Jackson Angell

A pipeline that pulls every match I play from the Riot Games API into Snowflake, then layers models and a dashboard on top. Prediction guesses my next champion (ranked games only); Statistics and Playstyle cover the social and stylistic side across every game.

Five models on a ladder of complexity, all trained on ranked solo queue only (queue 420). In ranked I pick what I believe is my best champion, so my history predicts my future; normals and ARAMs are experiments a model would just learn as noise. Features are recency-weighted (45-day halflife) so they follow a pool swap within a handful of games.

The Prediction page shows the two-step sampler: it samples a role, then a champion within it, rather than taking the most-likely answer. That draw happens once per new ranked game (seeded off my latest match, so it's stable until I play again) and lands on a pocket pick every so often, which is what makes checking back worthwhile.

Behind it: a logistic argmax, an XGBoost model (kept in to show how a data-hungry algorithm behaves on little data), a decay baseline, and a uniform floor. Each logs a pick for every past game beside the champion I actually played, so accuracy is measured, not asserted. See the scoreboard.

1. Match discovery: poll Match-V5 for new match IDs and upsert them into a match_ids discovery log, so re-runs are idempotent.
2. Match ingestion: fetch full detail for anything not already in Snowflake, then parse ~90 columns of participant stats into matches / participants.
3. Champion reference sync: check Data Dragon's latest patch against a pipeline_state key/value table, and refresh the champions table only when the patch actually changed.
4. Walk-forward prediction: for every ranked game lacking one, write a row to the wide predictions table (the actual champion beside all five models' picks) in a single batched insert.
5. Live prediction: publish each model's pick for the next, still-unplayed game to analytics_output under live_prediction.
6. Publishing: stats, model rollups, and the playstyle analysis land as JSON rows in the single generic analytics_output table, keyed by output name, so this site reads them all with one query shape.