Binary Forecaster

AI-powered forecasting pipeline for binary (yes/no) questions. Uses multiple LLMs to research a question from every angle, then aggregates their forecasts into a single probability.

How it works

The pipeline runs 5 stages:

1. Fetch — Loads and validates your input CSV
2. Research — Dispatches 6 AI research agents in parallel to gather: current state of the world, base rates, key factors, market/expert forecasts, YES thesis, and NO thesis
3. Forecast — Runs 6 LLM forecasters (Gemini, Claude, GPT at two temperature settings each) that synthesize the research into probability estimates
4. Aggregate — Takes the median probability across all models and selects the best rationale
5. Display — Prints results and exports the output CSV

Setup

pip install -r requirements.txt

Set your API key:

export EVERYROW_API_KEY="your-key-here"

Usage

Create a CSV file with a question column:

question
Will the US and Iran agree to a ceasefire before May 2026?
Will global average temperature exceed 1.5°C above pre-industrial levels in 2026?

Optional columns: resolution_criteria, background, fine_print

Run the forecaster:

python run.py questions.csv

The output CSV (in output/) contains two columns: probability and rationale.

Run specific stages

python run.py questions.csv fetch        # Just validate the CSV
python run.py questions.csv research     # Load + research only
python run.py display                    # View results from last run

Output

The final CSV is saved to output/forecast_YYYYMMDD_HHMMSS.csv with columns:

Column	Description
probability	Aggregated forecast (0-100%)
rationale	Selected rationale explaining the forecast

Intermediate stage files are also saved in output/ for debugging.