A Two-Stage Framework for Sports Betting Asset Construction: Probabilistic Prediction and Risk-Aware Portfolio Allocation

This manuscript presents a two-stage framework that treats sports betting markets as an alternative asset class, coupling probabilistic prediction with risk-aware portfolio allocation. The first stage produces calibrated probabilistic forecasts of match outcomes using heterogeneous learners trained on team-strength features, rolling form, and bookmaker odds. The second stage transforms these probabilities into a portfolio allocation problem under Conditional Value-at-Risk (CVaR) constraints, comparing multiple staking strategies including flat staking, fractional Kelly, mean-variance, and risk parity. The framework is empirically evaluated on multi-league, multi-season European football data (2015–2025) covering the top five leagues, using a strict rolling-origin protocol that eliminates temporal leakage. Evaluation combines machine learning metrics (log-loss, Brier score, ranked probability score) with financial metrics (return on investment, Sharpe ratio, maximum drawdown, probability of ruin), providing a joint assessment absent from most prior work in this domain. This work operationalizes the theoretical foundations introduced in our companion preprint on integrated uncertainty-aware decision-making, and extends our previous work on sports betting as an alternative asset class (Galekwa et al., IEEE Access 2026). It contributes to the doctoral research program of the first author on robust decision-making under model uncertainty at the Institute of Smart Systems Technologies, University of Klagenfurt.