The 2020 ZiPS Projection Wrap-up, Part I: The Teams

While there’s still a bit of baseball left to be played, this is always the time of the year when I dissect the current season’s ZiPS projections. Baseball history is not so long that we suffer from a surfeit of data, and another season wrapped means more for ZiPS to work with. ZiPS is mature enough at this point that (sadly) the major sources of systematic error have been largely ironed out, but that doesn’t mean that the model doesn’t learn new things from the results.

2020 was a highly unusual season (for very unfortunate reasons); its shortness will hopefully provide us some insight into baseball played in a truncated format. In terms of projections, I tend to have a conservative bent, and I like to be very careful about making sure I know which things have predictive value before I integrate them into the myriad models that make up the various ZiPS projections. A lot of my assumptions going into this season required far more guesswork than usual; I had no idea how teams would actually use prospects in a shorter season, what the injury rates would look like once we brought COVID-19 into the mix, or if we would even complete a 60-game slate.

In light of the risks involved, I kept player totals in the playing time model lower than I would have in a normal season, but I had little clarity into what the league’s COVID-19 case rate would be over the course of the season. Even the way-smarter-than-me epidemiologists didn’t know and I, alas, didn’t major in mathemagical science. With more volatility in projected roster construction, the ZiPS standings gave larger error bars than I’d expect over a “normal” 60-game season, but I wasn’t really sure if that was right.

In the end, the strangest thing to me was just how normal everything turned out being. After an inauspicious start to the season — with testing delays the first weekend of summer camp, early outbreaks on the Marlins and Cardinals, and poor team communication as to just what the rules were — I wasn’t optimistic. But in the end, 28 of the 30 teams played all 60 games, and the two teams that didn’t, the Tigers and Cardinals, were ready and able to play their missing games if they were needed to decided the standings.

Let’s start with a chart of the basic win totals vs. what reality handed down.

The average miss in win totals was 3.5 wins per 60 games, slightly better than the 3.7 I projected my average miss to be going into the season. Solid start! But is 3.5 good or bad? Those can be loaded terms when it comes to evaluating accuracy, so I always like to weigh ZiPS against the Vegas results. Disclaimer: this was the first time, since online betting became a thing, that I placed no bets on a season. While I’m not one to judge others, I personally felt a little queasy gambling on team results given the backdrop of this year.

As for the Vegas results, I used these from the same day I ran the projections. I’ve compared that to what ZiPS foresaw in the table below, highlighting hypothetical successes in yellow, misses in blue, and the pushes in red.

If you had bet the over every time ZiPS projected more wins than Vegas and the under every time the opposite was true, how would you have done? Out of 26 bets — ZiPS matched Vegas exactly for two teams, and most sportsbooks gave refunds to Cardinals/Tigers bets — you would have been correct 16 out of 26 times. If you only made bets in situations where ZiPS disagreed with Vegas by more than a win, you would have earned your cash eight out of 10 times. If the sportsbook did not cancel the Tigers bet given that they already clinched the over, that would have been nine of 11.

Unsurprisingly, the more ZiPS agreed with Vegas, the more ZiPS vs. Vegas became a coin-flip, with ZiPS losing five to six on the 12 teams with a half-win disagreement entering the season.

Another way to look at team results is the win probabilities for each team. ZiPS projected the Oakland Athletics to win 33 games while they actually won 36. In the preseason distribution of ZiPS win totals, 36 wins would have been the 70th percentile projection for the A’s. Let’s wash-rinse-repeat for the other 29 teams as well! For the Cardinals and Tigers, the percentile is based on a 58-game season.

In terms of beating expectations, the Blue Jays came out on top. While they didn’t have the largest positive separation from ZiPS (the Marlins did at six wins), the projection did not see the Blue Jays as a team that was as volatile as the Marlins, in this case because of the team’s large surplus of young, uncertain pitching. While ZiPS was more positive than most when it came to projecting the playoff probability of basement-dwellers like the Orioles and Mariners — 8% and 15%, respectively — both teams did outperform what ZiPS expected.

On the negative side, ZiPS did not foresee the degree to which the Pirates fell apart, and though the loss of Stephen Strasburg hurt the Nats by a win or two, the non-Max Scherzer portion of the rotation and a number of positions (1B, 3B, RF, CF) were far worse than the computer projected.

Ideally, you would want six teams in each bucket. I normally track Brier scores, but with a short season leaving me little to compare to, they’re not terribly useful here.

Coming up in Part II, we’ll look at the sources for team projection errors (player quality vs. roster composition) and which parts of which teams were the most responsible for the misses.