Predicting Speed Rating from Past Ratings
 

Predicting Speed Rating from Past Ratings

The original thread asked the question about how to assign weights to a horse's last four speed ratings in order to predict the speed rating for an upcoming race.

(Please stay on topic. If you are going to make negative comments about this topic being discussed again, please do not make them here. Make them in this thread http://www.paceadvantage.com/forum/s...76#post1740576)

So here are the key points of the first post of the closed thread:

I don't know if I answered this in the original thread, but I would find a set of races for which you have 4 past speed ratings for each horse and then do a standard regression. So you'd have four columns of independent variables:

SR1Back, SR2back, SR3back and SR4back.

You would have one column with the final speed rating for "today's" race. Call it something like SRTodayRace.

Next, you will perform standard regression and will get coefficients for each past speed rating that you can use to plug into a regression equation that will give you the speed rating of an upcoming race for which you have 4 past speed ratings.

Your data file would look something like this:

SR1Back, SR2Back, SR3Back, SR4Back, SRTodayRace
84,83,80,59,88
79,84,29,88,80
etc.

The topic of this thread is a good subject to try and get into.

Your suggestion of doing a regression analysis is a very logical one.
That is what I too would consider doing, but certainly I'd put in controls for surface and distance.

But telling people not to make negative comments in any thread seems narrow sighted. There may be others out there who do not think that regression analysis is the way to go and are willing to provide reasons why.

I feel like there would need to be some recognition that "third back" for instance is an ordinal consideration. One horse's third-back might be 29 days ago, and another's 220 days ago.

Good points. Surface, class, distance, track, etc. You can focus the model to be as specific as you like. The risk is having too many models. However, if you just play a couple of tracks then it probably makes sense to narrow your focus. In fact, I've always thought the way to beat the big syndicates is to be as focused as possible and take advantage of the errors in their models.

As far as negative comments, I meant the kind of comments that got the previous thread closed -- name calling and that sort of thing.

It should be possible to make a comment that is in disagreement and still be civil. :)

Be happy to hear any math, but will almost guarantee the outcome will be close to this when distance and surface are considered:

Last race, most likely.
Two back, second most likely.
Three back, third most likely.
Etc.

That said, it goes way deeper than that to win in this game. Things like trips, form cycles, trainer changes, horse age, figure patterns and so many other factors need to be weighed. I'm willing to hear others thoughts, but results will be very disappointing IMO just trying to predict speed ratings from nothing but past speed ratings. In a way, that is what everyone is trying to do every race, figure out, based on PPs, which horses will run fastest today.

Agree. Past speed figures probably don't give you enough information about how the horse is going to earn its speed figure today. However, the goal is to get enough winners by just betting the horse with the highest predicted speed figure. So maybe that is possible. I'm skeptical. Perhaps if you limited this method to only certain types of races, say, older male claimers, then it might work?

I hate to add to the confusion here, but I will.
Predicting speed figures can be fun.
But has anyone noticed this site is called Paceadvantage.com?

Let me just say that a horse's Speed figure will be, in part, a function of it's Pace figure, in addition to surface and distance.
To look at the Speed figure alone......well good luck.

It won't help much how selective your filtering strategy is if you intend to black box the fastest determined projected speed figure alone. Horse racing is too complex and dynamic for that. Study running styles and race/pace shape and watch replays.

To quote Brisnet:
"If you bet the highest speed figure horse to
win in every race, you will probably achieve
30 to 35 percent winners and a flat bet loss.
The loss means that the public is very
much in tune with speed handicapping which, of course, drives mutuel payoffs down. The percentage means that roughly two thirds
of the races are won by horses that are
not the top speed figure horses. Some of this
can be explained by the top speed horses
not being suited to the distance or surface,
going off form or being out classed.

By far
the main reason for this phenomenon, however, is that the top speed figure horse is not suited to the pace match-up of the race. It has been said: 'It is not how fast the
horse runs that is important, but how the horse runs fast.'"

Source:
http://www.brisnet.com/library/softw...e%20Shapes.pdf

I didn't follow TrifectaMike's original thread close enough, so I don't know if his point was that a profit could be made by predicting upcoming figures by assessing the horses' 4 most recent speed ratings; but if THAT was the point that TM was making...then I would bet him money that it couldn't be done. These speed ratings are dependent upon so many different things, that alone they tell us very little. Speed ratings are useful in understanding a race when looking at it in retrospect...but only if they are combined with ratings of a different sort. Alone, they cannot be effectively used in making future figure projections. Two speed ratings may LOOK identical...but they are seldom really the same.

I believe the thesis of this argument is not about the highest speedfigures/rating, but about deriving the optimal speedfigure/rating from the PPs for the horse by assigning weights given its last 4 figures

Pace is inherently the independent variable wrt speed of any moving object.

Therefore final time becomes the dependent variable of that moving object.

I know that, I saw the original thread also but i was addressing the specific concept in Highnote's third post.

Incidentally, I would say that in my own research, only the last 2 speed figures are of any predictive value.

Define optimal.

Suggestion: try weighting the speed ratings by number of days since the speed rating was earned.
Ie: if the last three were all earned within the last 45 days, all get equal weighting. If two were within 45 days they are equal and the third is in 45-90 so it gets a lesser weighting.

I may also suggest that this weighting is not linear ( ie: speed figure from 60 days ago is half as important as speed figure 30 days ago ) but would be some curve.

That is easy; optimal would be the "best" and to extrapolate, it would give the BEST rating with respect to the conditions the rating was earned.