Statistically evaluating a system and its significant variables

[Spankyf]

Hi guys and girls As of January past, I had decided to work on a few systems for draws. So far my selection policy for one system has been fruitful, albeit with an insufficient sample size (about 80 picks after 8 weeks with a strike rate of 37%). I've got it all nicely automated to calculate yield, strike rate etc after simply entering the scores in openoffice. I do want to improve the system and determine the most significant variables. So i'm collecting other data for each game too: home team's overall w/d/l records up to the match, their individual w/d/l at home and away, and the same records for the away side too. So closing in on halfway to my ideal sample size (i reckon 200 is a good number to separate the wheat from the chaff) and i'm thinking of ways i can test the many independent variables (home draws, away team draws away, recent form etc etc) versus the dependent variable - the outcome (if the result is a draw or not). I've read a lot on the subject since i started these systems but i am having a lot of trouble in working out the best ways to propose hypotheses and test those using multiple variables.* For example: lets say i want to test to see if the variables 'home teams draws at home' and 'home teams draws overall' have a significant effect on the outcome (draw or HW/AW). I don't know what test to use for this. I don't even know what value I should assign for a won bet (1 and 0 for loss?). I guess a 'LINEST' variance regression between the groups would determine if there was a degree of significance on the result, but i'm sure it does anyway (ie if the p value is close to zero, the null hypothesis - that neither the home draws @ home or the home draws overall in the season thus far - would be rejected. In otherwords, the % draws a team has will affect the outcome of x or HW/AW). *What i'm having trouble with is determining what tests to run to find out what variables have the biggest effect on whether the outcome but don't even know if it would be more effective to run individual tests rather than muddy up the correlation with unnecessary variables. I could run regression individually for each test, but would it be worthwhile? How should i do it? Any ideas? Please feel free to ask me more questions to clarify what i'm looking for as it may be apparent that i'm not exactly sure of what I should be looking for myself! I did do some stats in university and have a textbook with all the definitions and examples so I am ready to hear concrete suggestions such as 'Do a Pearson with p=0.3...' etc. Cheers

[Matthew]

Re: Statistically evaluating a system and its significant variables Place all your predictor variables in a logistic regression and using a fwd, bwd and stepwise selection method to assess which variables have the most significant bearing on the outcome. Or a simpler way Use Chaid (decision trees) to visualise how the proportion of draws shifts across the variables, I believe it uses the F test to do this test. It selects the most predictive variable with the split that drives the biggest change in the proportion of draws you observe in the sample. Or an even simpler way Take each variable and it's individual values (grouped in some way if they are many many different values in a given variable) and do a cross tab with H | D | A. Looking for variables that drive a large change in the proportion of draws from one extreme of the variable to another. Once you have the top 3 or 4 variables that drive the biggest difference between the proportion of draws, start layering them up together to find combinations of values in your top 3-4 variables that really drive the ability to predict draws. Then and only then, look at what sort of odds you need on the draw to find a profit. THE SINGLE MOST IMPORTANT THING TO REMEMBER - IT HAS TO MAKE SENSE. Don't put together day of week and attendance and just because a certain combination of the two see's profit at all odds on the draw think you have hit the jackpot. If it doesn't make sense then you are on a hiding to nothing. If you want to discuss more, PM me. An example of the crosstab I was talking about: [h=1]HOME TEAM: AVG SHOTS FOR PER GAME (ALL GAMES)[/h][TABLE=align: center]

[TR] [TH] [/TH] [TH=colspan: 3]RETURNS[/TH] [TH] [/TH] [TH=colspan: 3]STRIKE RATE[/TH] [TH=colspan: 3]YIELD[/TH] [/TR] [TR] [TH]HOME TEAM: AVG SHOTS FOR PER GAME (ALL GAMES) [/TH] [TH]HW[/TH] [TH]DR[/TH] [TH]AW[/TH] [TH]NO BETS[/TH] [TH]HW[/TH] [TH]DR[/TH] [TH]AW[/TH] [TH]HW[/TH] [TH]DR[/TH] [TH]AW[/TH] [/TR] [TR] [TD=align: center] 13.7[/TD] [TD=align: right]£95.98[/TD] [TD=align: right]-£585.21[/TD] [TD=align: right]-£623.78[/TD] [TD=align: right]4,677[/TD] [TD=align: right]58.37%[/TD] [TD=align: right]22.84%[/TD] [TD=align: right]18.79%[/TD] [TD=align: right]2.05%[/TD] [TD=align: right]-12.51%[/TD] [TD=align: right]-13.34%[/TD] [/TR] [TR] [TD=align: center]Not Played Yet[/TD] [TD=align: right]-£128.20[/TD] [TD=align: right]£11.92[/TD] [TD=align: right]£16.01[/TD] [TD=align: right]979[/TD] [TD=align: right]42.08%[/TD] [TD=align: right]29.01%[/TD] [TD=align: right]28.91%[/TD] [TD=align: right]-13.09%[/TD] [TD=align: right]1.22%[/TD] [TD=align: right]1.64%[/TD] [/TR] [/TABLE] If the Home Team has, on average, more than 13.7 shots in all games so far only 22.84% of games end in a draw. Conversely, if the Home Team has fewer than 8.7 shots, 27% of matches end in a draw. Not the most startling range but just an example for you to follow. Cheers Matt