Probabilistic Prediction of Football Matches - A Discussion

[Grex]

Thanks, Loon..... I will certainly put mine in. But just a few points:

Loon, please make sure that the cells that calculate the "weighting factor" and "weighted average" are protected to avoid unintentional deletion.

Also, presumably we need to quote our system's yield for that particular League only (i.e. English Premier, for the moment). So those of us who lump all leagues together, we are going to have to do some "homework" to separate them

I am not totally convinced that this is going to produce anything positive in the sense of increased accuracy.... after all, with the weighting system and the resultant "weighted averages", we are merely diluting the predictions of the strongest system with those of the weaker systems. It may be better to use this tool to establish which system is performing the strongest and not apply any weighting.

Anyway, early days.

[Grex]

BTW, if we access and open the Excel file as an attachment to somebody's post, then input our data, and close it.... is the version on the Forum automatically updated.. or does it retain the original version??

[I reached 200 posts..... I was expecting to upgrade to THREE rotating balls.... but NO ]

[lunatism]

Originally Posted by Grex

BTW, if we access and open the Excel file as an attachment to somebody's post, then input our data, and close it.... is the version on the Forum automatically updated.. or does it retain the original version??

[I reached 200 posts..... I was expecting to upgrade to THREE rotating balls.... but NO ]

what i'll do is this - protect everything that needs protecting to stop stuff from being accidentally modified. And when the file changes, we'll have to post a new version on here

which of course poses a problem of having upto date spreadsheets! So how about this, i do all the posting, and people send their spreadsheets to me, that what there will be no such problems?

[lunatism]

damn premiership!!!

i was just working out what my yield would be from the predictions that i have got (going back to 6th jan...you all know that by now, right?!!). And know what i find?

	OVERALL	HOME	DRAWS	AWAY
bets	98	47	25	26
wins	55	31	10	14
losses	43	16	15	12
strike	56.12%	65.96%	40.00%	53.85%
staked	980.00	470.00	250.00	260.00
return	1127.70	503.60	503.60	279.10
p/l	147.70	33.60	253.60	19.10
yield	15.07%	7.15%	101.44%	7.35%
average odds	2.25	1.65	3.53	2.12

obviously i'd be a lot happier about this if i'd known sooner those stats sooner - especially for the draws. And the reason for 98 selections is twofold - the first, already explained that i only go back to 6th jan. And the other, is that if any two ratings for a fixture were identical (eg. if Liverpool v Everton was 25% for home and 25% for away), then it would not be included.

[lunatism]

so now another question - with the yields, do we just look at overall yields? Or should we split it between homes, draws and aways? Because if we did split it, it would obviously be more "accurate"

[Grex]

Originally Posted by lunatism

so now another question - with the yields, do we just look at overall yields? Or should we split it between homes, draws and aways? Because if we did split it, it would obviously be more "accurate"

I think perhaps we are getting a little carried away here..... I suggest we just try to keep it simple for the time being i.e. use overall yield for the league in question. Later, when the logistics of inputting data etc. are more familiar, then we can look at tweaking the system.

Of course, this is the opinion of 1 out of 4.... I will go with the majority decision.

BTW, Loon, congratulations on the Premiership Draws.... 101% yield

[lunatism]

Originally Posted by Grex

I think perhaps we are getting a little carried away here..... I suggest we just try to keep it simple for the time being i.e. use overall yield for the league in question. Later, when the logistics of inputting data etc. are more familiar, then we can look at tweaking the system.

Of course, this is the opinion of 1 out of 4.... I will go with the majority decision.

BTW, Loon, congratulations on the Premiership Draws.... 101% yield

haha - dont know how that slipped under the radar. Of course, not from a big sample so i guess just watch it at the minute!

well, how about this - i give people the option to put in overall, home, draw and away. and if they dont want to put in one, then they just fill it with the same yield as the overall - fair, right?

ok - here it goes, i've made a sheet for each of us - all we need to fill in are the percentages. EVERYTHING is automatic - except for the manual input of fixtures (don't worry - i'll do it week by week), and the input of the percentages.

There is a sheet that looks at all percentages side by side, you can look at them on the user sheets and then the Average sheet gives the weighted odds (which you will notice is a one-sided because it only has my yield, and i've put my percentages in the place of all of yours so i could make sure it worked). let me know if it's ok

seems to work fine for me - but i guess one of you can fine a problem with it!

SHOCKING - seen how big the damn file is? Would be a shade of that if i'd made it in office 2007 i guess it's 3.6MB!! but for some reason PL won't let me post it, so just email me and i'll reply with the file (know yours Grex, so it's on its way!)

[Grex]

Received the File.... thanks, Loon.... looks impressive..... let´s hope the results justify all your hard work .

Got nothing better to do tonight (don´t suppose the France-England match will be particularly exciting - even if they show it in this part of the world!), so I will try to work out my yield percentages.

[Rushian]

Ok as promised, here is a more considered view of the posts previously in the thread.

First of all, this has indeed snowballed into a systems competition. Personally I believe that, that in itself is definitely an interesting challenge but only if done correctly. And without trying to sound like a pessimmist, I don't think comparing systems solely in terms of yield is necessarily the right thing to do. For example, a number of issues arise:

• what if the yield achieved by a particular system is a result of a long series of short-odds bets, whereas another system results in the same yield but with longshots (and maybe with an equal number of bets)? Which system is the most stable?
• What about the number of bets? A similar yield may be achieved with a few bets per year whereas another system may result in many picks each week? How do you compare these figures?
• What if a modeler uses a single bookie (with the overround that betting through a single bookie entails) and nevertheless achieves a 3% yield, whereas another punter has a 5% yield from a betting exchange? Which figure is higher?
• etc, etc...

Having the above in mind, my belief is that it would not be prudent to combine a number of system probabilities using figures which are not necessarily comparable as weights (such as the yield). Should people want a "systems competition" then a level playing field should be found, starting by setting a few rules in terms of the odds to be used, the number of bets, etc, etc .... Then you would know whether a 5% yield is indeed better than (or at least can be compared to) a 4% yield. Just my thoughts.

In terms of what the actual probabilities may indicate, is whether people are on the same wavelength with respect to predicting the chances of each team winning. So for example, one would know that his/her probabilities of a draw are sreiously out-of-line with the rest of the systems, or that the favourite odds are not as high as they should be etc. Now, if we wanted to combine these probabilities, maybe we would find some factors which would produce systems which are (on average) closer to the true (uknown) probabilities of a Home Win / Draw / Away Win in a given match. This could be done possible through something like regression or something fancier than that.

I think it was Grex who mentioned "Sniffy", suggested in the sticky thread on top of this forum. The main difference of this proposal to that one, is that "Sniffy" would be a way of maintaining all system picks under the same format. This would result in, as far as my technical knowledge allows me to understand, a structured database format of all bets which could be queried (is that a verb? to query?) to return either the most succesful systems, the most popular choices as well as the historical performance of these system etc. The main difference is that the odds on offer are taken into account. In this proposal, the odds on offer are completely ignored. We are just trying to see what are the estimated probabilities of HDA according to the different systems. And through this, to improve our systems overall.

So to keep it simple, and to start with, we could focus on the Premiership. A list of the Premiership fixtures can be set up in an Excel sheet which can be dstributed to anyone interested. People who agree to participate in this experiment could then send (by email or otherwise) their estimated probabilities in the common format to a "compiler" as lunatism put it, who in turn could summarise the results in tables in the thread. This could be done before the matches to ensure fairness, although we are all adults and I am sure one would not be so childish to change their figures if he/she missed the deadline. A regular (weekly) basis can be the selected timeframe to give the opportunity for the system mainterners to update their results / predictions for the next week. What do you think?

I am sorry if I put a dampener on things. I think however that this discussion is quite refreshing and hopefully something useful (whatever its final form) will come out of it.

[lunatism]

Originally Posted by Rushian

• what if the yield achieved by a particular system is a result of a long series of short-odds bets, whereas another system results in the same yield but with longshots (and maybe with an equal number of bets)? Which system is the most stable?

if somebody provides the percentages for likelihood of h/d/a then i would presume that they aren't doing so with the sole intention of taking longshots. But if that were the case, we'd soon notice and suggest they don't take part in this, because we're looking at consistent week by week bla bla bla?! If a system is always suggesting short odds, then it's likely that unless very very good, they wont be successful in the long term....so what? Isn't the point of this to see what's good and what's not?

Originally Posted by Rushian

• What about the number of bets? A similar yield may be achieved with a few bets per year whereas another system may result in many picks each week? How do you compare these figures?

Comparisons should just be made on what the system users provide based on this season, and then we'll get a clearer picture of how they perform as the season progresses. I was always under the impression you wanted every selection from every week - and so number of games shouldn't come into it too much. Obviously if someone only has results from start of jan, then that's their choice to include from then, or we could say a certain date is the cutoff point, but shouldn't be too much hassle. We could even start from scratch and assign a 0% yield for each to start with - so a level playing field (to ensure nobody is cheating...which i'm sure would'nt happen anyway)

Originally Posted by Rushian

• What if a modeler uses a single bookie (with the overround that betting through a single bookie entails) and nevertheless achieves a 3% yield, whereas another punter has a 5% yield from a betting exchange? Which figure is higher?
• etc, etc...

all that has to be done is that we use the b365 from football-data - that is neutral to this and would immediately provide clarity i think. that is what i have used (since i wouldn't have had the odds on file anyway). And that's what i'd recommend everyone else do for the purpose of this experiment. Choice of bookie doesn't affect strike rate, but we do need to look at odds and yield, otherwise how else could you differentiate between short odd favourites and the long shots?!

Originally Posted by Rushian

So to keep it simple, and to start with, we could focus on the Premiership. A list of the Premiership fixtures can be set up in an Excel sheet which can be dstributed to anyone interested. People who agree to participate in this experiment could then send (by email or otherwise) their estimated probabilities in the common format to a "compiler" as lunatism put it, who in turn could summarise the results in tables in the thread. This could be done before the matches to ensure fairness, although we are all adults and I am sure one would not be so childish to change their figures if he/she missed the deadline. A regular (weekly) basis can be the selected timeframe to give the opportunity for the system mainterners to update their results / predictions for the next week. What do you think?

already made the spreadsheet, so people who pm me their email addy are welcome to receive it and fill in predictions for this weekend. if you have a yield (go on.....you know you do - put one in ), then put this in so that the fair percentages can be calculated and then i can post both the percentages and the fair percentages too. simple!

i dont mean for this to be a competition either - it's just that in a competition format it will be more enjoyable!

[Rushian]

lunatism,

I'm trying to understand and I may have got this entirely the wrong way, so please excuse me, but a few of the points you are making have me confused:

Originally Posted by lunatism

if somebody provides the percentages for likelihood of h/d/a then i would presume that they aren't doing so with the sole intention of taking longshots. But if that were the case, we'd soon notice and suggest they don't take part in this, because we're looking at consistent week by week bla bla bla?! If a system is always suggesting short odds, then it's likely that unless very very good, they wont be successful in the long term....so what? Isn't the point of this to see what's good and what's not?

The way I see it, the estimated probabilities have nothing to do with the bets placed unless you know the odds. For example, if according to a model the probability of a home win is 50%, then a bet will be placed only if the odds on offer are higher than 2.00. If the odds available are only 1.91, then a bet would not be placed (since your guess of the probability is 50% and hence in the long run, that bet would return a negative yield).

Furthermore, when you say that you wouldn't want someone whose system highlights the lonsghots, I am not sure that I follow: maybe someone's system produces average strike odds of around 3.00 with a relatively low strike rate (say 35%) whereas someone esle has a high strike rate but with much shorter odds. I don't understand why you suggest that the latter is desirbale ("consistent" as you put it) but the former is not.

Originally Posted by lunatism

Comparisons should just be made on what the system users provide based on this season, and then we'll get a clearer picture of how they perform as the season progresses. I was always under the impression you wanted every selection from every week - and so number of games shouldn't come into it too much. Obviously if someone only has results from start of jan, then that's their choice to include from then, or we could say a certain date is the cutoff point, but shouldn't be too much hassle. We could even start from scratch and assign a 0% yield for each to start with - so a level playing field (to ensure nobody is cheating...which i'm sure would'nt happen anyway)

When I mentioned the number of bets, I was suggesting that it would not be logical to compare two systems, one which produced a 10% yield based on 10 bets (i.e. the bets that were picked after comparing the estimated probability with the odds on offer, as described previously) and another which resulted in 3% yield based on 200 bets. Maybe both systems estimated probabilities for 700 matches in total, but the odds on offer provided value for 10 and 200 bets respectively (hence the fewer bets).

Originally Posted by lunatism

all that has to be done is that we use the b365 from football-data - that is neutral to this and would immediately provide clarity i think. that is what i have used (since i wouldn't have had the odds on file anyway). And that's what i'd recommend everyone else do for the purpose of this experiment. Choice of bookie doesn't affect strike rate, but we do need to look at odds and yield, otherwise how else could you differentiate between short odd favourites and the long shots?!

Again, using odds from one bookmaker may highlight a certain number of bets while using the best available odds accross the board would result in more bets, for the same set of estimated probabilities. This would mean that both strike rate and yield would be dependant on the choice of bookmaker. I agree though that, should people decide that a system's league is needed, the comparison could be based on the odds offered by a commonly agreed bookmaker.

Finally, something that was not mentioned in my previous post, the staking plan would also result in different yield, even if the estimated probabilities are the same, i.e. fixed level staking would lead to different results to something like the Kelly criterion etc...

Finally, as for my system, I haven't updated for some time now as it is still under development . At the last update of data, the overall yield was -3.4%, which was broken down as follows: Premiership -35.9% , Championship -3.1%, League One +4.0% and League Two +14.4%.

I don't have figures for the Liverpool - Everton match, but if I find some spare time over the weekend, I'll try to update the data and get some probabilities for the next set of matches.

Maybe I got confused, maybe not. Just ignore me if I am completely off the mark

Rushian

[lunatism]

TOMORROW is Saturday guys! Won't you be submitting any percentages? So far only mines and Grex's are in the spreadsheet, still waiting on Rushian and . HooLooVoo

Rushian...I appear to be talkig jibberish, and the reply box on my screen is too small to fit all of your response in at the same time

[Grex]

C'mon, Rushian and HooLooVoo...... Loon has put a lot of hard work into this, so the least we can do is to avoid delays to getting the ball rolling.

[ozangumus]

Everything depends on probabilities in soccer-betting. Any person betting on any team with any odd should think that this bet is a valuebet. Maybe he does not know but he is making an estimation in his mind and compares his calculation result with odd, and he decides betting to that situation or not.

Every outcome have a probability in soccer matches even Rapid Wien can win Salzburg with 0-7 at away So important thing is if odd * probability passes 100 or not. If you only bet valuebets you will probably win in long term there is not another way for winning in betting, you have to make an estimation by using inputs (past stats, special cases, injuries ext) and compare your result with odds announced by bookies.

Please examine my page to make easier your estimation job, probs there are estimated by only analsysing results of past matches, you can add effects of other special situations to make probabilities more optimal/reliable.

[Rushian]

Just a quick note: I'll try to find some time to update my system data, hopefully tonight or early tomorrow at the latest and send my estimated probabilities. (Unfortunately my work is such that most days, I can only post at night...)

[Grex]

Originally Posted by ozangumus

Please examine my page to make easier your estimation job, probs there are estimated by only analsysing results of past matches, you can add effects of other special situations to make probabilities more optimal/reliable.

Thanks, Ozangumus, for your input..... looks like we´re all wasting our time, lads

[Rushian]

Ok, I rushed through the system (thank god for Excel formulae!) and produced these probabilities for the weekend Premiership:

Div	Date	HomeTeam	AwayTeam	Prob(H)	Prob(D)	Prob(A)
Div0	29/3/2008	Birmingham	Man City	35.0%	32.6%	32.4%
Div0	29/3/2008	Bolton	Arsenal	23.3%	26.8%	49.9%
Div0	29/3/2008	Derby	Fulham	31.1%	26.6%	42.3%
Div0	29/3/2008	Man United	Aston Villa	74.9%	16.9%	8.2%
Div0	29/3/2008	Portsmouth	Wigan	41.6%	28.1%	30.2%
Div0	29/3/2008	Reading	Blackburn	40.1%	29.6%	30.3%
Div0	29/3/2008	Sunderland	West Ham	45.4%	26.8%	27.9%
Div0	30/3/2008	Chelsea	Middlesbrough	76.5%	17.4%	6.1%
Div0	30/3/2008	Liverpool	Everton	57.1%	27.2%	15.8%
Div0	30/3/2008	Tottenham	Newcastle	56.2%	24.7%	19.1%

(if they don't add up to 100% exactly, it's the rounding error...).