For years our game models have lived on the site as predictions — win probabilities, projected scores, spreads. What we never showed you was the next step: what would we actually bet? This offseason we rebuilt our team models from the ground up (the V3 project — new nflverse play-by-play data pipeline, LightGBM ensembles, a proper walk-forward backtest), and the end product is a weekly betting card. As of this season, that card is on the site as the new Best Bets page, the first tab in the betting section.
This post covers what's on the card, how the model decides what makes it, and the honest backtest numbers behind it.
What the page shows
Every week the model publishes a short list of recommended actions — typically around 5 bets in a normal week (3-8 is the usual range):
- Moneyline value picks (~3 per week) — the core, validated product
- Wong teaser tickets (~1 per week) — complete two-team tickets with both legs listed, plus an upgrade tip when a third qualifying leg exists
- Totals picks (0-2 per week) — clearly badged experimental, more on that below
Each bet card shows the pick, the reference price, the model's win probability next to the market's implied probability, the edge between them, and a recommended stake as a percent of bankroll. There's also a track record view so you can see how the strategy has done — we're not interested in a picks page with amnesia.
How a bet makes the card
Some quick definitions so we are all on the same page:
- Edge — the model's win probability minus the market's no-vig implied probability (the price with the book's cut stripped out). If the model says 58% and the no-vig price says 53%, that's a 5-point edge.
- Quarter-Kelly staking — the Kelly criterion computes the bankroll fraction that maximizes long-run growth for a given edge; we bet a quarter of that, capped at 2% of bankroll per bet, because full Kelly assumes your probabilities are exactly right and ours aren't.
- Wong teaser — a 6-point teaser restricted to the zone where the math actually works: legs that cross both the 3 and the 7 (dogs +1.5 to +2.5, favorites -7.5 to -8.5).
A bet only appears on the card if the edge is at least 2% and the Kelly math says it's worth at least a quarter-percent of bankroll. Plenty of weeks the model likes 3 bets and that's it. A week with no card rows is a valid answer — we render "no recommended bets this week" rather than manufacturing picks.
The backtest numbers
From 2021 through 2025, walk-forward (the model only ever sees data available before each week), graded at closing prices:
- 1,196 bets, +14.1% flat ROI
- Quarter-Kelly staking turned $100 into $868 over those five seasons
- Maximum drawdown along the way: 15%
Now the honest part, because a headline ROI without the caveats is how bad betting content gets written. 2022 was roughly breakeven. Losing months are a normal, expected part of this strategy, and anyone who tells you otherwise is selling something. The backtest is graded at closing prices — your realized prices will differ, which is exactly why the card shows a "worst still playable" price on every bet (the model's edge survives down to that number; below it, pass).
The Wong teaser legs deserve their own number: in our evaluation window, legs in that qualifying zone covered their teased line 78.2% of the time (n=340), comfortably above the 72.4% you need to break even on a two-team teaser at standard pricing. We also tested whether the model could expand the teaser zone beyond Wong's rules with model-selected legs — it couldn't (70.7%, below breakeven), so we didn't ship that. The rules stay strict.
The totals picks are flagged experimental because that's what they are: the signal looked promising in backtests but hasn't proven itself on live 2026 games yet, so it runs at half stakes with a disclaimer until it does. If it doesn't hold up live, it comes off the card.
Shop your price — it's worth more than you think
One of the most repeatable findings from our backtesting had nothing to do with prediction: simply taking the best available price across books instead of the median added about +4.4% per bet on our moneyline picks. That's a bigger, more reliable boost than most modeling improvements we tested all summer. The card is built around this — every bet shows a reference price and the threshold where it stops being playable, and the single best thing you can do with the page is compare that against your books before betting.
How I'd actually use it
- Check the card when it publishes for the week.
- For each pick, find the best price across the books you have. If it beats the "still playable" threshold, the bet is live.
- Size it off the stake percentage — 1.8% of a $1,000 bankroll is $18. If you think in units, one unit per 1% works.
- Treat the experimental totals section with the skepticism the badge implies.
- Expect losing weeks. The backtest had plenty; the equity curve on the track record page doesn't go up in a straight line and the live one won't either.
The first moneyline pick each week is free to view; the full card is for subscribers.
One last disclaimer: this is a model, not a money printer. +14.1% over 1,196 backtested bets is a real result, but backtests are always the best case, and no bet on this card is ever more than a small, mathematically-sized fraction of a bankroll for a reason. Bet responsibly.
Thanks for reading — you can check out the Best Bets card at the Best Bets page, and if you want the model's full weekly output, create a free account by clicking this link.