It's the first week of April. A team is 7-1. The internet is calling them a World Series favorite. Their moneyline has moved 20 points toward favorite across every book.
Six weeks later, they're 28-28.
This happens every single year. And every single year, bettors lose money betting on early-season overreactions.
Here's the math behind why — and how we try to avoid getting burned by it.
The problem with small samples in baseball
Baseball has a well-known statistical property: it requires a lot of games before a team's record becomes predictive.
In the NBA, a team's win rate stabilizes after roughly 20–25 games. In the NFL, you have 17 games in a season and the signal is already weak. In MLB, the consensus among analysts is that you need somewhere between 50 and 100 games before a team's record is a reliable predictor of their true quality.
Why? Because each baseball game is won or lost by a thin margin. A ball lands an inch fair instead of foul. A reliever gives up a two-out homer. Run scoring in a single game is highly variable. Over 162 games, the good teams separate from the bad ones. Over 15 games, the variance swamps the signal.
What this looks like in the numbers
A true .500 team — genuinely, objectively average — will sometimes go 9-1 or 1-9 over 10 games just by chance. The probability of a .500 team going 7-3 or better in 10 games is about 17%. That happens to roughly five teams every season. Those aren't World Series contenders. Those are average teams who got lucky early.
The bettors (and sportsbooks) who overreact to those records are taking on pricing risk that the underlying performance doesn't justify.
What we do about it: shrinkage
The technical term is regression to the mean or Bayesian shrinkage. The idea: when your sample is small, you don't fully trust what the sample says. You blend it with a prior expectation.
Our approach works roughly like this:
- Before the season, every team gets a baseline — a reasonable expectation of their quality based on preseason signals (which we approximate as league average, since we don't use offseason projections).
- As the season progresses, we gradually shift weight from the baseline to the actual performance data.
- By roughly game 20, the current-season data has moderate weight. By game 50+, it's driving the model.
In practice this means:
- In the first two weeks, our edge readings are more conservative. We won't post a +15% edge on a team that's 8-2 if their underlying stats don't support that quality level.
- By June, recent performance dominates and the model is running on mostly real in-season data.
- Early-season picks look more like each other than July picks, because we haven't differentiated the teams yet.
This is intentional. The cost is missing some real early signals. The benefit is not chasing variance.
The pitching problem in April
Starting pitcher performance stabilizes even slower than team win-loss records.
A starter's ERA in his first three outings is basically noise. A starter could go 0 ERA over 18 April innings and have a 5.00 ERA by June — or vice versa. The sample is too small to know.
What's more stable, even early, is strikeout rate and walk rate. These stabilize faster than ERA and are better early-season predictors of pitcher quality. Our model uses ERA as an input, which means it's somewhat susceptible to early-season pitcher noise.
What to do with this: be more skeptical of picks that hinge on a single April start from a pitcher with a small sample. The edge calculation may look clean, but the underlying ERA could be misleading.
Why books don't fully solve this either
You'd think the sportsbooks would figure this out and price it correctly. They mostly do — but not completely.
Sportsbooks have market-making pressures that can amplify early-season narratives. If casual bettors flood one side because a team is 9-1, the book moves the line to balance action. That can create situations where the line has drifted past fair value, opening up the other side.
This is one reason we track edge rather than just picks. A team at -180 that "deserves" -140 based on current performance is overpriced even if they're the better team. The edge is negative, which is a signal that even "correct" picks can be bad bets.
What this means for following our card
In April, expect:
- More picks that look closer to coin-flip than blowout
- Smaller listed edge percentages on average
- More picks that shift around as sample sizes grow
In June and July, expect:
- Wider gaps between good and bad teams
- Higher-confidence picks on teams with genuine season-to-date data
- Larger absolute edge readings when they appear
The model isn't broken in April — it's being appropriately humble about what 15 games of data actually tells us.
One more thing: don't backtest April alone
If you're evaluating any picks system — ours or anyone else's — don't cherry-pick April for the backtests or the comparison. Small samples cut both ways. A system that looks great in April might just be getting lucky. A system that looks rough in April might be being appropriately conservative.
The honest evaluation window is a full season or multiple seasons at scale. Anything smaller is variance dressed up as signal.
We post every pick. We track the full record. We don't delete the losses. After a full season, you can make a real judgment call.
For informational use only. Past results don't guarantee future performance. Bet responsibly. If gambling is affecting your life, call 1-800-GAMBLER.