Subvertadown is in its 5th year, and the project has always been under continuous evolution.
The project became serious after realizing a first year of dabbling was enough to get above-average D/ST accuracy.
For the second year, I implemented the original vision: combining time-series data processing and machine learning techniques, to reach near-top accuracy and allow future week forecasting. The kicker model was unbeaten among tracked sources.
The third season was a wave of successful forecasting, after improvements to overfit reduction and better incorporation of betting lines. Plus a new QB model showed that top accuracy was possible.
For year 4, I created a new data processing method, while greatly widening the scope of input parameters. Accuracy was top-notch for all models, and significantly, the model was implemented online for “real-time” updated webpages instead of static table images. It was also the year I expanded to 25 models, for supplementary stats, and introduced a “closest historical match” tool.
Year 5 is special again, not just because of the website launch, but because of a few combined maturations: 12 full years of data are incorporated, the learnings from last year’s data expansion are taken in, and data processing has again received an overhaul after scientific testing during the off-season. While results are never guaranteed, Subvertadown models have never been so well prepared.
Tagged underBackground , Modeling