Jack Busher

The story

This started as a single question: are Polymarket prices ever wrong enough to bet against? It turned into three separate systems, each one building on what I learned from the last.

1. TSA Edge — the first attempt

I built a 6-layer model to predict daily TSA airport passenger volumes — day-of-week patterns, seasonality, year-over-year growth, holiday effects, and even weather impact across 10 major US airports using NOAA and FAA data. It compared predictions against Polymarket bracket prices and flagged undervalued bets.

After 13 settled bets, the result was clear: unprofitable. The model hit 46% accuracy with a poor Brier score. The market was already pricing in everything my model could see. But it taught me the fundamentals — Kelly sizing, expected value calculations, and how prediction markets actually work.

2. Football Edge — getting more sophisticated

Next I went after football (soccer) markets — Premier League and Serie A. This system pulls real team statistics from FotMob and expected goals (xG) data from Understat, then compares implied probabilities against Polymarket prices for Under 1.5 Goals and BTTS markets.

The model blends venue-specific rates with recent form (50/50 season vs. last 6 matches) and adjusts for xG deviation. It includes a walk-forward backtester with 13+ strategy variants and a daily scanner for live opportunities. More rigorous than TSA, but still learning whether the edge is real or just noise.

3. Poly Edge — the broad approach

Instead of predicting outcomes, I flipped the question: how often does the market leader actually win, and is that frequency higher than their price? If leaders at 24 hours before close win 85% of the time but are priced at 65c, that's a 20-point edge.

This system captures real-time prices every 10 minutes across 30 recurring Polymarket series — box office, Billboard, Netflix, Trump approval, Fed decisions, tweet counts, and more. Five strategies evaluate every market and log paper trades automatically. The whole pipeline runs unattended: capture, evaluate, trade, settle, repeat.

What was hard

Starting off I wanted to try find a market which wasn't as competitive as the top markets. Trying to find a balance between niche, but also a market where there was enough liquidity. This was how I arrived at TSA numbers markets. Instead of just relying on the official numbers which are published by US Gov, I wanted to introduce other data sources to help make a decision on what to bet on. I had the idea of using weather forecasts, to try predict downturn in traffic. I was able to find sources for weather events near airports and thought I was golden, until I realised my model was overestimating the impact of weather events on traffic. For example, a flood warning in downtown Chicago would flag as a weather event, the model would then apply some value to this event, and in turn, the model would predict downturn in traffic. I found that a flood warning in downtown Chicago, was having the same effect on the predictions, as a hurricane near O'Hare airport. So the problem became how do we weight these events. I decided on weighting them by:

What I learned

Your takeaways go here — market efficiency, the importance of sample size, when to pivot vs. persist, building automated pipelines.

The idea

Irish is taught in every primary school in Ireland but the resources for learning it are pretty dry. I wanted to build something that kids would actually enjoy using — a game that teaches Irish verbs, vocabulary, and sentence structure through 3D mini-games, all mapped to the actual primary school curriculum.

How it works

Students play as a customisable wizard character across four game modes, each teaching different skills:

• Magic Doors — read an Irish sentence and walk to the correct door. Quick recognition and vocabulary.
• Glass Bridge — build sentences word-by-word by jumping across glass tiles. One wrong step and you fall. Teaches grammar and sentence structure.
• Bomb Defusal — cut the wrong wires before the timer runs out. Forces fast recall under pressure.
• Word Connections — group 16 Irish words into 4 categories. Tests semantic understanding.

Teachers can log in, pause mid-level, reveal answers, adjust speed, disable lives for younger kids, and track student progress. The whole thing is deployed on Vercel with Supabase handling auth and data.

What was hard

Your story goes here — the 3D rendering, getting the game feel right for kids, curriculum mapping, the body tracking mode you had to scrap?

What I learned

Your takeaways go here — Three.js, building for a specific user (kids), working with teachers, shipping to production?

The idea

I wanted something like BeReal but with no social aspect — just a private daily record for myself. Take 3 photos throughout the day, whenever you want, and they sit side by side as a collage. Over time you build up a visual diary you can scroll back through.

How it works

It's a progressive web app — add it to your phone's home screen and it feels native. PIN to log in, tap the camera button, take your shot. Photos are resized and stored server-side. The main screen shows today's collage filling in as you go, with past days scrollable below.

Self-hosted on a VPS so photos stay private — no third-party cloud, no social features, no tracking. Just your photos on your server.

What was hard

Your story goes here.

What I learned

Your takeaways go here.