2 of 7

Probabilistic forecasting

The Problem

We want to be able to chart out the future.

Example: implementing rent control policies, what if the president dies, what if WW3 breaks out.

Our Vision

Create an environment that models the true probabilistic distribution of future events

3 of 7

PsychoHistory's Architecture

Solution

Probabilistic Forecasting

Research Agent

Gathers information on each potential future node or branch

Reasoning Model

Uses information to output accurate, optimized probabilities

Probability Output

Shows predictions for each outcome with confidence levels

4 of 7

Training for Truth: Supervised Fine-Tuning

Objective

Minimize difference between probability outputs by the tree and the trace of events that happened in real life

Data Curation

Real-World Events

Events that occurred outside the model's training data (GPT OSS so past January 2024)

Limited Search Scope

Model's research limited to data before the event occurred

Training Process

Methodology

Take model predictions at each node and measure similarity between the correct path probability and actual outcome

Scale

Trained on hundreds of curated examples of research, events, and outcomes (gathered by a research pipeline with DeepseekV3.1 and perplexity

Results

Before

Even Probabilities

— Essentially guesswork

After

Non-Uniform Distribution — Model excels at predicting future events

5 of 7

Reinforcement Learning: DPO

Method

Direct Preference Optimization (DPO) Pipeline

We deploy a DPO pipeline to force the model's end outputs to abide by binary constraints.

Goal

Binary Convergence

Force the model to output 0 or 1 for predefined market outcomes instead of continuous probabilities.

Impact

Market-Ready Predictions

Convergence to one of two outcomes

Direct actionability for prediction markets

Trained on hundreds of pairwise preferences - curves all predictions to the binary

outcomes

▸

Problem: Even with prompting we do not get binary convergence

There's currently a presidential election. Kamala versus Trump. Imagine we are in the phase 6 months before the election actually starts. Generate a tree for the next 6 months with events such that every ending node at depth 3 ends with either a kamala won or trump won

6 of 7

From Prediction to Profit

Capability

Our trained model excels at outputting future event probabilities.

Money-Making Track

Make Money on Prediction Markets

The Challenge: The model needs to convert continuous probability outputs to binary (0 or 1) for predefined market outcomes.

7 of 7

The Future is Predictable

PsychoHistory represents a breakthrough in probabilistic forecasting—a system trained to predict the future with high accuracy.

What We've Built

A machine learning system that transforms LLMs from outputting uniform guesses to generating precise, non-uniform probability distributions for future events through supervised fine-tuning and reinforcement learning.

Next Steps

Expand datasets with real-world events • Integrate with live prediction market APIs • Deploy DPO pipeline for binary market outcomes

The future of forecasting