Simon Rowlands / Simon Rowlands / 06 July 2010 / 8

BB's top racing bblogger gets technical again in the first part of a three week series - prepare for Fibonacci Sequences, Zipf's Laws and more

If you believe much of what you read and some of what you hear, handicapping is primarily a matter of guessing which horse in a race "ran to form" and basing an entire assessment on that premise.

Not only is such yardstick handicapping fundamentally flawed, due to what scientists call "extrapolation error", but it does not answer the question of what should be done when little or nothing is known about the form of the protagonists. You cannot pick a yardstick - even an imaginary one - when no yardstick exists.

Time analysis and some form of breeding analysis can both help to solve the problem, but a far more powerful and more widely applicable tool is race standardisation.

Race standardisation is as old as the Hills (think Barry, rather than Richard and Michael), and yet it seldom gets a mention in post-race analysis in the media. The reason it continues to go strongly all these decades on is because its core principles have been tested time and again and found to be an accurate measurement of past events and a reasonable indicator of future ones.

It can be shown that certain races and race types repeat the same sort of level of form, year in, year out, once adjustments have been made for field sizes, margins between horses, race times and so on. It is also fairly easy to discover which races do not conform to this principle as closely as is the norm and which therefore require a different approach.

When dealing with two-year-old racing, some form of standardisation is almost unavoidable. For instance, when assessing the Brocklesby Stakes, for unraced juveniles, it is highly desirable to have at least some idea of the ability the winner would usually be showing in winning.

It is also important, if somewhat less so, to know what rating the second would usually post. In turn, it is also important, if somewhat less so than with the second and the winner, to know what rating the third would usually post. And so on.

It is best to look at past races individually, so that the degree of variation becomes apparent, rather than simply to take averages or pars. However, in the absence of race-specific standards it makes sense to fall back on general standards, modelled from wider information (e.g. from the level of form usually achieved in maidens at a given course at a given time of year).

You then apply the figures to the race under consideration - the one just run - and weight the outcomes appropriately. What "appropriately" amounts to, is a matter of debate and may alter according to circumstances. My apologies for going all Da Vinci Code on you for a while here!

My experience is that weighting should usually be something close to a Fibonacci sequence (in which the significance of the winner is the same as the significance of the second and third combined, and the significance of the second is equal to the third and fourth combined, and so on) or Zipf's Law (in which the significance of each of the principals is weighted inversely according to rank: 1/1, 1/2, 1/3 and so on).

Here is an example of how standardisation works in practice. For the purposes of this illustration, I have used Zipf's Law for weighting and the evidence from only three years (it has also been assumed for convenience that no adjustments need to be made for field sizes etc).

2009: winner's rating, 100; second's, 98; third's, 95; fourth's, 90; fifth's, 76.
2008: 95, 94, 93, 88, 87.
2007: 102, 90, 80, 75, 73.

This year's race: winner 5 lb > second; winner 10 lb > third; winner 15 lb > fourth; winner 20 lb > fifth.

Judged on 2009, the winner this year would be rated 100, the second would be rated 98 (which would make this year's winner 103), the third would be rated 95 (making this year's winner 105), the fourth would be rated 90 (making this year's winner 105) and the fifth would be rated 76 (making this year's winner 96).

If you weight these figures according to Zipf's Law, then you get: ((100*1.00)+(103*0.50)+(105*0.33)+(105*0.25)+(96*0.20))/(1.00+0.50+0.33+0.25+0.20), which is 101.7 for the winner.

The figures for 2008 and 2007 by the same process would be 99.1 and 96.8.

What the number-crunching tells you is that a result like this, in a race like this, would throw up a winner's rating of between 101.7 and 96.8 (average of 99 to the nearest integer) based on the previous three years.

In the absence of other significant information - such as a notably good or poor time, or prior form of the protagonists - this is a good starting point for assessing the race.

The validity of this assessment is then tested against subsequent events and revised where necessary. This is a crucial part of the process in order that standardisation does not become self-perpetuating and divorced from reality. The evidence over many years is that standardisation has remarkable utility if proper care is taken.

Next week I will describe how standardised assessments can be made in handicaps. That is, in races in which we know much more about the horses involved.

'.$sign_up['title'].'

'; } } ?>

(8)

1. Zee Zoo | 07 July 2010

   Bloody hell, never realised there was so much to handicapping! Good stuff though, claps.

2. Stu | 08 July 2010

   Fascinating stuff. Normally find the subject of standardisation more than a little baffling but with a little effort this has helped make things clearer. Good work

3. Wayne C | 23 March 2011

   I see one problem with this.

   Suppose in a field of first time starters and other lightly raced horses the winner wins by 10 lengths and there is another large gap to the 3rd horse.

   There is no way to tell if the winner was great or the horses behind him were poor.

4. Simon Rowlands | 23 March 2011

   There is no way to tell - for sure - if the winner was great or the horses behind him were poor by any means of handicapping of which I know (other than, possibly, if the performance was backed up by a superlative timefigure).

   We are always working with incomplete information, but we have to estimate achievement or sit out an awful lot of races.

   However, we do know that 10-length winners are quite rare on the Flat and that 10-length winners tend to be better than, say, 2-length winners of the same race.

   How much better can be established from retrospective analysis of a large sample of races. This shows that 10-length winners are not as much as 8 lengths better than 2-length winners of the equivalent races, on average. 10-length winners tend to show up the limitations of their rivals as well as their own superiority.

   This is all factored into race standardisation. If an average winner of a race "should" be rated 80 and the second 75 then a 10-length winner might be rated, by race standardisation, 92 and the second 67. It would be unlikely to rate the winner 100 and the second still 75.

   Race standardisation tends towards the "third way" in your scenario, namely that the winner is a little bit great and the horses it beat were a little bit poor. Experience has shown that this is a sensible approach to take.

   Simon

5. Simon Rowlands | 23 March 2011

   winners of the same race.

   How much better can be established from retrospective analysis of a
   large sample of races. This shows that 10-length winners are not as
   much as 8 lengths better than 2-length winners of the equivalent races,
   on average. 10-length winners tend to show up the limitations of their
   rivals as well as their own superiority.

   This is all factored into race standardisation. If an average winner of
   a race "should" be rated 80 and the second 75 then a 10-length winner
   might be rated, by race standardisation, 92 and the second 67. It would
   be unlikely to rate the winner 100 and the second still 75.

   Race standardisation tends towards the "third way" in your scenario, in other words, namely that the winner is a little bit great and the horses it beat
   were a little bit poor. Experience has shown that this is a sensible
   approach to take.

   Simon

6. Wayne C | 30 March 2011

   Simon,

   Thank you for the explanation.

   I am a US based handicapper. I have been making similar ratings for over 30 years for personal use.

   I tend to incorporate pace and final time ratings to help make the tougher decisions about quality like in the example we've been discussing, but figures even partially based on final time tend to drag you towards betting favorites here in the US without necessarily providing a more accurate rating.

   Are there any books or probability studies available on this subject that you can recommend?

   I'd be happy to purchase anything worth while. I am always looking for new insights to improve my own process.

   You can contact me here

   classhandicapper@nyc.rr.com

   Wayne

   
   

7. Simon Rowlands | 30 March 2011

   Wayne,

   Many thanks for your continued interest. I see the problem as essentially two-fold.

   In the first instance we try to assess the abilities and aptitudes of horses under the prevailing conditions as precisely as we can. In this respect, form (class) and time analysis is crucial. The more accurately you can perform such analysis, the better.

   It is then necessary to convert this analysis into percentage chances, and then to execute bets accordingly.

   The two tasks are separate. The former is, indeed, likely to end with obvious selections if all you do is take the top- or near-top rated horses by form and time analysis. The purpose of the latter is to identify value bets, which can occur at all types of odds, and which may ignore the obvious from the very same initial information.

   It seems to me that you have a lot of experience of the former. If you have not already read my remarks on the latter then I direct you to my blog "Converting ratings to odds" ( https://betting.betfair.com/horse-racing/bloggers/simon-rowlands/post-222-040810.html )

   I have since refined the process slightly, with different weightings and power functions, but it is a guide which can be adopted to your own purposes.

   Unfortunately, I do not know of any books or probability studies on this subject. Worthwhile literature on horseracing analysis is particularly sparse in the UK.

   Incidentally, since the initial article on race standardisation was published, I have applied these and other methods to racing in US and Canada, on behalf of Timeform, with very encouraging results. I think there is a great deal of potential there.

   Simon

8. Wayne C | 31 March 2011

   Simon,

   Thank you again.

   I have made an effort to convert my ratings to percentage chances in the past and read your blog entries on the subject.

   I find the process fraught with peril because I lack detailed enough probability studies on my own ratings. I just know they work well for isolating probable winners. :-)

   What I have been doing instead is ranking the horses in terms of probability of winning and then looking for discrepancies between my rank and rank by the actual odds.

   When I believe a horse is misranked based on the odds, I take a closer look at the actual gap in ability between various horses to refine my decision making process.

   I wish I could automate the process a little. I'll read you entry again.

   Thanks

   Wayne

:

:

:

?

: