Wagr

Wagr is a play-money prediction game for friends, built on a real forecasting engine underneath. You predict how events will turn out, and Wagr estimates the probabilities by combining historical data, live odds, and market signals.

What I built

• A probabilistic forecasting system that blends historical event data, live odds, and market signals into event-probability estimates.
• A custom PyTorch model that combines historical patterns with external market signals into a single probability.
• AI-generated news summaries via the Perplexity API, so players get real context on an event before they predict.
• Peer-to-peer and global prediction markets with market creation, odds updates, and event summaries, wrapped in an engagement layer of streaks, XP and leveling, daily quests, and a leaderboard with head-to-head duels.

Why it matters

Calibration is the whole game here. Folding clean historical data and messy live market signals into a single probability is a fun problem on its own; making that probability actually well-calibrated, instead of just confidently wrong, is the hard one, and it’s the part I’d happily do again. The streaks and leaderboards mostly existed to get friends generating enough predictions to test the model against.