The next game
Right now Logistic argmax expects him to play 
Lulu next, at 21% confidence.
Across 427 scored games it runs at 34%, against 27% for decay frequency, the model to beat. McNemar's test puts that gap at p = 0.001, so it is performing significantly better than the baseline.
See how all six scoredThe short version
I wanted to know whether I could predict my own League of Legends games. The Prediction page keeps score, including the models that lose.
Every ranked game I play comes off the Riot API into Snowflake overnight, and six models take a guess at what I'll play next. Each is scored walk-forward, trained only on games that happened before the one it's predicting. The accuracy carries a 95% confidence interval, and the gap to a baseline that does no learning at all gets a significance test. On a few hundred games, that baseline is hard to beat.
It runs itself on a nightly cron and serves from cache, so browsing never touches the warehouse. Under $10 a month.
Prediction is the part I'm proudest of: six models guessing my next ranked champion, scored against what I actually played. Statistics covers win rates, and which friends I win more with. Playstyle measures seven axes of how I play against everyone else at my rank.
Data pipeline
Riot Match-V5 API
Pulls any games I haven't downloaded yet.
Ingest
~90 stats per player; a match saves all ten or none.
Snowflake warehouse
Holds every match, and what the pipeline has already done.
New-game gate
Models rerun only after a new ranked game.
Walk-forward models
Six guesses at my next champion, trained only on earlier games.
Publish
Results written back in one format the site can read.
This website
Serves cached results; browsing never hits the warehouse.
Why once a night
Nothing changes between runs, so running it more often would only wake the warehouse to find that out. The cron is best effort, so a run can be delayed or skipped. The heartbeat above is how you'd know.
Stack
| Layer | Choice |
|---|---|
| Warehouse | Snowflake, key-pair authenticated |
| Pipeline | Python, scheduled on GitHub Actions (requests, pandas, scikit-learn, XGBoost) |
| Website | Flask + gunicorn |
| Deployment | Docker container (slim Python base, non-root), deployed on Render |
| Assets | Champion art pre-fetched from Data Dragon, served as static files |
| AI summaries | Claude API, called on demand and rate-limited per visitor |