We Said Ninety Wins. They Won Seventy-Four.

How wrong were we about the Mets, and which kind of wrong was it? We published five forecasts this season. They finished 74–88, fifth in their division and thirteenth of fifteen in the league.

Our March number was ninety wins. We missed by sixteen. You are owed the whole list, so here it is.

But the interesting failure is not the size of that miss. It is the pattern underneath, and you would probably guess it backwards. Our gloomiest number was our most accurate. Every time we talked ourselves up from a bleak figure, we got further from the truth. By the end of this you will have a cheap test for whether a forecast is being reasoned with or reasoned away.

The Five Calls, Graded

March, before a pitch. We ran a Beta-Binomial on a 3–3 start using a preseason prior of about ninety wins, and concluded six games was noise. That part was correct. The prior was not.

April, at 7–14. We said 80–82 wins with a 90% credible interval of [70, 89]. The point estimate missed by seven. The interval contained the answer. Hold on to that one; it is the only thing on this page that worked.

May, at 20–26. The arithmetic gave a 76-win pace. We then wrote that the projection “plausibly climbs from 76 wins back into the low-to-mid 80s.” It did not climb. This was the worst call of the year, and it is the one worth studying, because the number was nearly right before we improved it.

June, at 25–33. We said “on pace for approximately 72 wins” and left it alone. Final answer 74. Off by two — the best forecast we made all season, and the one with the least argument attached.

September. We ran a piece called “The Mets Must Pass Seven Teams.” They passed none.

The Prior That Would Not Die

Watch the order. Ninety in March. Eighty-one in April. Eighty-three in May. Seventy-two in June. The number walks down all year, which is what a model is supposed to do as evidence arrives — except for May, where it walks back up.

That is the signature of a prior holding on too hard. We began with a ninety-win team in our heads, so each bad month read as something to regress away rather than something to learn from. The May piece is the cleanest specimen. The Pythagorean told us 76; we spent the rest of that paragraph explaining how it would get better — healthier arms, an easier schedule, regression to a mean we had assumed.

None of those mechanisms were measured. They were reasons, and reasons are how a forecast gets talked out of what it just computed. The unadorned number was nearly right. The reasoning laid on top of it was the error.

The One Thing That Worked

April's forecast was wrong by seven wins and is still the most useful thing we published, because it came with [70, 89] attached and 74 sits comfortably inside.

That is the whole argument for publishing error barsi rather than numbers. We get to make it at our own expense, which is the only way you should accept it from anybody. A point estimate can only be wrong. An interval can be honest.

Notice what that interval cost us, though. Nineteen wins wide is not a bold forecast; it is close to admitting we did not know. That is exactly why it survived, and it is the trade nobody wants: the forecast that holds up is the one that sounds least impressive in April.

What to take home

Here is the test, and it costs nothing. When you meet a projection, find the raw number first — the one the arithmetic produced — and then read what was said about it afterwards. If the commentary moves the number, ask what new measurement moved it. If the answer is a list of reasons rather than a quantity, you are watching someone talk a model out of its own result.

We did that to ourselves in May, in print, and it cost us nine wins of accuracy. The number knew. We argued with it.