File size: 6,588 Bytes
65e805f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
980a808
65e805f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ef80786
65e805f
ef80786
65e805f
 
ef80786
65e805f
 
 
 
 
ef80786
 
 
 
65e805f
 
 
 
 
 
 
 
 
 
 
ef80786
65e805f
 
 
 
 
 
 
 
 
 
 
 
 
 
980a808
 
 
 
 
 
 
65e805f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
980a808
65e805f
 
 
 
 
 
ef80786
65e805f
 
 
ef80786
65e805f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ef80786
65e805f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ef80786
 
 
 
 
 
 
 
65e805f
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
---
license: cc-by-nc-4.0
task_categories:
  - question-answering
  - text-generation
language:
  - en
tags:
  - finance
  - agent
  - deep-research
  - point-in-time
  - benchmark
  - rubric-evaluation
pretty_name: FinanceGym
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: test
        path: benchmark_400_public.parquet
dataset_info:
  features:
    - name: task_id
      dtype: string
    - name: question
      dtype: string
    - name: cutoff
      dtype: string
  splits:
    - name: test
      num_examples: 400
---

# FinanceGym — Point-in-Time Finance Deep-Research Benchmark (400 questions)

**FinanceGym** measures whether a finance research agent can produce a grounded, cited
research report *under a point-in-time constraint*. Every question carries a `cutoff` date,
and an agent may use **only** information published on or before that date.

Each question is graded against expert-annotated rubrics on two axes:

- **Pre-cutoff** — did the report correctly ground itself in the evidence that existed *before*
  the cutoff?
- **Post-cutoff** — did it correctly anticipate what actually happened *after* the cutoff?

The split between those two axes is the point of the benchmark. Across every system evaluated
so far, post-cutoff scores land far below pre-cutoff — anticipating consequences from pre-cutoff
evidence alone remains the open problem.

This repository holds the **public 400-question slim set**: the questions and their cutoffs.
**Rubrics are withheld** (see [Rubrics are not public](#rubrics-are-not-public) below).

- 💻 [Code and environment](https://github.com/google-research/google-research/tree/master/finance_harness/FinanceGym)
- 🏆 [Leaderboard](https://financegym.github.io/#leaderboard) — current standings
- 📥 [Participate](https://financegym.github.io/#submit) — run your agent and submit a report
- 📄 [Paper](https://arxiv.org/abs/2607.27853) — FinanceHarness: Autonomous Financial Deep Research Framework

## Data Statistics

| | |
|---|---|
| Questions | 400 |
| Split | `test` (single split) |
| Language | English |
| Cutoff range | 2024-12-31 → 2025-11-10 |
| Question length | 32–48 words (median 41) |
| Withheld rubric items | 2,464 total · 4–8 per question (median 6) |
| Rubric axes | 1,496 pre-cutoff · 968 post-cutoff |

Cutoff dates by month:

| Month | Questions | | Month | Questions |
|---|---:|---|---|---:|
| 2024-12 | 3 | | 2025-06 | 60 |
| 2025-01 | 11 | | 2025-07 | 63 |
| 2025-02 | 10 | | 2025-08 | 57 |
| 2025-03 | 20 | | 2025-09 | 29 |
| 2025-04 | 66 | | 2025-10 | 22 |
| 2025-05 | 52 | | 2025-11 | 7 |

## Data Format

The same 400 records ship in two identical forms:

| File | Use |
|---|---|
| `benchmark_400_public.jsonl` | Canonical raw file — byte-identical to the copy in the code repository. |
| `benchmark_400_public.parquet` | What `load_dataset` and the dataset viewer read. Same rows, all three fields typed as strings. |

One JSON object per line in `benchmark_400_public.jsonl`.

| Field | Type | Description |
|---|---|---|
| `task_id` | string | Stable unique id for the question. |
| `question` | string | The research question. Used **verbatim** to match a submission to its rubric during grading. |
| `cutoff` | string (`YYYY-MM-DD`) | The point-in-time date. An agent may use only information available **on or before** this date. |

Example line:

```json
{
  "task_id": "69f904b728538874c086db16",
  "question": "How does Air France-KLM's accelerated transition toward a 60.5% majority stake in SAS impact its strategic bandwidth and capital allocation for the privatization of TAP Air Portugal, given the group's 2026 full-control target and ongoing European airline consolidation?",
  "cutoff": "2025-08-05"
}
```

Load it with `datasets`:

```python
from datasets import load_dataset

ds = load_dataset("Rujun/FinanceGym", split="test")
print(ds[0]["question"], ds[0]["cutoff"])   # cutoff is a "YYYY-MM-DD" string
```

## The Point-in-Time Rule

Agents retrieve evidence from a frozen news corpus served behind a search API that enforces a
`max_date` filter — the server never returns documents published after the cutoff. That single
hard rule is what makes the post-cutoff axis meaningful; you may otherwise run any agent
architecture you like.

For each question, pass its `cutoff` as `max_date` on every search call, gather evidence, and
write a cited report.

## Rubrics Are Not Public

The scoring rubrics are **not** part of this dataset. Grading is run by maintainers against a
private full-rubric set, which keeps evaluation single-blind — no tuning to the rubric is
possible, and leaderboard rows stay comparable across systems.

## How Submissions Are Graded

A judge scores every rubric item on a **0–4** scale:

| Score | Meaning |
|---:|---|
| 0 | NOT ADDRESSED |
| 1 | MENTIONED (no substance) |
| 2 | PARTIAL (directionally correct, missing specifics) |
| 3 | SUBSTANTIVE (correct with specifics, minor gaps) |
| 4 | FULLY GROUNDED (correct, specific, plausibly sourced) |

The headline metric is **macro-averaged**: each question's normalized rate is `s / (4n)` over its
`n` rubric items, and the score is the mean of those per-question rates, so every question is
weighted equally. **Pre-cutoff** and **post-cutoff** are the same macro mean restricted to
antecedent and consequent rubric items respectively, reported with a 95% bootstrap CI.

## License

This dataset is licensed under the
[Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0)](https://creativecommons.org/licenses/by-nc/4.0/).

**Non-commercial use only.** FinanceGym, its benchmark datasets, and retrieval tools are
provided strictly for non-commercial research and evaluation. Commercial use — including
integration into commercial financial advisory services, or using generated outputs to provide
commercial investment advice — is strictly prohibited.

**Not investment advice.** The questions and any model outputs derived from them are research
artifacts, not financial advice.

## Citation

```bibtex
@misc{xiao2026financeharnessautonomousfinancialdeep,
      title={FinanceHarness: Autonomous Financial Deep Research Framework},
      author={Yijia Xiao and Rujun Han and Yanfei Chen and Zifeng Wang and Ke Jiang and Zhongying CuiZhu and Vishy Tirumalashetty and Wei Wang and Burak Gokturk and Tomas Pfister and Chen-Yu Lee},
      year={2026},
      eprint={2607.27853},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2607.27853},
}
```