OracleProto Forecasting Benchmark

300 real-world questions. A fixed score. A declared information boundary.

GitHub · Leaderboard · Paper · 中文

OracleProto is a held-out forecasting benchmark of 300 resolved events, dated 2026-03-12 to 2026-09-21. It combines model knowledge cutoffs, time-masked retrieval and a separate retrieval-content auditor. The released test split matches the fixed question set used by the complete 27-configuration evaluation.

Dataset at a glance

Question type	Rows	Points per question	Available points
Yes/no	100	1	100
Named binary	9	1	9
Single-answer multiple choice	154	2	308
Multi-answer multiple choice	37	3	111
Total	300		528

The dataset contains evaluation questions only. The source retains the fixed 80-question anchor and 220 additional questions; their source configurations remain separate strata in the experimental analysis. The repository's bundled example is a different 300-question set. Use this release for the published paper results.

Load the test split

from datasets import load_dataset
import json

ds = load_dataset("MaYiding/OracleProto", split="test")
assert len(ds) == 300
sample = dict(ds[0])
sample["options"] = json.loads(sample["options"])
print(sample)

The dataset viewer and load_dataset use the CSV explicitly assigned to the test split. options remains a JSON string for SQLite/CSV compatibility. The full database, including the stored prompt recipe, is also available:

from huggingface_hub import hf_hub_download
import sqlite3

path = hf_hub_download("MaYiding/OracleProto",
    "forecast_eval_set_example.db", repo_type="dataset")
with sqlite3.connect(f"file:{path}?mode=ro", uri=True) as conn:
    rows = conn.execute("SELECT * FROM test_cases").fetchall()
    assert len(rows) == 300

Scoring

Score is the percentage of the 528 available points earned. Yes/no and named binary questions earn 1 point for an exact answer, and single-answer multiple choice earns 2. Multi-answer multiple choice earns 3 × 2TP / (2TP + FP + FN). Extra and missed selections both reduce credit; an exact set earns all 3 points.

Average each question's earned points over three repetitions, sum those averages and divide by 528, then multiply by 100. Invalid answers and model refusals earn zero without reducing the denominator. Missing samples or infrastructure errors make official Score unavailable. Accuracy, repeated-answer agreement and cost are supporting diagnostics. Oracle best-of-three results are not a deployable answer-selection method.

The manifest records the scoring contract, file hashes and 2,000-replicate stratified question bootstrap settings. Complete scores include all 27 configurations; type-level scores expose the four question types. Scores use a 0–100 scale, while accuracy and agreement CSV fields use 0–1. Cost fields reprice forecasting inference at public tariffs and divide three-repetition totals by three; search, detector usage and invoice totals are excluded.

Replay the information boundary

Use the OracleProto harness, with SOURCE_TABLE=test_cases, PROMPT_TEMPLATE_STYLE=shared, three repetitions, at most six ReAct steps, four searches per sample and five results per search. Declare each model's cutoff and preserve the route, reasoning parameters, prompt hash and detector settings. Provider defaults, explicit effort and ON/OFF switches are distinct configurations.

The SQLite table dataset_metadata contains features_json.prompt_reconstruction; prompt_reconstruction.json exports the same recipe. Its strict-answer instructions are preserved exactly as collected. The scorer's multi-answer F1 partial-credit rule is specified separately above. The recipe can be reproduced by the harness's shared-template renderer without changing stored prompt text.

Resolution dates have day precision and are read as GMT+8 in the prompt. A calendar cutoff is not proof that an event's outcome was unavailable at retrieval time. Retain raw retrieval evidence and check result publication times when interpreting prospective forecasting.

Leakage audit

The final author-reported audit uses GPT 6.1 Sol's 137 leakage labels out of 10,000 reviewed filtered results: 1.37%. Human reviewers checked all 137 labelled positives and confirmed all leaked. They also checked 1,000 labelled negatives and found no leakage. The negative check is a sample, so it does not establish overall detector recall. This is an audit of retrieved content, not of model training data.

Files, schema and provenance

File	Contents
forecast_eval_set_example.db	300 rows in test_cases, plus dataset_metadata
forecast_eval_set_example.csv	The same seven columns and 300 rows, used by the viewer
prompt_reconstruction.json	Stored shared prompt recipe
dataset_manifest.json	Counts, checksums and evaluation contract
results/	Complete-configuration and type-level aggregate results

Field	Meaning
id	Stable question ID
choice_type	single or multi
question_type	yes_no, binary_named or multiple_choice
event	Event to forecast
options	JSON-encoded option array; A is index 0, B is index 1
answer	Gold letter or comma-separated letter set
end_time	Resolution date in YYYY-MM-DD

The questions derive from Futurex-Past, with closed answer options and normalized letter-set answers. All 80 anchor rows retain their existing fields and gold answers. The release is limited to resolved, closed-choice events and does not represent all domains or forecasting horizons. Declared model cutoffs are operational assumptions, not independently verified training-corpus boundaries.

Citation and contact

@article{OracleProto,
  title={OracleProto: A Reproducible Framework for Benchmarking LLM Native Forecasting via Knowledge Cutoff and Temporal Masking},
  author={Yiding Ma, Chengyun Ruan, Kaibo Huang, Zhongliang Yang, Linna Zhou},
  journal={arXiv preprint arXiv:2605.03762},
  year={2026}
}

• Yiding Ma: yidingma@bupt.edu.cn
• Chengyun Ruan: ruanchengyun815@bupt.edu.cn
• Kaibo Huang: huangkaibo@bupt.edu.cn
• Zhongliang Yang: yangzl@bupt.edu.cn

Released under the MIT license.