THE BOOK--Playing The Percentages In Baseball

I had thought about that, but not sure it changes my argument from last week, or whenever it was.  Say you take all the teams that round to a 1% playoff chance, every day.  Say the 2005 Astros meet that qualification on 25 different days.  Then we’ve got 40 days of the Pirates, Orioles, etc. (those teams also have many days with a zero percent chance).  That’s your opportunities, but then the Astros also count 25 times to your successes.  Does it really work out to 1%?  Maybe, I don’t know.

I do know I’ll practice restraint in criticizing teams that find themselves in those 1% reports, and still make moves to try and go for it.

Quote from this article: “GA: Originally we didn’t take hot streaks into account. When fans brought that to our attention, we started emphasizing recent scores as opposed to older scores.

P2: Because the way a team is performing most recently is a better predictive indicator than how they were performing a few months ago?

GA: Exactly.”

Is this really true?  I had the opposite opinion.  And if it is true, can it be measured somehow? ( I know that projections change daily, but I don’t think that’s what he’s saying.  I think he’s saying that recent team performance is a better indicator of how they’ll perform in the near future than their stats would indicate.)

I was bothered by it.  But, it depends how much.  If you follow the Marcel weighting of weight equals .999n ^ daysAgo, (where n is a digit somewhere between 0 and 4), then we can see that yes you need to weight recent data more.

But, it depends how much CoolStandings is doing it.

Yes, you want to weight more recent games more heavily, as Tango says, but the CS guy implies it is because of hot and cold streaks.  Hot and cold streaks for teams have no predictive value (pretty much same as for players).

However, Tango, for teams, I think you want to use a more aggressive weighting system because team personnel changes a lot and recent games reflect those changes. For some teams, the team they fielded in May is completely different from that fielded in August.

Anyway, as I said, if you have player projections, I hate using team w/l records to predict their future. Hate it.  Heck, we see so much fluctuation in player component stats, to me, using a win loss record, which is obviously filled with noise, for anything going forward just seems ludicrous.  If you are going to be using to coarse, at least use pythag record.  And then again, while you are at it, at least use team OTS (for offense and pitching) or something like that.  Etc.

I agree about the personnel changes, and argued as much in the past.  Teams in contention around the trading deadline change their rosters more than those not in contention, and so, we have a shift in talent, which may be captured by recent team performance.

Ideally, CoolStandings would update their forecasting engine using Marcel’s day-to-day.

I did this study a few years ago—the optimal weighting for teams from 1900 to 2006 was .9970^(days back), and it was more accurate to use days back than games back. I think this was discussed in another thread.

.997^(days back) seems reasonable.  Probably need something even more aggressive for teams in contention, as Tango says.  Although, teams not in contention often experiment with lots of players.  Look at the starting pitchers for teams like PIT and BAL.  Almost every day, I see a name that is unknown to me.

In any case, as I said, I would personally NEVER use a team’s w/l record for anything predictive.  Ever. The least I would ever use is ERA (or OTS against) and OTS.

OTS?