Pre-game paired-comparison modeling of professional League of Legends map outcomes

Battle Count: 68

Directly relevant through: (1) prediction market benchmarking against Polymarket per-game contracts, demonstrating a transparent statistical forecaster can match market pricing; (2) paired-comparison forecasting methodology applicable to any binary-outcome market; (3) calibration diagnostics (walk-forward slope 0.995) relevant to probability estimation in trading; (4) Diebold-Mariano inference framework for comparing forecasters. However, the domain is esports rather than financial markets, and the model does not address price formation, order flow, or portfolio construction.

The proposed model is deliberately minimal: 4 regression coefficients plus a ridge-penalized team-strength block, fit via a single convex L-BFGS minimization converging in seconds. Feature construction involves EWMA recursion per team per side. Hyperparameter selection is a grid search over (lambda, tau). The two-stage candidate is more complex (REML fit, BLUP prediction, Platt calibration). Main complexity lies in data ingestion from three sources with no common key, pagination handling, and timestamp anchoring conventions.

Data sources are public APIs (lolesports, Leaguepedia, Polymarket). Code was written with LLM assistance (Claude) and reviewed by authors. Hyperparameter selection protocols, grid searches, and evaluation workflows are fully specified. However, no explicit code repository URL is provided for the paper's own pipeline. The community API documentation link is given for data ingestion. Two holdout protocols (global time-based split and per-game walk-forward) are clearly defined.