[Stocks]

I'm working on a Pythagorean theorem based model - goals for versus goals against.

I have goals for, goals against, and goal difference for every team.

So for the Bruins its 2.67 goals for per game and 2.50 goals against per game for a 0.17 goal differential.

When you add up the goal differentials for all of Boston opponents it comes out to -2.14 which seems really high to combine into Bostons goals for and goals against numbers.

How should I combine a teams opponents into their goals for and goals against numbers.

Should I average them instead? The final numbers make a lot more sense like that.

[Axis]

Goal scoring is such a random event in a hockey game...I'd probably focus a model on something different.

If you're going this route, I'd use the averages.

[a4u2fear]

I don't know a lot about pythagorean models but I hope you've at least considered the following:

Splitting out goals for and against for home/away.

Compare the team averages to the league average and then adjust accordingly.

In my own personal research I've found that moving averages are VERY GOOD when modeling the NHL. There are lots of games and EVERY TEAM goes on many game losing streaks (while scoring very few goals). So a team that scored a lot in the beginning, it may not make a ton of sense to include those numbers in your current average.

[Stocks]

Thanks fellas

I plan to split home/away once a few more games are played and the numbers are more accurate. For now I'm going use .30 goal difference for home/away and adjust that up or down slightly for say a team like the Minnesota Wild are probably better then home on average and worst on the road then average compared to their overall record.

I'm also going to use exponential smoothing for goals for and against so it will adjust to hot and cold streaks more then just using averages.

That's going to be the basics of the model and I do plan to track some other stuff and then just use good old handicapping for the final numbers.

[James Marques]

Honestly, I would use medians as opposed to averages. Just my 2 cents.

[a4u2fear]

Honestly, I would use medians as opposed to averages. Just my 2 cents.

In hockey? Crazy? In a game where where 2 or 3 goals are almost always scored, it would be useless to use 2 or 3 as a standard. It's different in the nfl where there are many more points scored and it is drastic from week to week, but in the nhl you have tons of games and very little change to goals scored

[Miz]

db cooper

[Miz]

db cooper

[Stocks]

What do you guys think about capping blowout losses? And when I say capping I don't mean handicapping I mean taking a 9-0 win and readjusting it down to say 6-0 or 7-0.

This is obviously a bigger deal in football with so few games one blowout win or loss will skew the numbers too much and in hockey it will skew the numbers but I wonder how much of a difference it will make.

[James Marques]

In hockey? Crazy? In a game where where 2 or 3 goals are almost always scored, it would be useless to use 2 or 3 as a standard. It's different in the nfl where there are many more points scored and it is drastic from week to week, but in the nhl you have tons of games and very little change to goals scored

Well, in just a 2 second glance of last year's stats:

Team Goals Against Average (League Avg): 2.690
Team Goals Against Average (League Median: 2.699

Probably not enough of a difference to be significant. But again, I know nothing about hockey. But I'd be willing to bet if you used the league median rather than the league average for all the relevant stats (penalty/penalty kill, home/away, etc), your standard error would be reduced significantly.

[Stocks]

I just finished up my model and here are the first results

St Louis Blues 2.67 (-172)
NY Rangers 2.03

The numbers after each team are expected goals.

Well I do like the fact it's telling me to bet St Louis but -172 is a bit high I would have like to see between -140 and -150.

[James Marques]

How did you arrive at those numbers? I know you said Pythagorean, but did you use a Log5 formula (or some variation of it) to compare the offensive stats for each team vs the defensive stats for each team? Just curious.

[Stocks]

How did you arrive at those numbers? I know you said Pythagorean, but did you use a Log5 formula (or some variation of it) to compare the offensive stats for each team vs the defensive stats for each team? Just curious.

For my final numbers I had

St Louis 2.53 goals for and 1.98 goals against
Rangers 2.58 goals for and 3.03 goals against

So I divided St Louis goals for (2.53) in league average (2.72) = 0.93 then multiply that by Rangers goals against (3.03) = (2.82)

I did the same with New York and got 1.88.

Once I got those numbers I added 0.15 to New York and subtracted 0.15 from St Louis for home ice advantage.

So then that gives me

St Louis 2.67
New York 2.03

Would this method be correct?

[yak merchant]

Correct? I don't know, I'm not smart of enough to figure it out close form, and from some limited testing I think it is close but no cigar. When I first used Bill James log5 years ago in a baseball model, I did some half ass backtesting to validate it worked for strikeout percentages or some other baseball stat other than batting average (that was plastered all over the internet) and it seemed to hold. I would love to hear if anyone has ever done any validation for "non-percentages" (i.e. hockey goals, YPC, Soccer goals etc). When I first tried, I tried converting everything to "percentage" of league average and plugging it into log5. Fail. I've used the exact same formula as Stocks above but in my limited testing it seems that near the league average it underestimates and overestimates when you get too far from league average. i.e. if league average is 2.72 and team A offense is 2.9 and team B defense is 2.9 it predicts 3.09 (which from some football stats testing seems to be a tad low). However if league average is 2.72 and Team A offense is 5 and team B defense is 6 the formula works out to 11.02. Which might work for hockey (I doubt it), but from some YPC testing seems a bit too exaggerated. So I too, would love to hear any experiences with "log5" for whole numbers.

[James Marques]

For my final numbers I had

St Louis 2.53 goals for and 1.98 goals against
Rangers 2.58 goals for and 3.03 goals against

So I divided St Louis goals for (2.53) in league average (2.72) = 0.93 then multiply that by Rangers goals against (3.03) = (2.82)

I did the same with New York and got 1.88.

Once I got those numbers I added 0.15 to New York and subtracted 0.15 from St Louis for home ice advantage.

So then that gives me

St Louis 2.67
New York 2.03

Would this method be correct?

I think so, except I would use the league average (again I prefer median, but average is probably fine) goals against, not goals for. However, I doubt a model could be that simple and be profitable long-term. For instance, what if the Rangers goals against average is so high because of the high degree of competition they have faced (and vice versa for St. Louis)? You need some way to account for that. When I do college football modeling, not only to I look at each team's stats - I also create a power rating system to use as a weighting structure for those stats. I don't exactly know how you would go about doing that for hockey (I don't really do much modeling for pro sports outside of the NBA), but I would strongly advise you to look into it.

By the way, I did tail your St. Louis pick last night for solidarity's sake. Thanks for the winner!

[Stocks]

I think so, except I would use the league average (again I prefer median, but average is probably fine) goals against, not goals for. However, I doubt a model could be that simple and be profitable long-term. For instance, what if the Rangers goals against average is so high because of the high degree of competition they have faced (and vice versa for St. Louis)? You need some way to account for that. When I do college football modeling, not only to I look at each team's stats - I also create a power rating system to use as a weighting structure for those stats. I don't exactly know how you would go about doing that for hockey (I don't really do much modeling for pro sports outside of the NBA), but I would strongly advise you to look into it.

By the way, I did tail your St. Louis pick last night for solidarity's sake. Thanks for the winner!

lol thanks the Blues were a pick for me even before I did the model.

By the way I do account for competition and some other things I put those into my numbers before I divide and multiple them with opponents. I am thinking about doing something a little different with my competition numbers I'm just not sure how I'm going implement it yet. After I run the numbers for today I might make that change and compare it seeing there's a lot of games today.

[charris0528]

St Louis (2.71) @ New Jersey (2.29) ***Blues -105***Toronto (3.84) @ Arizona (2.54)*** Leafs -105***Nashville (1.80) @ Winnipeg (1.38) ***Under 5.5***

[James Marques]

lol thanks the Blues were a pick for me even before I did the model.

By the way I do account for competition and some other things I put those into my numbers before I divide and multiple them with opponents. I am thinking about doing something a little different with my competition numbers I'm just not sure how I'm going implement it yet. After I run the numbers for today I might make that change and compare it seeing there's a lot of games today.

Best of luck. Someone will no doubt come into this thread and beat you over the head about back-testing. I won't be foolish enough to say ignore them, but don't focus too heavily on it. Track all the games, all of your picks, totals, etc on a spreadsheet. Compare it to the Vegas line every week. If you want to use it to bet right now, sure go for it. Be very careful early on though. I didn't make a single college football bet for almost 3 years when I built my model (well, I'm lying... I did bet, just not solely based on what my model said); instead, I simply tracked everything. My numbers vs Vegas numbers, means, medians, mode, standard deviation, mean square error, bias.... everything. I learned that my model is not as good as Vegas's numbers, not by any stretch. But I found that some of the biases my model had could be used to my advantage. For instance, my model has a heavy bias towards favorites covering. After a few years of tracking that, I realized that when it does pick a dog to win outright (usually about 8 or so a week), it was hitting almost exactly 50% on those bets. Any gambler will tell you if you're only picking dogs ML and getting close to a 50% hit rate, you're making money. Now, those are pretty much the only sides bets I make. If your model is any good, you'll really see the pay off in betting totals.

Again, you're not gonna find an easy score modeling sports. But over time, you'll start to see where your model succeeds and where it fails. The most important aspect isn't so much the back-testing... it's the testing and tracking as you go along. Even if you don't have a nickel on the Bruins game, you should feel like you do because your model said they would win. The second most important thing: while you're tracking, don't lie to yourself. You pick a day/time to measure your numbers against the line at that moment. You can use opening lines, closing lines, mid-day, whatever... but they all need to be the same. You can't ignore the line in one game moving against you while at the same time taking another number that's moving in your favor. It'll make you think your model is radically successful, when in reality it actually is not.

[Stocks]

Thanks for the advise James.

Here's todays lines with no extra handicapping put into it just by the numbers.

Edmonton 2.81
Philadelphia 4.04 (-207) I do like Philly tonight pre model

Carolina 2.32
Columbus 3.28 (-199) Columbus has some major injury problems line not adjusted for that.

St Louis 3.12 (-167) I also like St Louis tonight pre model
New Jersey 2.41

Detroit 2.74 (-131) Pretty close to actual line.
Ottawa 2.40

Nashville 1.82 (-169) Nashville was one of my plays of the night pre model
Winnipeg 1.40

Vancouver 2.69
Colorado 3.03 (-127) close to actual line

Calgary 3.37 (-154) This one is interesting to me.
Washington 2.71

Florida 1.15
Boston 2.34 (-415) I like the Bruins but -415 is a bit much (not counting injurys)

Chicago 1.87
Montreal 1.94 (-107) Probably what the actual line should be.

Pittsburgh 2.72
Minnesota 2.97 (-118) With a little extra home ice line would be very close.

Los Angeles 2.57
Dallas 2.95 (-131) I like this line.

Toronto 3.47 (-190) Leafs deserves to be favored but -190 a bit high
Arizona 2.52

[Stocks]

So it seems like using Pythagorean on each team to get winning % for each side then use log5 gets much more accurate results.

For the Toronto at Arizona game I had

Toronto 2.58 goals for and 2.55 goals against
Arizona 2.67 goals for and 3.63 goals against

Using Pythagorean

Toronto 50.6%
Arizona 35.1%

Then use log5

Toronto 57.4% or -134

For the Bruins game I still got a high number -379 but it was lower then before. Did they deserve to be that big a favorite? probably not I could have adjusted the goals for and against slightly for some key injuries on offense and defense which would have brought it down some and that may have been a pretty good line. The Panthers do suck and I had them as having the easiest schedule in the league by far.

The Calgary and Washington game I actually got the exact same number -154.

I do like this method better for getting the final line its a lot more accurate.

[James Marques]

I think you're onto something here. Nice work.

[Stocks]

Here are the results for today

Montreal -186
Buffalo

NY Islanders
Anaheim -410

Detroit -163
NY Rangers

And here are the results after I changed my strength of schedule format.

Montreal -176
Buffalo

NY Islanders
Anaheim -376

Detroit -146
NY Rangers

1. I think the Montreal line is pretty dead on.

2. Anaheim line is obviously high even without the Ducks injury problems.

3. I actually like New York to win tonight but that has nothing to do with numbers just my gut. As for the line seems high for Detroit. I'm guessing Bookmakers are using some of last years stats for the Rangers especially on defense where my lines are 100% based on this season.

Any thoughts or feedback is definitely welcome.

[Axis]

Are you accounting for backup goalies starting? Like tonight Gus is starting for Detroit and Montreal is trotting their backup out there as well.

[Stocks]

Are you accounting for backup goalies starting? Like tonight Gus is starting for Detroit and Montreal is trotting their backup out there as well.

No these are just base numbers using goals for and goals against. To account for goalies or injuries or whatever else I would adjust the goals for or goals against like say if Montreal's goals against were 3.15 and Price wasn't starting I would probably adjust it up to like 3.20 or something like that depending on the situation.

As for injuries and what not is there any advice on how much to account for certain players? Say if Crosby was out for the Penguins how much would you adjust the Penguins offensive numbers?

[DrunkHorseplayer]

What do you guys think about capping blowout losses? And when I say capping I don't mean handicapping I mean taking a 9-0 win and readjusting it down to say 6-0 or 7-0.

This is obviously a bigger deal in football with so few games one blowout win or loss will skew the numbers too much and in hockey it will skew the numbers but I wonder how much of a difference it will make.

If you're going to try and formulate a hard-core system, you should completely eliminate all blowouts, win or lose.

[Stocks]

If you're going to try and formulate a hard-core system, you should completely eliminate all blowouts, win or lose.

Why completely eliminate blowouts? Would'd it be better to cap it instead.

Say if the Penguins won 9-0 why take away all credit for that game they still played really good. I figured it would be better to somehow cap the amount of goals at say 6 or 7.

[trytrytry]

I don't know a lot about pythagorean models but I hope you've at least considered the following:

Splitting out goals for and against for home/away.

Compare the team averages to the league average and then adjust accordingly.

In my own personal research I've found that moving averages are VERY GOOD when modeling the NHL. There are lots of games and EVERY TEAM goes on many game losing streaks (while scoring very few goals). So a team that scored a lot in the beginning, it may not make a ton of sense to include those numbers in your current average.

Whoa move this person to the "Read everything and go back and read every past post category"..

ok note made

[Stocks]

Here are todays results. Still a lot of really big favorites and dogs. I don't mind if it gives me opinionated results makes it easier to bet if it picks the right side but I still think I'm missing something to make things a little more evened out. Maybe it's still too early in the season to be 100% using this regular season results.

Edmonton
Boston -256

Minnesota
Ottawa +132

New Jersey
St Louis -273

Nashville
Dallas +182

NY Islanders
LA Kings -151

Florida
Philadelphia -320

Calgary
Tampa Bay -114

Pittsburgh
Winnipeg +286

Toronto
Colorado +102

Vancouver
San Jose -121

[flsaders85]

As pointed out in earlier posts, factoring for player injuries, minor leaguers, etc NEEDS to be accounted for. This is something that is also applicable for basketball and baseball. Using team based stats won't give you an edge IMO.

With this said, here's my take on a player based model for NHL that I use:

1. Learn to make your own player projections or acquire from another source. The important player stats for skaters IMO are projected shots, goals, missed shots, blocked shots. For goales, projected shots against and save %.
2. You must make some sort of prediction of a team's forwards, defense, and goalie for a given game.
3. Learn what Corsi and Fenwick are....and how those translate to actual Shots on Goal and Goals. Basically, this is what I am predicting when I simulate a game using my projected lineups and projected player stats.
4. Learn how to develop a Monte Carlo simulation or use regression analysis to develop this sort of model.

Again, this is all opinion and what I use. You'll realize when you can actually predict a game using a method similar to this, your lines will be more accurate to closing lines. Not poking fun here, but Anaheim won't be -410 over the Islanders...I actually had Anaheim as a -130 fave after Perry and Getzalf illnesses were announced.

[Stocks]

As pointed out in earlier posts, factoring for player injuries, minor leaguers, etc NEEDS to be accounted for. This is something that is also applicable for basketball and baseball. Using team based stats won't give you an edge IMO.

With this said, here's my take on a player based model for NHL that I use:

1. Learn to make your own player projections or acquire from another source. The important player stats for skaters IMO are projected shots, goals, missed shots, blocked shots. For goales, projected shots against and save %.
2. You must make some sort of prediction of a team's forwards, defense, and goalie for a given game.
3. Learn what Corsi and Fenwick are....and how those translate to actual Shots on Goal and Goals. Basically, this is what I am predicting when I simulate a game using my projected lineups and projected player stats.
4. Learn how to develop a Monte Carlo simulation or use regression analysis to develop this sort of model.

Again, this is all opinion and what I use. You'll realize when you can actually predict a game using a method similar to this, your lines will be more accurate to closing lines. Not poking fun here, but Anaheim won't be -410 over the Islanders...I actually had Anaheim as a -130 fave after Perry and Getzalf illnesses were announced.

Thanks bud I'll look into everything you mentioned.

[DrunkHorseplayer]

Why completely eliminate blowouts? Would'd it be better to cap it instead.

Say if the Penguins won 9-0 why take away all credit for that game they still played really good. I figured it would be better to somehow cap the amount of goals at say 6 or 7.

Because blowouts usually happen in circumstances that you simply can't predict; one team phones it in, lucky breaks, etc.