Human vs Algo: Tracking Rick's Personal Picks Against the Algorithm

Analysis · by Rick's Picks Analytics

One of the most interesting questions in sports betting analytics is does the human pickers' intuition beat the data? We quietly tracked this through the entire 2025 season at Rick's Picks. Here's what the numbers say now that the season is complete.

The Setup

Every week, Rick personally enters his picks in the admin panel -- spread side, totals side, and a free-form note explaining the reasoning. Separately, our algorithm (the engine we've documented extensively in other posts) generates its own picks using CFBD ELO, validated Walters edges, and the conservative confidence framework.

That gives us two public scoreboards measured against the same closing lines: human conviction on one side, a data-driven model on the other.

The Headline Numbers (2025 Season, Final)

Rick's Personal Picks (213 graded picks): - Against the spread: 97-94-5 (50.8%, ROI -3.0%) - Over/Under: 17-14-0 (54.8%, ROI +4.7%)

The Algorithm (2025 backtest replay, same closing lines): - Against the spread: 274-236-3 on 510 plays (53.7%, ROI +2.6%) - Over/Under: zero plays -- with every unvalidated constant zeroed out, no 2025 game's projected total moved far enough past the market line to clear the engine's 2.5-point edge threshold

One honesty note before anyone scores this as a knockout: these are not yet a true head-to-head. Rick picked the games he wanted; the algorithm's record comes from replaying the engine against every eligible 2025 game. A same-game, side-by-side comparison is on the roadmap but not built yet, and we won't pretend otherwise.

What This Tells Us

Four honest observations:

1. Rick's spread record is essentially market-equivalent 50.8% over 191 decided picks is well within statistical noise of 50% (p = 0.70 against the 52.4% break-even). The 95% confidence interval ranges from 43.8% to 57.8% -- that's the inherent uncertainty of a 191-game sample. Rick isn't getting demolished, but he's also not yet showing demonstrable spread alpha.

This is actually the right result given how efficient closing lines are. The famous handicappers in the world hover in the 53-55% range over thousands of picks. Beating Vegas consistently is one of the hardest things in sports.

2. The totals look promising but the sample is too small to call 17-14 (54.8%) on totals with +4.7% ROI is what you'd want from a profitable bettor. But 31 picks is not enough to distinguish skill from variance. We need 100+ totals picks to confidently say "Rick beats the over/under market."

3. The algorithm's 2025 number is above break-even -- and we still won't celebrate it The engine's replayed 2025 record (274-236-3, 53.7%) sits above the 52.4% break-even rate, which sounds like the machine beat the human. But the binomial test against break-even gives p = 0.29: a season this good happens about 29% of the time to a model with zero real edge. Worse, the engine's lowest-confidence picks hit 61.8% while its highest-confidence picks hit 45.6% -- exactly backwards from what a well-calibrated confidence score should do. One season proves nothing for the algorithm for the same reason it proves nothing for Rick. We wrote up the full engine backtest, including the eight prior seasons where it lost, in a separate post.

4. The algorithm sits out; Rick plays The other structural difference is volume and abstention. The engine issued no totals plays at all in 2025 -- its validated factors never moved a projected total far enough from the market line to justify one -- while Rick put 31 totals picks on the board and went 17-14. An honest system needs room for both: a model that only speaks when its validated edges apply, and a human who can take positions the model has no opinion on, as long as both records are tracked in public.

Per-Team Tendencies

Some interesting patterns from Rick's 2025 picks:

Some of these are real reads (Rick rode Ohio State and Big Ten teams well), some are small-sample variance. With 3-4 picks per team across a season, we can't separate "Rick has a feel for Ohio State" from "Rick happened to pick the right side 7 times in a row." That comes with more data.

How This Affects Our Picks Going Forward

For now, we don't bake Rick's pick into the algorithm's confidence score. The sample is too small to claim a stable signal, and the spread hit rate isn't yet beating break-even.

What we do show on every game: - The algorithm's pick and confidence breakdown - Rick's pick (if entered) with his confidence and notes - When they agree, we flag it as a "double-confirmed" spot - When they disagree, we flag it as a "Human vs Algo" interesting matchup

Over multiple seasons, if Rick's hit rate stabilizes above 52.4% (break-even after vig), we'll add a small confidence boost when he agrees with the model. If it stays at coin-flip, it remains a transparent display element -- a story for our users, not a math input.

Why We're Doing This Publicly

A lot of pick services hide their record. They cherry-pick winners to tweet about and quietly delete losers. Rick's Picks shows the full scoreboard -- the wins, the losses, the slightly-below-break-even ROI. Because the only honest way to claim "alpha" is to track every pick and show the full data.

With the 2025 season in the books (and re-confirmed in our July 2026 re-validation of every signal), the verdict is: Rick finished at market on spreads, showed some early signal on totals, with not enough sample to claim either definitively. The scoreboard resets and keeps running for 2026.

This is what honest analytics looks like.

Rick's Picks publishes statistical analysis of college football for informational and entertainment purposes. Nothing here is betting advice. 21+. If gambling is affecting you or someone you know, call 1-800-GAMBLER.