[benrama]

Setting up this thread for That Guy, angelo, rohan and anyone else who wants to brainstorm approaches for modelling NRL games. Aim is to try to see how one might use stats combined with situational analysis to pick accurately ATS at a rate of 55%+.

I'll start by giving my current thoughts.

1) You need at least 5 games worth of stats before you even think about using a model accurately. You also have to tread carefully the week before and after state of origin, and in the last games of the season when teams have nothing top play for. That leaves you at most only 18 weeks or so of "normal" betting conditions to use a modelling approach.

2) NRL is a small market, and heavily driven by "squares" who go off what the commentators on Ch 9 and FOXTEL say and what they read in the Daily Telegraph - these means there is a much higher probability of using modelling to determine an inefficiency in the lines being set than sports like NFL, MLB and NBA. The key is to not "become" a square, and stats can help you do that by not over-emphasising last weeks game or plays that stood out, games you watched closely vs over 10 beers at the local, etc

3) Focus should be on only the absolute, most important stats. Here is my current list, interested to see what others have here.

Offensive efficiency
Metres gained/set
Points gained/set
Line breaks/set

Defensive efficiency
Metres allowed/set
Points allowed/set
Line breaks allowed/set

Error rate
Errors/set
Missed tackles/set

Others? Kicking game?

My thinking is to work out these numbers for all teams, assign 1 to the lowest number, 10 to the highest and allocate numbers to other teams depending on where they fit along the range - i.e. let's take points gained/set, let's say the best team gets 1 point/set, the worst gets 0.5 point/set. Range is 0.5, divide by 10, so each increment is 0.05. A team then that gets 0.8 points/set would be ranked a 6.

All these stats should be able to be broken down by:

Home vs Away
Top 8 vs non-Top 8

And once you have (say) 10 games you might arguably also eliminate any outliers - the games on either end of the extreme for a given team.

4) Along with these stats you assign each team a "Home" ranking and an "Away" ranking depending on the strength of their home ground advantage, and their relative away vs home performance.

Home ranking can be done by looking up historical stats, going back no more than 2-3 seasons, and allocating it a number from 1-10. I haven't done it yet, but I expect NQ, NZ, Melbourne, Brookvale to be the top home ground advantages.

5) Along with the stats above you want a sense of a teams momentum, to do that I'd again calculate a number out of 10 to factor in how they performed in the last "X" games, I think I'd start with "X" being 5.

So you'd use:
Wins in last 5 games
Key stat improvement in last 5 games vs season average

Is a team getting better and better? Regressing? You want to somehow capture this.

6) Once you have all these numbers out of 10 you then just need to determine how then it calculates what you think the "true" game spread should be This requires a bit of trial and error and - ideally - back-testing and regression analysis. This is really the "magic" so I expect it to take the most time.

Even after all this you still need to factor in situational elements - motivation, weather, etc - and then of course injuries. The ideal world is where you have a model that can very quickly show you were there is potential line value and then allow you to focus your time accordingly.

These are just initial thoughts - interested to hear other points of view on them, devil's advocate, etc.

If I get time I'm going to run through the methodology above with a given game given the stats from this season so far and see how it might be applied.

[rohan22no]

Hi Benrama,
Firstly - All sounds great.

Just want to quickly mention my experience with NRL and other models for different sports I've created.

As I mentioned really briefly in the other thread, my NRL models record is 63.2% (mean error 14.01 pts) over the last 15 seasons. Im purely working off Team scores, and a relatively simple Home ground advantage formula.

simple example..(team A playing at home)


Team A rating = 8
Team B rating = 5
Team A's HGA = 4

(Team A + Team A's HGA) - Team B = Home team predicted margin

(8 + 4) - 5 = 7

Predict Team A by 7

...................

another example

Team A rating = -2
Team B rating = 9
Team A HGA = 3

(-2 + 3) - 9 = -8

Predict Team B by 8.

Teams ratings adjust when the actual result of the games are input into the spreadsheet and the error is calculated. The function is called exponential smoothing.

Of course, team scores are not the best measure of how good a team actually is, due to the massively random mess that is the nature of NRL games. The method you're proposing above I have no doubt would be significantly more accurate than the method I've used, or any other NRL model available (free or paid) available now.

I totally agree with your Points in your OP. Models work the best in the "normal" rounds that are not interrupted by Origin weeks, or games right at the start of the season. During the interrupted weeks, I think a model will be useful to an extent, but will need subjective analyisis on top of that to consider the factors that are not easily quantifiable into a number, such as...

- players unavailable due to injuries, suspension, origin etc
- known changes within the club (new coach or whatever)
- games where teams will have different motivation levels ("must win" games, or teams playing for nothing)


The biggest issue in taking all the stats you mentioned into account is simply the time taken to record and input them all. Have you got past data for the stats or do you plan to begin recording from now? IMO, we would ideally want 5 seasons of data in which to back-test the model on and make the tweaks and ajustments. Have you got this data or can you get it (anyone?) ?

Re: point 6 - I disagree that back-testing takes the most time. In my experience what takes the most time (by far) is the manual data entry of scores/fixtures etc in the correct format...once all the data is entered I've found its quite smooth sailing. Once set up correctly, you can tweak a "weighting" of a certain stat and see it instantly applied to thousands of games in the database, and instantly look at how it changed the end results (% games tipped correctly, avg error etc)

Anyway, thats all the things I can think of at the moment, look forward to discussing it more with you soon.

If any of you guys have skype, add me. (user: rohan22no)

Cheers

[That Guy]

Hey mate,
Sorry - must have missed this thread yesterday.
I think you're onto something with all the info above.. would be interesting to see how the methodology would perform with the stats so far this season.
This bloke has managed some pretty good figures forecasting games and assigning 'power play' numbers based on offense / defense to teams: http://footyforecaster.com
What are you hoping to achieve as an output?
Most models predict a winner / loser but obviously if it can be further refined for OVER / UNDER or margins it would be much more valuable.

Awesome work so far. Thinking out aloud.. I wonder whether comparing past closing lines to the actual result is beneficial?
I have lines / actual game results for all games going back about 5-6 years.

[goty0405]

Gents,

The NRL is one of my key areas of work around modelling so very keen to see this chat progress. I think I talked to rohan last year too (if so mate, very sorry I never ended up getting back to you...things went a busy in my life and sport modelling took a back seat)

Anyway I'm back and just checking in now so I can easily find this thread tomorrow when I'm awake and got more time to comment deeper.

My first contribution is a very simple one but very good for NRL. Tested over 3 years it hit 69.8% (with 1 simple filter). It uses a simple stat I created which I dubbed "Relative Efficiency"...

Relative Efficiency (RE) = Team Points Scored Per Possession + Opponents Points Conceded Per Possession

So for each match you calculated a RE rating for each team based on the previous stats (points for/against) of these teams. Then you just consider which team has the better RE. If you only bet on home team when the home team has the best RE then it went 213/305 (69.8%) from 2008 to 2010. I've since developed a better model so didnt run this last year but someone remind me to run it over my database for 2012 tomorrow. If you also considered away teams I think it dropped to about 61-62% strike rate.

Some other points I've noted:

• 3 games is ABSOLUTE minimum for modelling but yeah 5+ is best.
• Location specific performance is generally overrated
• Wet weather is not strong indicator of low match score

Ummm that's all I can think of off the top of my head. Ended up spending more time writing this response than I planned. Now its definitely time for bed

[Blackroc78]

[benrama]

Great conversation here guys, have to log off and hit the hay here, but will get back with more detailed thoughts in the next few days. Just quickly though: I don't believe using more than 1 season - 2 max - is relevant for modelling purposes - too much changes in that time. Maybe if you assigned a "relevancy factor" depending on how similar teams are personnel and coaching wise, etc it might be different - but that would take way too much time. Having dabbled in some Forex technical stuff I also know that back-testing accurately means nothing about what will happen in the future. Really I think we are just trying here to convert some logical thinking about analysing games into a short-cut method to then apply situational/subjective factors.

What I was thinking was that we could choose one game each for the coming weekend and explain how we might model the game to bring some of this to life, using real stats from this season and last seasons. More on this when I get some time.

[rohan22no]

@ Thatguy.
Footyforecaster is run by a guy called Greg, Ive been in semi-regular contact with him via email over the years. His model is very similar to mine, using exponential smoothing and the simple T1 + HGA - T2 = T1 margin formula. Its simple, but very effective. From what I understand he mainly put the website up as a bit of an exercise for himself and isn't particularly interested in actively improving it.

As for Models, In my experience most of them predict a margin, rather than just the team. The exeption to this is Alan Mcabes model which uses neural network (far too complicated for me to understand, but maybe some of you guys can get your head around it). www.mymait.com

When you say do have the lines, do you mean you have the +/- pts start for each team? (sorry, im not too familiar with the lingo). Do you have any other stats such as 1-12 or 13+ margins, or the over/under scores? They would be most valuable. I've got all the H2H odds since 2004 and all game results since 1997.

@goty.
Yes, we did talk a year or so ago I was disappointed I never heard from you. Think I've still got you on Skype though, so hit me up when you have a chance.

Re: Relative Efficiency, I'm not sure I understand you fully.

If Roosters have 40 Possessions in a game and score 20 points, does that mean their efficiency is .5 ? How do you calculate their opponents points conceded per possession? Would it be averaged out over the year (e.g they've had 350 possessions against them and conceded 200 points) or are more recent games given a bigger weighting?

Re: Your 2008-2010 results. Could you please explain exactly what your figure 69.8% represents? Is this the amount of times the home team has the best RE and you bet on them? If I'm correct, Is that profitable? From my records, over the same period the average home team odds was $1.83...so im guessing your mainly backing the favorites here? Anyway, I'm possibly way off track with what you were saying, so Ill let you respond again and hopefully ill understand it a bit better.

Re "Location specific performance is generally overrated", are you saying that too much weight is given to a perceived home ground advantage? If you could go into more detail with this and explain how you arrived at your conclusion that would be awesome.

Re "Wet weather is not a strong indicator of a low match score", are you saying you've got a database of past games including the time of game and weather conditions? And you've found the average PPG of these to be no different from games in dry conditions? Again, if you could go into detail here that would be awesome

Hopefully you guys haven't found my questions too demanding, I promise I'm not nitpicking - I'm just extremely interested in these kind of statistics and am keen to learn.

Anyway, met Angelo in real life for a coffee today, was great to chat to someone who's approach is very very different to my own. Whereas I think my strength and love is crunching the numbers and doing statistical analysis, Angelo's insight and subjective observations on the game was fantastic. Look forward to collaborating with him and the rest of you guys further with this.

Don't mean to derail thread, just thought I'd post some hopefully thought provoking and interesting stats for the NRL that I've derived from my database (games recorded since 97, odds since 04).

Games since round 1, 1997 - 2853
No of Home Team wins - 1671 (58.6%)
No of Away Team wins - 1138 (39.9%)
No of Draws - 44 (1.5%)

Average home team score - 24.15
Average away team score - 19.43

Average winning score - 29.29
Average losing score - 14.3
Average winning margin = 14.92

average total game points - 40.5

biggest win margin - 70 (Eels v Sharks rd 25 2003, 74-4)
most points in a game - 102 (Raiders v Knights, rd 2 2006, 32-70)


(stats since rd1 2004, sample of 1554 games)
Home team are favourites - 67%
Away team are favourites - 33%
Favourites win the game - 63%

biggest upset - Roosters v Titans, rd 18 2008. $1.09/$10.40. Titans won 32-28
biggest difference in starting odds - $1.03/$12.50 (Storm v Rabbitohs, rd 26 2008)

If you guys are interested in any specific stats during this period, just ask and ill do my best to dig it up for you. Look forward to chatting to you guys soon.

R

[benrama]

Sorry for the delay in replying:

Hi Benrama,
Firstly - All sounds great.

Just want to quickly mention my experience with NRL and other models for different sports I've created.

As I mentioned really briefly in the other thread, my NRL models record is 63.2% (mean error 14.01 pts) over the last 15 seasons. Im purely working off Team scores, and a relatively simple Home ground advantage formula.

Is that winning % ATS or on the ML? I'm assuming ML as otherwise you'd already be rolling in the $ winning at 63+%!

simple example..(team A playing at home)


Team A rating = 8
Team B rating = 5
Team A's HGA = 4

(Team A + Team A's HGA) - Team B = Home team predicted margin

(8 + 4) - 5 = 7

Predict Team A by 7

...................

another example

Team A rating = -2
Team B rating = 9
Team A HGA = 3

(-2 + 3) - 9 = -8

Predict Team B by 8.

Yep - that's exactly what I had in mind with all my "ranking out of 10" categories. I haven't yet figured out what my "formula" will be with all these rankings, that's what I need to do some work on.

Teams ratings adjust when the actual result of the games are input into the spreadsheet and the error is calculated. The function is called exponential smoothing.

Great point, hadn't thought of this. It would make sense for exponential smoothing to give better results - more recent performance should be weighted higher.

As I've looked at this too I've realised that I need a "schedule toughness" ranking to gauge how tough the competition has been for a given team.

Of course, team scores are not the best measure of how good a team actually is, due to the massively random mess that is the nature of NRL games. The method you're proposing above I have no doubt would be significantly more accurate than the method I've used, or any other NRL model available (free or paid) available now.

Cheers - definitely interested in getting this model up and running. I plan to use it starting from week 6, so have two and a bit more weeks to get something up and running. I have no interest in getting $ for the model, but I'm most interested in having a few key people to work with the model and generate picks for a round. There's an opportunity cost of sharing the model that is outweighed by the feedback you get. Plus - I don't think books would take any notice of such a model even if successful. Hence why I'm liking having this discussion out in the open.

I totally agree with your Points in your OP. Models work the best in the "normal" rounds that are not interrupted by Origin weeks, or games right at the start of the season. During the interrupted weeks, I think a model will be useful to an extent, but will need subjective analyisis on top of that to consider the factors that are not easily quantifiable into a number, such as...

- players unavailable due to injuries, suspension, origin etc
- known changes within the club (new coach or whatever)
- games where teams will have different motivation levels ("must win" games, or teams playing for nothing)

Agreed. Just like with sports like the NBA though - for example - I think having a star player out through injury actually tends to put the value firmly on the team affected as punters and therefore books over-compensate.

Motivation and situational factors are so subjective - I'd be looking to use them more to rule out a play than rule it in. There will always be key spots however - coaching change as you mention, comments said during the week, and most importantly games towards the end of the season where finals motivation comes into play.

The biggest issue in taking all the stats you mentioned into account is simply the time taken to record and input them all. Have you got past data for the stats or do you plan to begin recording from now? IMO, we would ideally want 5 seasons of data in which to back-test the model on and make the tweaks and ajustments. Have you got this data or can you get it (anyone?) ?

Agreed this can be time consuming to do. All the stats are available it's just a case of writing simple software to do screen scrapes and put them into Excel or a DB for analysis. I don't currently have this data, but I think I have a source for getting it.

Until I do I've just got this seasons stats and will work off that.

Re: point 6 - I disagree that back-testing takes the most time. In my experience what takes the most time (by far) is the manual data entry of scores/fixtures etc in the correct format...once all the data is entered I've found its quite smooth sailing. Once set up correctly, you can tweak a "weighting" of a certain stat and see it instantly applied to thousands of games in the database, and instantly look at how it changed the end results (% games tipped correctly, avg error etc)

Sorry should have been clearer, I was only referring to the "magic formula" that combines all the key stats together as what will take the most time to get right. Agree that the back-testing is straight forward once everything is in a DB. What DB/tools do you currently use?

[benrama]

Hey mate,
Sorry - must have missed this thread yesterday.
I think you're onto something with all the info above.. would be interesting to see how the methodology would perform with the stats so far this season.
This bloke has managed some pretty good figures forecasting games and assigning 'power play' numbers based on offense / defense to teams: http://footyforecaster.com
What are you hoping to achieve as an output?
Most models predict a winner / loser but obviously if it can be further refined for OVER / UNDER or margins it would be much more valuable.

Awesome work so far. Thinking out aloud.. I wonder whether comparing past closing lines to the actual result is beneficial?
I have lines / actual game results for all games going back about 5-6 years.

Output I'm after is really just a projected score to compare against the actual line. A capper I really respect in the NBA doesn't work with over/unders and actually assumes the total line is correct and generates his score relative to that. There's real merits to that.

Definite some possibilities in comparing closing lines to results to see trends. In the NBA home dogs are undervalued historically by nearly 4% (about half comes from undervaluing the dog and half from undervaluing the home advantage), that alone could make you a profitable sports better if you used bankroll management.

I'd definitely be interested to know, for e.g.:

ATS results on Monday night football
Results the weak after a top 8 team gets beaten by more than (say) 10 points
ATS results for each team home and away

Going back 5-6 seasons I think isn't useful as teams change too much. What would be great to have though would be some of these historical ATS results weighted by "team similarity" to see how applicable they might be - "team similarity" being derived by looking at current and past playing and coaching personnel.

What format do you have the data in?

[benrama]

Gents,

The NRL is one of my key areas of work around modelling so very keen to see this chat progress. I think I talked to rohan last year too (if so mate, very sorry I never ended up getting back to you...things went a busy in my life and sport modelling took a back seat)

Anyway I'm back and just checking in now so I can easily find this thread tomorrow when I'm awake and got more time to comment deeper.

My first contribution is a very simple one but very good for NRL. Tested over 3 years it hit 69.8% (with 1 simple filter). It uses a simple stat I created which I dubbed "Relative Efficiency"...

Relative Efficiency (RE) = Team Points Scored Per Possession + Opponents Points Conceded Per Possession

So for each match you calculated a RE rating for each team based on the previous stats (points for/against) of these teams. Then you just consider which team has the better RE. If you only bet on home team when the home team has the best RE then it went 213/305 (69.8%) from 2008 to 2010. I've since developed a better model so didnt run this last year but someone remind me to run it over my database for 2012 tomorrow. If you also considered away teams I think it dropped to about 61-62% strike rate.

Some other points I've noted:

• 3 games is ABSOLUTE minimum for modelling but yeah 5+ is best.
• Location specific performance is generally overrated
• Wet weather is not strong indicator of low match score

Ummm that's all I can think of off the top of my head. Ended up spending more time writing this response than I planned. Now its definitely time for bed

We're thinking along the same lines here, if you have a better model than one just using RE for home teams then definitely interested to hear more about it.

With location I tend to agree - you only have a small sample size of games played on certain grounds (for away teams) as I think anything beyond 2 years back is irrelevant data

Cheers for popping into the thread and adding some comments.

[benrama]

@ Thatguy.
Footyforecaster is run by a guy called Greg, Ive been in semi-regular contact with him via email over the years. His model is very similar to mine, using exponential smoothing and the simple T1 + HGA - T2 = T1 margin formula. Its simple, but very effective. From what I understand he mainly put the website up as a bit of an exercise for himself and isn't particularly interested in actively improving it.

I've had a brief look at footyforecaster, in some ways I'm a fan of keeping things simple, but as the great quote goes, "everything should be made as simple as possible, but not simpler"

Part of what has inspired me to do some NRL modelling is that I am a big fan of a NBA modeller that has had some awesome results, and I think his approach is right on the money for ANY sport. In the NBA there are four key factors:

FG efficiency
Offensive rebounds per possession
Turnovers per possession
FG per point scored

That combined with defining home and away factors can be highly effective in predicting outcomes; just wish I'd found all this a few years ago

As for Models, In my experience most of them predict a margin, rather than just the team. The exeption to this is Alan Mcabes model which uses neural network (far too complicated for me to understand, but maybe some of you guys can get your head around it). www.mymait.com

Yep - seen that, and I have a background in computer engineering so I get the neural network stuff. Far too complicated for what we are looking at here - at least initially. The basic idea though - to have a model that "self-learns" to produce better predictions is a good one. I think we can do pretty well without that extra complexity however.

Anyway, met Angelo in real life for a coffee today, was great to chat to someone who's approach is very very different to my own. Whereas I think my strength and love is crunching the numbers and doing statistical analysis, Angelo's insight and subjective observations on the game was fantastic. Look forward to collaborating with him and the rest of you guys further with this.

Awesome! It's been great having angelo join this forum, I think situational capping always has it's place, and if you are watching enough of the games you can pick up things that stats will never tell you. Part of my issue though is that I simply don't have the time to watch enough footy these days - and being a fan of both AFL and NRL makes it even harder! Modelling for me is a way of trying to short-cut situationally. I've seen early in this season what happens when I make picks without being able to watch enough of the games - I've started with probably my worst results in the last 3 years.

Interesting stats you posted, I wonder if you calculated the draw % when it was two top-8 teams playing what it would look like? (no need to answer it just thinking out loud)

[kingsr]

Can the geniuses on here tell me if NEWCASTLE will make the top 8 this year? LOL

Brilliant thread guys, good to see the different angles you guys use in betting. I'm personally a situational + stats better, I use a bit of both to determine my bets. More so the situational though, because I watch a fair bit of the games.

Great thread thanks OP!

[Farlad]

Hi Guys

I thought i might try and revive this thread about one year later. I am very interested in trying to develop an NRL model along the lines of what you guys have been discussing here. My first hurdle has been getting ATS data for the past. I am in the fortunate position of having some spare time to be able to devote to this project.

I am new to the forum so i should say a little about myself. I am a lawyer who doesn't practice anymore. I have been a profitable poker player for the last 5 years, mostly live play. I have taken sports betting more seriously in the last 2 years and was profitable on NRL last year. This makes me new to the game.

Anyway I have noticed the contributors of this thread are mostly still posting here and im wondering if anyone would like to update with their progress or if anyone is interested in a collaboration to discuss the idea further or helping me try to track down the data i need.

I hope we can revive this very promising thread......

[Tim Gerry Mander]

I have some spreadsheets I can flick you.

The original one I got had the Sportstab spreads for NRL from 2001 to 2009 and AFL from 2003 t0 2009.

I was updating it after that for season 2010, but I am missing 2011, 2012 and the current season.

Happy to discuss my model too, even though it is more situational focused than statistical.

[Farlad]

TGM

Thanks a lot. That would be great. At present i just have basic score data for 2009-2012 with more detail for this year. I dont have PM privileges yet perhaps you could PM me with an email or something. Id love to talk further about your model as well. Look forward to it.