• 

  Prediction Accuracy by Category

PredictPulse — AI-Powered Prediction Market Intelligence

ZerveHack 2026 | Data Science / ML / AI Track | April 2026

The Problem

Prediction markets handle $1B+ in annual volume across platforms like Polymarket, Kalshi, and Metaculus. Yet 68% of participants lack tools to assess whether a forecast is actually reliable. Current approaches treat all predictions equally — ignoring the rich metadata signals that separate trustworthy forecasts from noise.

Our Solution

PredictPulse is an end-to-end data science pipeline built on Zerve that answers: "Can we predict which prediction market forecasts will be most accurate?"

By cross-referencing prediction market metadata with economic indicators and social signals, PredictPulse trains an ML model to score the reliability of any active prediction — then deploys it as a live API with AI-powered explanations.

Key Features

1. Cross-Platform Intelligence — Ingests 500+ resolved Metaculus predictions, correlates with FRED economic indicators and social trend data to identify accuracy patterns across domains
2. ML Accuracy Scorer — Gradient Boosting classifier trained on 20+ engineered features (participation density, confidence levels, question complexity, economic context) achieves AUC-ROC above baseline
3. Deployed API with AI Analysis — Live API endpoint accepts any prediction question and returns a reliability score, confidence tier, top contributing factors, and a Claude-generated natural language explanation

Architecture

graph TD
    A[Metaculus API] -->|500+ questions| B[Data Collection]
    C[FRED API] -->|Economic indicators| B
    D[Google Trends] -->|Social signals| B
    B --> E[Feature Engineering]
    E -->|20+ features| F[Model Training]
    F -->|GBM Classifier| G[Reliability Scorer]
    G --> H[Claude AI Analysis]
    H --> I[Deployed API]
    I -->|JSON response| J[User/Application]

Pipeline Walkthrough

Quick Start (Zerve Platform)

1. Create a Zerve account at zerve.ai
2. Create a new Canvas and add Python blocks
3. Copy each block from src/01_* through src/06_* in order
4. Set environment variables:
   • ANTHROPIC_API_KEY — for Claude AI analysis (optional, has fallback)
   • FRED_API_KEY — for economic indicators (optional)
5. Run blocks sequentially — each builds on the previous
6. Deploy Block 6 as an API endpoint via Zerve Deployment

Tech Stack

API Usage

# POST to deployed Zerve API endpoint
request = {
    "title": "Will global temperature exceed 1.5°C before 2030?",
    "community_prediction": 0.62,
    "prediction_count": 245,
    "description_length": 1200,
    "num_comments": 48,
    "question_age_days": 180,
    "category": "Science"
}

# Response
{
    "reliability_score": 0.82,
    "reliability_tier": "High",
    "analysis": "This Science prediction has high reliability...",
    "top_factors": [...],
    "metadata": {"model": "GradientBoosting", ...}
}

Key Findings

• Participation density (predictions per day) is the strongest predictor of accuracy
• Extreme predictions (>90% or <10%) are less reliable than moderate ones
• Question complexity (description length) correlates positively with accuracy
• Older questions with sustained engagement show higher reliability
• Economic volatility periods reduce prediction accuracy across all categories

Team

• Built with Zerve AI, Claude API, and a passion for turning crowd wisdom into actionable intelligence

Built With

• anthropic
• api
• claude
• fastapi
• fred
• fred-api
• machine-learning
• metaculus
• numpy
• pandas
• plotly
• prediction-markets
• python
• scikit-learn
• zerve