I’m Building a Prediction-Market Bot — Here’s What Information Theory Taught Me
I've been experimenting with automated prediction-market systems and building a Python-based Polymarket Trading bot.
At first, I thought the core problem was straightforward:
Find a market where my probability estimate is different from the market price, then trade.
But the deeper I went, the more I realized that probability isn't the whole problem.
The harder question is:
How much should I trust the information behind my probability estimate?
The Problem With “Just Find the Edge”
Imagine:
Market probability: 60%
My model: 70%
It looks like a 10% edge.
But what if my model is uncertain?
Maybe the real probability could reasonably be:
58% — 70% — 78%
Suddenly, that 10% edge doesn't look so reliable.
That's where concepts from information theory become interesting.
Thinking About Information, Not Just Prices
Instead of asking:
“Did the market move?”
I'm increasingly interested in:
“What information caused the market to move?”
A price change could come from:
- genuine new information
- low liquidity
- a large order
- speculation
- temporary noise
- correlated markets
A bot that reacts to every movement can easily overtrade.
My goal is to build a system that tries to distinguish information from noise.
Where Entropy Fits
For binary markets, entropy provides a simple way to think about uncertainty.
50% Yes / 50% No
↓
High uncertainty
95% Yes / 5% No
↓
Low uncertainty
This doesn't automatically tell us whether to trade.
Instead, it gives another dimension to consider alongside:
Probability
Market price
Confidence
Liquidity
Execution cost
Risk
The Architecture I'm Experimenting With
My current mental model looks something like:
Market Data
↓
Information
↓
Probability Model
↓
Uncertainty
↓
Market vs Model
↓
Risk Management
↓
Position Sizing
↓
Execution
The interesting part isn't necessarily the order execution.
It's everything that happens before the order.
What I'm Learning From Building It
One of the biggest lessons so far:
A trading bot being technically correct doesn't mean the strategy is statistically correct.
There are many ways to build something that successfully places trades but still has a poor strategy.
I'm paying much more attention to:
- probability calibration
- uncertainty
- liquidity
- slippage
- position sizing
- correlated positions
- model errors
- risk limits
My open-source experiment is here:
https://github.com/Benjam1nCup/Polymarket-trading-bot-python-V2
And the deeper write-up on information theory is here:
What I'm Exploring Next
I'm particularly interested in combining:
Information theory + probability calibration + Monte Carlo simulation + adaptive position sizing.
The idea is to move from:
"Do I have an edge?"
toward:
"How confident am I that I actually have an edge,
and how much should I risk if I'm wrong?"
That's a much more interesting engineering problem.
I'd be interested to hear from other indie hackers and quantitative developers:
How do you deal with uncertainty when your product or model has to make probabilistic decisions?
Contact Info
https://telegram.me/BenjaminCup
