菜单
Benchmark LLMs on real-world forecasting questions from Metaculus, comparing against human crowds and expert forecasters. Identifies which domains LLMs handle well and where they fall short relative to human intelligence.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ADu2021/skillXiv --skill llm-forecasting-evaluation命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Convert arXiv papers into ready-to-use agent skills using category-aware extraction. First classifies the paper into one or more of 11 research categories, then applies a specialized extraction pipeline for each category — because different types of papers produce different types of usable knowledge. A single paper can yield multiple skills if it spans categories. Use this skill whenever the user wants to turn a paper into a skill, extract practical techniques from research, build a skill library from papers, convert arXiv papers into reusable agent instructions, or batch-process multiple papers into skills. Also trigger when someone asks about extracting actionable knowledge from papers, making research practical for LLM agents, or systematically converting academic contributions into structured agent capabilities.
Establish regret bounds for behavior cloning with discretized actions combining statistical error and quantization error terms. Prove smoothness requirements for safe quantizer design, show that learning-based quantizers fail these requirements, and propose model-based augmentation to reduce error dependence from H² to H.
Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment.
基于 SOC 职业分类
正在显示 SKILL.md
| name | llm-forecasting-evaluation |
| title | Evaluating LLMs on Real-World Forecasting Against Human Superforecasters |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2507.04562 |
| keywords | ["Forecasting","LLM Evaluation","Prediction","Human Comparison","Brier Score"] |
| description | Benchmark LLMs on real-world forecasting questions from Metaculus, comparing against human crowds and expert forecasters. Identifies which domains LLMs handle well and where they fall short relative to human intelligence. |
Evaluating whether LLMs can forecast future events reveals critical gaps in reasoning and planning capabilities. Unlike benchmarks that test memorization or in-distribution reasoning, forecasting requires understanding causal mechanisms, weighing uncertain information, and updating beliefs as new evidence emerges. This framework uses real Metaculus forecasting questions to assess frontier LLMs against human crowds and expert superforecasters—revealing that while frontier models exceed crowd performance, they dramatically underperform human experts.
The challenge is that forecasting is inherently out-of-distribution: answers depend on events that haven't occurred yet, cannot be contaminated in training data, and require genuine reasoning rather than pattern matching.
The evaluation framework operates on three key insights:
Real-world tasks matter: Use actual forecasting questions from Metaculus (a tournament platform with incentive-aligned forecasters) rather than synthetic benchmarks.
Multiple evaluation modes: Compare frontier models under different prompting strategies to understand reasoning failure modes.
Heterogeneous baselines: Compare against both crowd forecasters (wisdom of crowds) and expert superforecasters (domain specialists) to understand capability gaps.
The Brier score (0 = perfect, 1 = worst possible) quantifies forecast accuracy and enables direct comparison across all forecasting attempts.
Set up the evaluation framework by loading questions and preparing news context:
from forecasting_eval.data import MetaculusDataset
from forecasting_eval.context NewsContextualizer
dataset = MetaculusDataset(
date_range=(, ),
n_questions=
)
contextualizer = NewsContextualizer(provider=)
question dataset:
news_articles = contextualizer.fetch_context(
query=question[],
max_articles=
)
question[] = news_articles
Implement direct forecasting evaluation with a frontier LLM:
from forecasting_eval.models import LLMForecaster
from forecasting_eval.metrics import BrierScore
forecaster = LLMForecaster(model="gpt-4o")
scorer = BrierScore()
predictions = []
for question in dataset:
# Use direct prompting: ask for probability estimate
prompt = f"""
Question: {question['title']}
Context: {format_news(question['context'])}
Based on current information, what is the probability this occurs?
Provide a number between 0 and 1.
"""
prediction = forecaster.forecast(prompt) # Returns probability
predictions.append(prediction)
# Calculate Brier score (lower is better)
brier = scorer.calculate(
predictions=predictions,
ground_truth=[q["outcome"] for q in dataset]
)
print(f"Direct forecasting Brier score: {brier:.4f}")
Evaluate using narrative (story-based) prompting to test robustness:
# Narrative prompting: ask model to write reasoning, then extract forecast
predictions_narrative = []
for question in dataset:
prompt = f"""
Consider this question: {question['title']}
Relevant context:
{format_news(question['context'])}
Write your reasoning about what will likely happen, then estimate
the probability as a percentage.
"""
response = forecaster.forecast(prompt)
# Extract probability from narrative response
prob = extract_probability(response)
predictions_narrative.append(prob)
brier_narrative = scorer.calculate(
predictions=predictions_narrative,
ground_truth=[q["outcome"] for q in dataset]
)
print(f"Narrative prompting Brier score: {brier_narrative:.4f}")
Compare LLM performance against human baselines:
from forecasting_eval.baselines import MetaculusBaselines
baselines = MetaculusBaselines()
# Get crowd forecast aggregates (wisdom of crowds)
crowd_predictions = baselines.get_crowd_predictions(dataset)
crowd_brier = scorer.calculate(crowd_predictions, ground_truth)
# Get expert superforecaster predictions
expert_predictions = baselines.get_expert_predictions(dataset)
expert_brier = scorer.calculate(expert_predictions, ground_truth)
print(f"LLM (GPT-4o) Brier: {brier:.4f}")
print(f"Crowd average Brier: {crowd_brier:.4f}")
print(f"Expert superforecasters Brier: {expert_brier:.4f}")
# Analyze performance by domain
domain_analysis = analyze_by_domain(
predictions=predictions,
ground_truth=dataset,
domains=["politics", "economics", "healthcare", "sports"]
)
for domain, score in domain_analysis.items():
print(f"{domain}: {score:.4f}")
Use LLM forecasting evaluation for:
Avoid forecasting for:
| Model | Brier Score | vs. Crowd | vs. Experts |
|---|---|---|---|
| o3 (frontier) | 0.1352 | Better | +11.2pp worse |
| GPT-4o | 0.1540 | Better | +13.2pp worse |
| Claude-3.5-Sonnet | 0.1680 | Similar | +14.6pp worse |
| Human crowd avg | 0.1490 | Baseline | +2.5pp worse |
| Expert superforecasters | 0.0230 | Much better | Baseline |
Critical insight: Frontier LLM o3 beats crowd average but lags expert forecasters by ~11 percentage points of Brier score—suggesting that domain expertise and empirical judgment remain superior to raw reasoning ability.
LLMs typically perform better on:
LLMs typically perform worse on:
| Prompting Strategy | Best For | Pitfall |
|---|---|---|
| Direct probability | Speed and consistency | Anchoring effects |
| Narrative + extraction | Interpretability and verification | False reasoning in stories |
| Structured reasoning | Transparent multi-step reasoning | Length bias toward extremes |
Warning: Narrative prompting can degrade accuracy if models generate plausible-sounding but incorrect reasoning.
"Evaluating LLMs on Real-World Forecasting Against Human Superforecasters" - arXiv:2507.04562