Baseball - Idea for statistical prediction model

[BobJoeJim]

I've always been very interested in computer rankings and prediction models in sports, purely as a hobby. Baseball has always been my favorite sport. Now that I've developed a keen interest in sports betting, as well, it's no surprise that I've been thinking a lot lately about how to put together a statistical model that I could use to predict games. I feel like I've come up with an idea that could turn into a pretty decent system, and I wanted to run it by the board, and see what you all think about it.

Essentially, I would have a spreadsheet or database with all the players on each team, and their RC/27, or some other offensive measure designed to predict runs. Then I'd be able to, before any game, select the most likely starting lineup, and get out an expected run total for that team, with those starters. Similarly, I could use the ERA of the starting pitcher, his average innings per start, and the bullpen ERA, along with some sort of defensive metric to predict unearned runs, to calculate a team's expected runs allowed.

Once I have runs scored and runs allowed, a pythagorean calculation could easily give me an expected winning percentage for that team, given that starting pitcher and that starting lineup. Armed with an expected winning percentage for each team it would be easy to calculate the odds of each team winning, and there we are!

Of course there are many more technical details that would be involved for it to be accurate (for example, ballpark adjustments would be pretty critical), and I'd be happy to discuss those, but I'd like thoughts on the basic concept, first.

Putting together a system like this will take a bit of time (not too much, for the basic idea, but the details that will make it more accurate, those will take a lot of work), but once I have it I'll be able to get predictions on any game very quickly, so if it has a good chance of being accurate it would probably be worth the effort. Thoughts, questions, comments, suggestions? Fire away!

[gomberg]

I like the direction you're heading, but don't you think the linesmakers are already doing this? And if so, do you think yours will be better than theirs - which is what they do as a full time job? It's tough...

Now, what you can do is make a model like this, figure out what line movements mean, and bet the really big deviations in your model to the current line. So what you're really doing is betting against public money, which is what you'd do w/out a model as well [img]/images/graemlins/smile.gif[/img] Good luck though if you try and put it together. Let us know how it's working if you use it to bet.

[mj555]

I'd shy away from using ERA for pitchers. There are much better ways of measuring their production, something as simply as WHIP, or something more complex like DIPS (Defense-Independent Pitching Stats).

[eternalnewbie]

Just a quick thought - I would think that your system might be more profitable if you use it during games rather than after.

I would think, for example, that (depending obviously on the predictive power of your model) you might have a decent edge during the game on other bettors on an exchange like tradesports. This is because you have already put in all of the analysis, and can just quickly plug the numbers into excel.

Presumably these other bettors aren't putting as much thought into into their during the game bets as you are (for example, I would expect that many of them seriously undervalue the probabilities of a team's making up a large deficit which could be quite profitable for you in hitter friendly ballparks).

Also, you have the advantages of a lower vig and not having to go up against a line generated ahead of time by people with many more resources than you.

Just a thought.

[BobJoeJim]

[ QUOTE ]
I'd shy away from using ERA for pitchers. There are much better ways of measuring their production, something as simply as WHIP, or something more complex like DIPS (Defense-Independent Pitching Stats).

[/ QUOTE ]
Absolutely, but that's one of those technical details I wanted to stay clear of at first in this discussion. However, while ERA isn't a perfect stat for pitchers, since my goal here is to predict a team's average runs allowed, for the sake of a pythagorean winning% prediction, ERA will probably be a good place to start, though I'll likely want to shift it up or down for individual players if their periferal stats are noticable better or worse.

[stephan]

[ QUOTE ]
I like the direction you're heading, but don't you think the linesmakers are already doing this? And if so, do you think yours will be better than theirs - which is what they do as a full time job? It's tough...

Now, what you can do is make a model like this, figure out what line movements mean, and bet the really big deviations in your model to the current line. So what you're really doing is betting against public money, which is what you'd do w/out a model as well [img]/images/graemlins/smile.gif[/img] Good luck though if you try and put it together. Let us know how it's working if you use it to bet.

[/ QUOTE ]

Our job as betting junkies isn't to find the mistakes that the books made with their lines, it is to find out what the books know to be the true value of the line.

[VickreyAuction]

If your goal is to predict average runs allowed, consider using RA instead of ERA. That factors in unearned runs by the pitcher without the filtering of an official scorer to arbitrarily determine what is an error and what isn't.

Are you familiar with Bill James' log5 method? He came up with a probability model for the likelihood that team A would beat team B given the winning percentages of team A and team B. If you can figure out expected winning percentages for each team, then using that method would give you a firm number for the probability that one team wins.

And there's definitely more advanced run stats than RC/27 now. VORPr and MLVr are 2 that I know of.

[King Yao]

I think that's a nice start. As you go along, you'll probably pick up a few things to tweak your model. Stuff like "is there something better than ERA?"

Of course, you'll have to compensate for these other issues too:
- how useful is historical data for the players and how far do you go back? Do you go back the same amount for every player or adjust on a player-by-player situation? what about a player aging? What about a young player improving?
- lefty/righty issues....platoons?
- bullpen issues.
- Is there something better than Runs Created? Can RC overestimate or underestimate certain type of players?
- defense
- injuries...how good are the backups if they are playing full-time?
- when do regular starters get days off? some play 162 games every year (Tejada), others only 140 (Bonds)

There's a lot of work and there are so many variables that you'll be constantly adjusting your model.

[silentbob]

A stat that I used with moderate success starting around June last year -- one that the linesmakers probably don't consider -- was BABIP. Fading the SPs with low BABIP and betting the SPs with high BABIP (provided that their peripherals were decent). Sorta along the same lines as DIPS mentioned above.

[mj555]

BABIP = Batting Average for Balls in Play?

If so, then that sounds like a good stat. Basically, finding the pitchers who have been lucky / unlucky.

Edit: Lots of good info here