The Two Totals Edges We Found: Cold Weather and Big Ten Football

Analysis · by Rick's Picks Analytics

Most of the betting "edges" we tested turned out to be market noise. But two effects on totals (over/under markets) survived our entire backtesting gauntlet and now drive the OVER/UNDER picks in Rick's Picks. They're not flashy, they're not new, but they are the only two totals effects our data actually supports -- validated in a training window, confirmed in a held-out test window, and re-confirmed against a season neither had ever seen.

Edge #1: Cold Weather Suppresses Totals

When the temperature at kickoff drops below 40°F and the game isn't in a dome, scoring goes down. We tested this rigorously against 5,649 train games (2015-2022) and 2,313 test games (2023-2025):

Train: -2.73 points to the total on 359 cold games (p = 0.007, Cohen's d = -0.15) Test: -2.57 points to the total on 111 cold games (94% retention, sign agrees with train)

That's a textbook promotion: the train effect is statistically significant under Bonferroni correction, the test holds up in direction and magnitude, and Cohen's d shows a small-but-real effect size.

This signal actually got STRONGER with more data. When we first promoted it on 2023-24 test games alone, retention was only 53% (-1.46 test effect on 77 games) -- right at our threshold. Adding the 2025 season as fresh holdout lifted retention to 94%, and 2025 by itself was the strongest cold-weather year we've measured: -5.12 points on 34 cold games (d = -0.31).

To be safe, the live engine applies about three-quarters of the train effect (roughly -2.0 points) instead of the full -2.7. Two reasons for the discount: the test confidence interval still crosses zero (n = 111 is honest but not huge), and the eye-popping 2025 number comes from just 34 games -- we won't upsize a coefficient on a 34-game season. If the live edge is closer to 2 points than 3, we're still betting on the right side, just more conservatively. (We originally shipped this at -1.4, sized to the old test effect; the re-validation justified moving it to -2.0.)

What about wind and rain?

We tested wind > 15mph (no significant effect), precipitation (train showed -3.83 points but test FLIPPED sign to positive -- and stayed flipped in 2025, at +2.03 -- classic data-mining illusion), and dome venues (effect was -0.05 points, basically zero). None promoted. The market correctly prices wind and rain into the open line; only the cold-weather effect is consistently under-weighted.

Edge #2: Big Ten Games Go UNDER

This one is the cleanest result we have on the totals side.

Train: -2.68 points lower total scoring in Big Ten games (p = 6e-06, Cohen's d = -0.15) Test (2023-2025): -4.27 points lower (p = 1.3e-06, d = -0.26, n = 401) -- actually GREATER than train, with the 2025 season alone at -2.90 on 136 games (108% of the train magnitude)

When test retention is > 100% (the effect got stronger), you have very high confidence the signal is structural, not statistical noise. Big Ten football really does play to the under at this scale.

The reasons are well-understood: defensive recruiting profiles, slower offensive systems on average, weather (the conference plays in November/December cold), and a coaching tradition that emphasizes possession football. The market knows this in aggregate, but doesn't fully bake it into the over/under -- so the under cashes more than it should.

The live engine applies -2.68 points to the total whenever either team is from the Big Ten. That's the validated train coefficient -- we deliberately do NOT chase the stronger test number (-4.27), because resizing a coefficient to fit your holdout data is just overfitting with extra steps. If a game has BOTH a Big Ten matchup AND cold weather, the adjustments stack: -2.68 - 2.0 = -4.68 points off the total.

Why So Few Totals Edges?

We tested many other totals hypotheses that didn't make it:

• General UNDER bias claim (-1.6% edge from legacy code): UNDER hits 51.9% in train but 48.8% in test (sign-flipped). Rejected.
• Big Ten UNDER bias as standalone signal: UNDER hits 52.0% / 51.3% -- below the 52.4% break-even.
• Pace-based totals (fast offenses → OVER, slow → UNDER): sign-flipped between train and test in both directions.
• Drive efficiency proxies: ~50% across the board.

Cold weather and Big Ten unders survived. Everything else didn't. That's the methodology working.

It's worth pausing on why the survivors are both totals effects rather than spread effects. A point spread is fundamentally an opinion about which team is better, and team quality is the single most heavily modeled quantity in the sport -- every rating system, every projection, every syndicate is grinding on it. Totals get less of that attention, and game-environment factors like temperature interact with totals in a way that's easy for a team-centric model to under-weight. A book can know it's cold in Madison and still not move the number quite far enough, because "how much does 28°F subtract from combined scoring" is a narrower, less-contested question than "who wins and by how much." Small pockets of under-correction persist where fewer sharp models are pointed.

The other thing both survivors share is a physical mechanism. Cold hands, harder footballs, and wind-adjacent November weather genuinely make passing and kicking harder. Big Ten rosters and schemes are genuinely built around defense and possession. When an effect has a mechanism you can explain to a coach -- not just a pattern in a spreadsheet -- it's far more likely to persist, because it comes from how the games are actually played rather than from a temporary quirk in how they're priced.

What This Looks Like on the Site

When you see a Rick's Picks UNDER recommendation on a cold-weather Big Ten game, the engine isn't guessing. It's compounding two effects that have been validated across 11 seasons:

1. CFBD ELO suggests the over/under is approximately fair
2. Walters cold-totals adjustment: subtract 2.0 from the over/under
3. Big Ten defensive adjustment: subtract another 2.68
4. If the adjusted total is more than 2.5 points below the market line, we recommend the UNDER

That's three honest components, each documented in our integration log, each backed by the methodology. No guesswork, no narrative fitting -- just the two totals edges that survived 11 years of testing.

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.