \
    python-pptx

RUN mkdir -p /workspace /workspace/docs && \
    chown -R user:user /workspace

# LibreOffice Calc shortcut on the desktop
RUN cp /usr/share/applications/libreoffice-calc.desktop /home/user/Desktop/ && \
    chmod +x /home/user/Desktop/libreoffice-calc.desktop && \
    chown user:user /home/user/Desktop/libreoffice-calc.desktop

# Screensaver blanks the screen mid-run and the model loses the GUI
RUN mkdir -p /home/user/.config/autostart && \
    echo '[Desktop Entry]' > /home/user/.config/autostart/xfce4-screensaver.desktop && \
    echo 'Hidden=true' >> /home/user/.config/autostart/xfce4-screensaver.desktop && \
    chown -R user:user /home/user/.config/autostart

# Session restore pops a recovery dialog over the browser on restart
RUN mkdir -p /home/user/.mozilla/firefox-esr/profile.default && \
    echo 'user_pref("browser.sessionstore.resume_from_crash", false);' >> /home/user/.mozilla/firefox-esr/profile.default/user.js && \
    echo 'user_pref("browser.startup.page", 0);' >> /home/user/.mozilla/firefox-esr/profile.default/user.js && \
    chown -R user:user /home/user/.mozilla

ENV PYTHONPATH="/usr/lib/python3/dist-packages:/app"
ENV DISPLAY=:1

USER user
WORKDIR /workspace

compose.yaml:

services:
  default:
    build: .
    init: true

You can run it with:

inspect eval hf/kensho/frontierfinance --solver solver.py@computer_use --sandbox docker:compose.yaml

The sample submit tool requires inspect_ai>=0.3.91.

Output files

submit(files=[...]) copies the named files into a directory named after the eval log, one subdirectory per sample:

logs/2026-08-20T20-08-35-00-00_..._RSJGKLv2uWnAX8VNSVfrno.eval    <- the log
logs/2026-08-20T20-08-35-00-00_..._RSJGKLv2uWnAX8VNSVfrno/
├── 1/Enerpac_Model.xlsx
├── 1/README.md
├── 2/Merck_model.xlsx
└── 2/README.md

Scoring

There is currently no automated scorer available. Rubrics live in the dataset under config: rubrics (also as CSVs in rubrics/) but applying them requires custom code.

Inspect's eval.yaml schema requires at least one scorer, so the task declares includes as a placeholder. It reports a constant accuracy 0.000 regardless of output.

Each type has 5-11 evaluation categories and up to 141 criteria, 491 in total. Rubrics are calibrated to remediation effort, so mistakes needing substantial rework cost far more than easily fixed ones.

Citation

@misc{krumdick2026frontierfinancelonghorizoncomputerusebenchmark,
      title={FrontierFinance: A Long-Horizon Computer-Use Benchmark of Real-World Financial Tasks}, 
      author={Michael Krumdick and Varshini Reddy and Shivani Chaudhary and William Day and Maarij Ahmed and Hayan Haqqi and Muhammad Ahsen Fahim and Hanzallah Amjad and Ahmad Orakzai and Aqsa Gul and Chris Tanner},
      year={2026},
      eprint={2604.05912},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2604.05912}, 
}