Quote:
|
Originally Posted by traynor
What kind of data source do you use? From your description, it seems you are downloading PPs and results separately. I assume you create/generate your own track variants on the fly using the above data sources? That was a key issue in my own apps, and it took a bit of work to get right.
When you back test models, do you set filters mainly for accuracy (win%) or (possible) return (ROI)? Most lean toward the ROI side, to their disadvantage. In most cases, ROI (in smaller samples of whatever size) is derived from what are essentially anomalies. Rarely repeat going forward, yet everyone seems obsessed with the "woulda coulda shoulda" type modeling.
You might try parsing for winner attributes (ignoring ROI) and test that on future races. ROIs in the 90s (in past races) with relatively high win rates can be especially productive. My conjecture is that bettors "seeking value wagers" tend to avoid the obvious best choices for win, and those obvious best choices for win may often generate a decent profit going forward (that may be concealed or missing in the sample used to build the model).
The big advantage is that the results of a frequency in the 45% and up is much more likely to be productive than a lower frequency is likely to reproduce a paper ROI from a sample of past races.
|
The PP's and results are 2 separate files. I don't use track variants or create them but I did write a program for someone who wanted to automate their own manually derived variant formula.
I have done thousands of tests over 12+ years. I look at both value and strike rate. In the beginning I was probably more focused on ROI. Any time I see an outlier I convert it to the win average of the rest of the group. The main thing I have realized is that static testing of the most common or obvious factors such as speed figures, distance, surface and class or any combination of them will generally result in a negative ROI in the long term. For example looking at a long list of past data and just filtering factor "A" with factor "B" then moving on to factor "A" with "C" and so on. Without developing some external formula or calculation that creates a new metric that is not in the raw data per se but is derived from the data (or not), I would say there is no hope of developing anything of any value (unless possibly you are extremely selective with great discipline or are very intuitive with visual cues or something like that). The novice and most handicapping software/services usually end up with a selection in the top 2 or 3 M/L or post time odds. Since this category is over bet, you end up with underlays consistently. I look at the more competitive races that are more confusing to the public and find an edge there. The general public will gravitate towards the easier or obvious choices without fail. So part of what I do is handicap the handicappers and use "outside" factors that can still be derived from the data through their theoretically (hopefully objective) implied meaning. So a big part of the battle is overcoming the majority consensus which dictates a high percentage of the finish order in the results. It is no secret that the post time odds are extremely efficient. When analyzing data, the trick is to distinguish the things that truly have a real effect from the ones that are merely illusions. The problem is that the illusions can be very convincing when looking at patterns. The hardwired human ability to detect patterns can be detrimental in this scope. I would agree about trying to increase the win rate and looking at attributes.
