| 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. |
LLM Forecasting Evaluation: Assessing Predictive Reasoning on Out-of-Distribution Tasks
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.
Core Concept
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.
Architecture Overview
- Question dataset: 464 real forecasting questions from July-December 2024 spanning politics, economics, healthcare, sports, technology
- Information pipeline: Integration with AskNews to provide relevant news articles for each question
- Multiple model evaluation: 12 frontier models including GPT-4o, o3, Claude variants, and DeepSeek
- Prompting strategies: Direct forecasting vs. narrative (script-based) prompting to test robustness
- Aggregation method: 5 predictions per question, averaged for stability
- Human baselines: Crowd forecaster aggregates and individual expert predictions from Metaculus
Implementation
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
