Michael Clarke's 4-Point Strategy
I've never really looked at course-specific models Michael. I have a very basic grasp of Bayes but I've never found it great for predicting prices. Our old friend co-variance scuppers it I think.
The paragraphs below I posted on Michael S's thread but they're better suited to this thread as they deal with combining predictive factors.
Some statistical software that does logistical regression will find the coefficients for each factor and by using several iterations will quickly eliminate those that bring nothing to the party in terms of predicting winners (strike rate). You're left with the most significant combination of factors for SR but you still need to measure the combined A/E.
Models I've seen take these coefficients apply them to fresh sample data which hopefully gives similar SR results to the original sample. Then the A/E calculation formulas are set up and an algorithm like Excel Solver is used to maximise the A/E by tweaking the coefficients whilst keeping the SR within certain bounds. It's not guaranteed to find a solution every time but if it doesn't you try some other factors (or composite factors) until something robust is found.
So the process is,
- Find the most predictive factors and their coefficients (weights)
- Apply these to fresh data and verify predictive power
- Calculate A/E of both datasets
- Apply back-solving algorithm to coefficients to maximise A/E by tweaking coefficients whilst keeping SR within bounds.
- Rinse and repeat
I could never get to grips with Excel Solver. Co-variance is always the big pain. As you say, with model building you want to spend the most time on correlation tests and then on keeping the factors your including in the model to a minimum by only focusing on those which prove to be most effective.
Solver should work for your power rating method Michael.
Once you have guestimated your factor weightings and calculated your power rating, If you then set up a spreadsheet applying the PR to your sample and calculating the top rateds, making these the bets, calculating through to A/E and SR. You should be able to tell solver to maximise A/E by tweaking the weights whilst staying within certain strike rates.
Tweaking the weights changes the power rating which changes the top-rateds. This changes the bets which changes the expected winners and actual winners which changes SR & A/E.
I'm not saying A/E will be greater than 1 but it should maximise it for your chosen strike rate range.
I have now finished my analysis on the longer distance turf handicaps and unfortunately I have arrived at the same conclusions. Those ratings with an AE above 1 in the 5 year base period produced collectively an AE of less than 0.99 in the result period.
I will not proceed with this approach.
I will look at the foundation strategies which look promising. My initial concern is that there is an element of back-fitting used in these strategies.