What Does Vegas Really Think?

About a month ago I wrote two posts looking at the differences between how PECOTA and CHONE expected each team to perform over the first few weeks of the season, compared with the level of performance expected by the Vegas lines. The results of that study were fairly predictable; for the most part the two metrics tracked each other pretty closely (average difference: .021, or 3.4 wins over a full year), with a few large disagreements (most notably the Marlins and Nationals).

That was interesting, but it wasn't clear how seriously we should take these differences. There was bound to be some random variation thrown in there, between facing unusually good/bad pitchers, and also an unusual number of teams that were under- or overrated. We didn't know how these differences would hold up in the future; it was obvious that there was some logic to them, but it seemed possible that the correlation might end up being quite weak.

It turns out that that's not the case at all. Here are the results of a regression with the early season differences as the x-variable, and the differences from the last month of play as the y:

It turns out that, even with just a few weeks of data, there isn't that much random variation at all. Basically, to figure out what a team's strength is going forward, you just need to take their PECOTA/CHONE W% and add 2/3 of the difference between that and their Vegas lines. 

Here's an example.  In late April the Marlins' PECOTA/CHONE strength was .448, and their difference was .062, so we could've implied that their actual strength in the future, adjusted for the Vegas odds, was .490. That would've gotten us pretty close; over the last few weeks their lines have indicated a team strength of .486 (unadjusted for schedule), and their difference has dropped from .062 to .040, almost exactly what we would've expected.

With the Marlins leading the way, here's what happened to them and the teams with the next four highest differentials:

Old Diff: Differential in the first three weeks of the season.
Pred. Diff: "Recent Diff" predicted by regression shown above.
Recent Diff: Differential over the last month.

The differences here are all in the same direction for both periods, and all five are still sizable in the more recent sample (the Rockies' recent difference amounts to a little over two wins per 162 games). The regression doesn't always nail the difference, but that's to be expected.

The next question is what to do now that we have seven weeks of data rather than three. The difference now should clearly be weighted more than the 68% we used a month ago, but we have no real way of knowing how much more. 

There is also the factor that some of the lines may be a bit skewed because of recent performance; for example, Toronto's difference went from +.004 in the first three weeks to +.047 over the last four. To deal with that we would want to regress these to PECOTA/CHONE even more; in other words, to lower the percentage weight given to the difference.  It is also possible -- if not probable -- that we should be weighting the differences in recent weeks higher than those from weeks 1-3, but lets treat them all equally for now.

There is no scientific way to figure out the correct weight going forward, but 75% sounds about right to me. Using that figure, here is the current strength of each team.

Tomorrow I'll use all this information in an updated simulation to see what it means for each team's playoff chances when combined with the current standings.