System for over/under analyses

[wadea]

So, I had some free time and some help from a spreadsheet wizard today and I put together I spreadsheet that I believe should get me pretty close to predicting over/unders for baseball games (and can be adapted to anything that is "Poisson distributable"). I haven't yet tested it, but I have the optimism that is usual before a well-conceived brainchild is proven utterly useless if not very costly. I will describe it below and post the picks it has selected for today. I know it's not taking every factor into account, so I'd like to hear everybody's opinions about what the biggest conceptual "leaks" are. I realize that injuries will tend to favor the under and weather may play a role. What else? Keep in mind, the number are generated from real game data, where players have been missing from the lineup for at least some games and the weather has not always been perfect. Okay, here goes:

1) I take the home and away "average runs scored per game" for each team. I build a grid where, going across is the poisson distribution from zero to 15+ of the home teams average and going down is the road team. The inside of the grid is the product (multiply) of the two components. Thus, theoretically, each cell of the grid represents the probability of the game ending at that exact score (assuming R/game is the only appropriate metric involved).

2) I erase the diagonal since the games can't end in ties and total all cells in the grid (some number <1 since ties were removed). A new grid was made where each grid value was divided by the sum of the last, such that the sum of all cells is once again =1.

3) The next two essentially identical grids (one for over and one for under) say, "if [homescore]+[roadscore]>[overunderscore], then this cell should = the cell from the above grid, if not then it =0". The sum of these grid cells equals the total probability of all outcomes that yield a score over the set O/U score. The probability for over and under are then divided to yield a line for each side in the usual way.

4) But what if Johan Santana is pitching today? I have another identical spreadsheet for pitching factors. Instead of a Poisson distribution around the average runs scored (offense) for each team, I used the following formula. (([starterERA]+[starterERA]+[teamBullpenERA])/3)*1.085. I figure the starter to go six innings (2/3 of a game) and the bullpen to go 3 innings on average. The factor of 1.085 is what I found to be the MLB average ratio of [total runs allowed]/[earned runs allowed]. In other words, it's the unearned runs factor. Finally, because the visiting team doesn't play defense in the 9th if the home team leads (but this isn't figured into ERA since it's calculated by innings pitched), I figured this would happen about half of the time so I multiply the visiting teams ERA by (17/18). So, replacing [runs scored/game] with [ERA * unearned run factor] this new set of grids uses the same methods as before to produce a moneyline for the O/U.

5) I just add the two lines (i.e. offensive over + defensive over) divide by two and use that as my line.

So, here are the formula's predictions for today's games. For the purposes of testing, I'll include every pick with even the slightest edge.

FLA/PHI(9) - NO PICK
PIT/MIL(7) - O -125
ATL/NYM(8) - O -104
COL/CIN(9.5) - U -117 (my line here is -118)
WSH/MIL(7.5) - O +118
HOU/STL(8.5) - NO PICK
ARI/SD (7.5) - O +118
SF/LAD (8) - O -109
KC/DET (8.5) - U +103
CWS/CLE(8.5) - O +125
NYY/BOS(10.5)- U +125
TB/TOR (9.5) - O +115
ANA/MIN(8.5) - O -112
TEX/OAK(8.5) - O -108
BAL/SEA(8.5) - O +110

As you can probably already see, there's possibly some bias towards picking the OVER.

While writing this, I realized that there will be a slight but small bias in that for all the times the 9th inning ends in a tie, the next highest 1+ scores should be favored. If teams keep matching each other run-for-run in extra innings, it could go much higher. In other words, the score goes up from there, but never down. I'm not sure how significant this is, but it will favor the over. Maybe it'll offset runs lost to injuries, etc. Perhaps I'll account for that when I have more time to fiddle with it. I'll continue to keep track of both this formula's ability to beat the O/U money line, as well as to just correctly pick over or under. As always, it'll be interesting to see if it works over time. Please let me know what you think about this whole system...

-w.a.

[Losing all]

I'd be a little worried about all the overs as they often represent square action. It could be a one time anomoly, or it could be a garbage system.

Do you plan on posting these picks? If you do I think we need to see the line this thing comes up with for it to have any value.

Good luck. It will be interesting to follow the progress if you continue to post picks/results.

[GrekeHaus]

Very interesting idea. I think that you could account for extra inning games by just originally taking the numbers along your diagonal and counting these times as the number of runs scored +1 since the vast majority of extra inning games are decided by 1 run. If you didn't want to do this, you could come up with some formula for the expected number of times that each number of runs is scored.

I'd say the biggest oversight of your method is that it doesn't take into account the ballpark factor. A game played between the Dodgers and the Rockies is likely to be higher scoring if it's played in Colorado rather than LA.

[wadea]

It's a good point about the line generated by the spreadsheet. If I have time tomorrow (better than even money), I'll post the picks it generated for Saturday's games and include my lines in parentheses or something. Thanks for the feedback,
-w.a.

[wadea]

[ QUOTE ]
Very interesting idea. I think that you could account for extra inning games by just originally taking the numbers along your diagonal and counting these times as the number of runs scored +1 since the vast majority of extra inning games are decided by 1 run. If you didn't want to do this, you could come up with some formula for the expected number of times that each number of runs is scored.

[/ QUOTE ]

I agree with this. It's a somewhat low probability event, but probably still significant. Probably the most straightforward thing would be to decide on an average number of runs that are scored after the game goes past the ninth and add that to the total. I have no data for this, but it seems logical that more extra inning runs might be scored following a 9-9 tie than a 0-0 tie. On average, of course.

[ QUOTE ]
I'd say the biggest oversight of your method is that it doesn't take into account the ballpark factor. A game played between the Dodgers and the Rockies is likely to be higher scoring if it's played in Colorado rather than LA.

[/ QUOTE ]

Again, I agree, but only partially. The ballpark factor is already factored into the offense numbers for the home team. For defense numbers and road offense, I might be easy to generate a "ballpark multiplier" and just reference it in the spreadsheet according to who the home team is. I know that data for runs/game for each park is (or at least used to be) available online.

Thanks for the suggestions. When I have more time I'll continue the tinkering and backtest it a bit.

-w.a.

[wadea]

[ QUOTE ]

FLA/PHI(9) - NO PICK
W-PIT/MIL(7) - O -125
L-ATL/NYM(8) - O -104
W-COL/CIN(9.5) - U -117 (my line here is -118)
L-WSH/MIL(7.5) - O +118
HOU/STL(8.5) - NO PICK
W-ARI/SD (7.5) - O +118
L-SF/LAD (8) - O -109
W-KC/DET (8.5) - U +103
L-CWS/CLE(8.5) - O +125
L-NYY/BOS(10.5)- U +125
W-TB/TOR (9.5) - O +115
L-ANA/MIN(8.5) - O -112
W-TEX/OAK(8.5) - O -108
W-BAL/SEA(8.5) - O +110


[/ QUOTE ]

So the picks went 7-6 (+120) if my math is right. For picks NOT against the line - just picking O/U straight up and taking whatever the line happens to be - it went 9-6 (+144). I'll keep you posted,
-w.a.