I tested my own method on 1,743 games before asking you to trust it
Before this site asks anyone to care about its record, it should be able to answer an obvious question: is the method any good?
So I tested it. Every completed game of the 2026 season, 1,743 of them, walked in order so that nothing from the future leaks backward. At each game I reconstructed exactly what was knowable beforehand, then scored several ways of picking a winner. The script is in the repository and you can run it yourself.
Here is what came back, and the first half is not flattering.
Picking single games is close to a coin flip
| How the pick was made | Accuracy | Games |
|---|---|---|
| Better run differential | 52.8% | 1,360 |
| Better record | 52.5% | 1,337 |
| Better Pythagorean record | 52.5% | 1,364 |
| Always take the home team | 52.0% | 1,364 |
| Take the more "due" team | 51.8% | 1,364 |
Read that table honestly and it says something uncomfortable. The best signal I have beats picking the home team every single time and thinking about nothing at all by eight tenths of one percentage point. Across 1,360 games that difference is inside the noise. It is not an edge. It is a rounding error wearing a lab coat.
This is not a failure of effort. It is what baseball is. One game is nine innings of coin flips with a good pitcher's thumb on the scale, and anybody promising you they hit 65 percent on single games is either counting selectively or lying.
So when Pick No. 1 said low confidence, that was not modesty. That is the honest ceiling of the format.
The part that did work
Now the other half. Pick No. 1 rested on a specific claim: that Detroit, sitting ten wins below what its run differential deserved, was due.
That claim is testable too, just not one game at a time. So I grouped every team-snapshot in the season by how far the team sat from its Pythagorean expectation, then looked at what they actually did over their next twenty games.
| Group | Win rate, next 20 games | Snapshots |
|---|---|---|
| Very unlucky (more than 6 points below) | 61.0% | 151 |
| Unlucky (2 to 6 points below) | 49.2% | 659 |
| Neutral | 50.6% | 660 |
| Lucky (more than 2 points above) | 47.7% | 688 |
Teams buried furthest below their expectation went on to win 61 percent of their next twenty. Teams running hot went 47.7 percent. That is a real spread, and it runs in exactly the direction the Pythagorean argument predicts.
Now the caveat, because leaving it out would make that number a lie. Those 151 snapshots are not 151 independent events. If a team sits deeply unlucky for a month, it contributes dozens of overlapping rows describing nearly the same stretch of baseball. The true sample is a handful of team-seasons, not 151. One season, one league, and the effect could shrink hard with more data.
So: suggestive, directionally right, not proven. I would rather tell you that than print 61 percent in a big font and hope you do not ask.
What this site is actually for, then
Put the two findings side by side and the strategy writes itself.
Single-game calls will hover near 50 percent forever. Mine, yours, everybody's. When this record sits at 8-7 in a month, that is not me being bad at this. That is the game. The reason to keep making the calls anyway is that a public, timestamped, ungameable record is the only way to prove nobody is cooking the books, including me.
The interesting work is one level up. Not "who wins Tuesday" but "what happens to a team ten games below its expectation," and questions like it. That is where the numbers actually say something, and it is where this site is going to spend most of its time.
Detroit is the most unlucky team in baseball right now. The history says teams in that position tend to play better from here. That is a far more useful thing to know than my opinion about tonight.
Method: MLB Stats API, all 1,743 completed regular season games through 2026-08-07. Pre-game state reconstructed by walking the season chronologically, scoring only once both teams had 25 or more games played. Pythagorean exponent 1.83. Full script at scripts/backtest.py in the repository.
Not betting advice. Just calls, made in public and kept in public.