CRAIG: Bayesian Parameter Tuning for Predictive Elo Models
In December 2025, I began adapting my disc golf ratings algorithm for broader applications, ultimately focusing on NCAA Division 1 basketball. By mid January I developed CRAIG (a Collegiate Ratings Algorithm via Intelligent Guessing). The system utilizes Bayesian optimization to fine-tune a ten parameter ranking model which is trained on 2 years of historical data. To maintain accuracy, CRAIG is retrained every few days, ensuring team ratings and parameters reflect current performace trends. In the time since the initial deployment I have streamlined the workflow by automating the bet selection process and refining the underlying predictive logic.
Note on Transparency:While the source code, training data, and specific parameter weights remain proprietary to protect the model's integrity, I will share a simple excerpt of CRAIG's ouput and bet selections and will also note its recent win/loss percentage and weekly ROI below
Above: A sample of CRAIG's moneyline selections for January 22, 2026.
Update (January 26, 2026): I performed rigorous out-of-sample verification on a population of 1016 games sampled over the past month. To test the model under high volume conditions, I focused on the 15 days with the highest game density. During this period, CRAIG maintained a 70.2% hit rate, carving out a consistent edge over the -230 average implied hurdle (which requires a 69.7% win rate to break even).
While the raw signals alone were profitable, the significant gains came from the execution layer. The sample day shown above serves as a case study: by transitioning from flat betting to a modified Kelly criterion, I was able to scale into positions where the model showed the highest divergence from market prices. This improved capital efficiency fourfold, converting a baseline 9% return into a 37% net ROI for that session. To manage tail risk and mitigate idiosyncratic noise, the system enforces a cross-sectional diversification filter, only triggering when it identifies a bucket of at least 10 concurrent opportunities.
Update (March 3, 2026): As March Madness approaches, I thought I'd post the top 68 teams that the algorithm has been trending towards. To give away too much information would be counterproductive, but the rankings and ratings come from a model that is very similar to the one I currently use, and this outdated model saw a ~40% return on investment during the period that it was active. These rankings are built using that same model, but with updated data (through games played on March 2, 2026).
Selection Sunday Update: Now that Selection Sunday has concluded, I wanted to share two early brackets I've created with CRAIG. A highlighted cell means that the game is a tossup and a bold cell means that the model picked an upset.
Above: CRAIG's bracket using Multivariate TPE .
Above: CRAIG's bracket using CMA-ES.