personal project · 2026
i like watching professional League of Legends and wanted to study how a completed draft changes a matchup. i built a research app around a CatBoost model that brings champion selections, team and player profiles, market comparisons and prediction tracking into one place.
Predicting which team wins and understanding whether the market prices that outcome accurately are different problems. A strong favourite can still be priced too confidently, and an apparent discrepancy may simply reflect uncertainty or incomplete information.
In League of Legends, each team selects five champions before play begins. This app evaluates those completed selections and keeps the resulting prediction available for review after the match.
01 - JavaScript
Responsibility: Captures two teams and ten role ordered champion picks, then presents the estimate and fixed-nine alternative-pick comparisons.
How it works: Completing a draft reconstructs the model inputs from the frozen browser bundle and runs inference; an alternative-pick check changes one champion while holding the other nine fixed. I used Drafting.gg's interface as a visual reference.
02 - Python
Responsibility: Builds pre-match team, player, champion-pool, role-matchup and composition features from historical games.
How it works: Games are processed in time order: the pipeline builds a row from previously available state, saves that row and only then allows the completed result to update future state.
03 - CatBoost + calibration
Responsibility: Learns nonlinear relationships between the engineered features and match outcomes, then maps the raw score onto a probability scale.
How it works: The feature vector passes through the exported CatBoost trees to produce a raw score, then a separately fitted logit calibration step converts that score into the displayed probability.
04 - Descriptive analytics
Responsibility: Shows role-aware player percentiles, raw statistics, coverage, champion pools and team history for investigation.
How it works: Prepared historical summaries are filtered by player, role and competition scope, then compared with same-role peers while retaining raw values and missing-data coverage.
05 - SQLite ledger
Responsibility: Stores the original prediction before play, accepts the result later and compares model versions on matched predictions.
How it works: Saving creates a timestamped forecast record and model id, settlement appends the outcome later, while reports score only eligible matched records.
06 - Odds arithmetic
Responsibility: Converts manually entered two-sided decimal odds into market probabilities and removes the bookmaker's proportional margin for comparison.
How it works: The interface converts both prices to implied probabilities, normalises their combined margin and places the resulting market view beside the model estimate.
I train on earlier matches, calibrate on a later period and evaluate on later unseen matches because a random split would let future game states influence an earlier prediction task. This leaves fewer recent maps for evaluation, but makes the test closer to how the model would actually be used.
The feature builder reveals each result only after its recorded availability time, and games sharing a timestamp are featured before either result is added. Historical reconstruction becomes more demanding, but it prevents the model from learning with information that was unavailable at prediction time.
CatBoost learns the ranking patterns, while a later held-out period fits the probability mapping. Reserving data for calibration reduces the examples available to the classifier, but a score described as 70% should be evaluated as a probability rather than only as a ranking.
The main model excludes market probability because the historical data does not contain consistent, timestamp-matched two-sided post-draft prices. Manual odds can still provide context, but the missing history means a displayed discrepancy is not evidence of a profitable strategy.
A local version connects the statistical research to a browser interface. Draft analysis, profiles, market comparisons and evaluation are implemented, with automated checks around prediction consistency.
Development and evaluation are ongoing. Draft entry and result recording are still manual, and an apparent pricing difference does not establish profitable performance. Time-ordered backtests are being used to compare training windows and evaluation splits.
Rare champion picks and combinations have limited historical support. In those cases the model can fall back toward broader team-strength patterns, so the interface keeps sample support beside alternative-pick comparisons rather than presenting every estimate with equal confidence.
203.32u
settled bankroll
+103.32u
net profit
Continue collecting saved forecasts and results on later, unseen matches.
date TBCInvestigate different historic odds providers to improve training.
date TBCImprove champion features to account for low sample draft picks without overrelying on team strength.
date TBCCompare alternative model designs under the same point-in-time evaluation contract.
date TBCExperiment with neural networks to model draft to better account for low sample draft picks.
date TBC