In today’s world we are increasingly aware that data can be powerful in informing decisions; very often, large sets of data from past events are used to help predict the likelihood of things happening in the future, given a certain set of circumstances.
Unsurprisingly, this concept applies to horse racing and punting, and plenty of people use trends (essentially data from past renewals of given races) to help them make decisions on which horses are likely to win these races in the future.
I have seen a huge variety of trends metrics used to narrow down potential main players in races, including: mean (average) winning rating; mean weight carried; mean runs per season prior to race; mean SP; proportion of wins over a given number of renewals of different ages, sexes, and with previous course or distance-winning form. Sometimes, trends summaries combine lots of these metrics to produce an ‘ideal template’ for what you should look for in a selection (e.g. a horse rated 140-145, weighted between 10-12 and 11-03, run 3 times this season, SP of 6-1 (7.0) – 10-1 (11.0), aged eight, won at the track before…).
While trends analyses can (sometimes) be informative, there are a number of potential pitfalls. The three main ones are listed below, along with tips for lessening their impact and resources that will help you perform your own, statistically rigorous trends analyses.
Pitfall 1: Using a lack of data
Any statistician will tell you that the more data you have to work with, the more confidence you can have in an apparent trend actually having an influence on results. But frequently the trends you see or read in racing conversations are woefully scant in terms of data.
Given that there are so many variables at work in racing, any trends that consider individual variables (e.g. the winner’s official rating band, age etc.) should incorporate at least 50 data points (and preferably more than that).
Tip 1: Instead of limiting data to ‘winners’ only, incorporate ‘placed’ horses too
Although this may not be appropriate for non-handicap races (due to the smaller fields that are typical), it ought to help you accrue a larger data sample for big handicaps.
There is so little margin between winning and coming second, third or fourth in a big handicap with lots of runners (e.g. the Coral Cup at Cheltenham), that it is easy to argue that any horse filling the places could be treated as succeeding for your trends analysis. For example, if ‘weight carried’ really is a significant factor in affecting the likelihood of success, you would expect horses with similar weights to the winner to also perform well in behind him/her. And, if they don’t, the odds are that weight carried isn’t all that important.
Pitfall 2: Ignoring race evolution
It is easy enough to gather large data samples for most big races if you go back far enough, but for that data to be valuable it must come from races that are similar in make-up; after all, the whole point of a trends analysis is to see if there is anything common about the typical winners (and losers) that can help you predict likely future winners (and losers). Of course, this can only be possible if these races are similar year to year (in terms of the number of runners, the abilities of the competitors, and the course conditions).
Racing evolves, though. The make-up of a race in 2016 might be considerably different to the 1996 — or even the 2006 — renewal of that same race. In these instances, including data from older renewals will make any emergent trends irrelevant.
Such evolutionary changes are easy to visualise in some of the Cheltenham Festival handicaps, where the official rating required to get into these races has risen considerably in the last 10-15 years. For example, in 2000, the lowest-rated horse in the handicap in the Grand Annual Handicap Chase ran off an official handicap mark of 125; in 2006 it was 126; but in 2016, the lowest-rated ran off 137. That’s a big jump in a short space of time, which is typical of Festival handicaps that are increasing in quality according to the ratings each year; clearly it would be crazy to include all that data in the same pot when calculating the mean handicap mark of the winner. Yet each year, you will see trends analyses suggesting the ideal handicap mark for a potential winner of the Grand Annual that incorporate some of these out-dated data.
Tip 2: Check races for similarity before including data in any analysis
This is so simple to do; all you need is separate your data by year/renewal and then see whether there is a consistent upward or downward pattern. You can use the below spreadsheet to see if there is a correlation between time (year/renewal) and the factor you are interested in (e.g. official handicap rating, age etc.).
< Correlation spreadsheet >
Simply input the data into the two blue columns and the calculations will take care of themselves. The correlation result will be displayed between -1 and 1, with numbers closer to 0 implying no correlation, and numbers closer to either -1 or 1 implying a stronger correlation. Also note the p-value / significant number. Only if this is less than 0.05 is any correlation deemed to be statistically significant.
If there is a significant correlation, then your data are showing evolutionary change because the factor you are interested in is affected by year/renewal. In this instance, these data should not be used in trends analyses.
Note that you can work backwards from such a result, removing the oldest years/renewals and their data, and then recalculating to see if the significant correlation disappears. If it does, you can use the data from the more recent years/renewals.
Pitfall 3: Not assessing whether a trend is statistically significant
A trend is useless in helping you narrow down potential winners if there is no statistical confirmation that it is unusual in a mathematical sense. For it to be meaningful it needs to show that something has happened far more or less often than expected in past renewals. And there is more in that phrasing than first meets the eye.
As an example, if you look at the past 10 renewals of a race and find that 9 winners were aged three and only one was aged four, you might assume that four-year-olds had performed really badly and that they are therefore likely at a major disadvantage once again. But, to be sure they actually did underperform relative to what you would expect based on chance alone, you need to consider how many three- and four-year-olds ran in these renewals and what kind of chances they had.
If around 90% of the total runners were aged three and their SPs were similar to the 10% aged four, you would expect around 9/10 races to have been won by the three-year-old group (assuming of course that the difference in age is important in influencing success).
The SPs of these horses are really important. Imagine that there were 10 horses in one of these races and five were aged three, and five were aged four. Clearly, if the three-year-olds took out a huge proportion of the market (say they included the favourite and three of the next four in the betting), the winner would be much more likely to come from that group than from the four-year-old group.
Assessing statistical significance
I use the term ‘statistical significance’ fairly lightly here, as the below guide and resources have been greatly simplified from a mathematical point of view; they will, however, provide a fairly accurate means for assessing how significant an apparent trend is, and they will do this without requiring any statistical know-how or complex statistical software packages to compute.
Use the below spreadsheet to perform your analyses very painlessly. The blue sections indicate regions that you must input data in, whereas the pink sections will generate numbers automatically based on your data. You will see I have included enough columns to take data from 10 renewals, and for up to six groups (A to F). If you want to assess trends for things such as weight carried, or official handicap rating, just group horses into six bands (e.g. official handicap rating 120 – 124, 125 – 129, 130 –134, 135 –139, 140 – 144, 145 – 149).
< Using trends spreadsheet >
Below is a three-step guide that provides some insight into what is going on in the spreadsheet, and why it needs to be going on to provide you a reliable measure of whether a trend is meaningful or not.
Step 1: Access your past data of interest
I suggest you pull up the results from the most recent 5 – 10 renewals of the race you are interested in generating trends for. Ideally, you are aiming for a minimum of 50 data points (a minimum of 50 total runners across the renewals you are assessing).
The more data the better, but the more recent the renewals the better, as well; if you are able to get 50 or more data points from the seven or eight most recent renewals, that may be preferable than getting lots more data but including the last 20 renewals (see the pitfall of race evolution, above).
Step 2: In each renewal, group all the runners together that fell into the discrete category you are assessing (e.g. age, or weight band etc.). Then calculate each group’s chance of success in that renewal based on their SPs
To calculate the different chances of success for each group, add the individual chances of success for all horses that fall into each group. For example, in our three-year-old versus four-year-old example mentioned above, suppose that in one renewal there were only two horses representing the three-year-old group.
To calculate that group’s chance of success, you would add Horse A and Horse B’s chances together (e.g. Horse A: 4-1 or 5.0 = 20%, Horse B: 7-1 or 8.0 = 12.5%, so the three-year-old Group chance of success = 32.5%, or about 2-1 or 3.0).
Because you should include places as ‘successes’ in the same way as ‘wins’, you must then multiply each group’s % chance of success by the number of places available in the race. Suppose there were three places available; you should multiply 32.5% by 3 to get 97.5% for the three-year-old group.
Note that the race totals for all groups will exceed 100% in the win market (sometimes by a long way in handicaps with lots of runners), so you need to approximate their real chance of success as if all groups added to 100%. This is simple enough; you just need to divide 100 by the total win % for all groups and then note that figure down before using it to multiply each individual group’s % chance of success. Suppose the total win % for all groups was 124. We would divide 100 by 124 to get 0.806. We would then multiply 97.5 by 0.806 to get 78.6% as the final % chance of success for the three-year-old group in this renewal.
Step 3: Add each group’s % chance of success across all the renewals and compare the frequency of expected ‘successes’ with real-life ‘successes’
If any group performs much better or much worse than expected across all races, there is a chance that you have found a meaningful trend.
In the spreasheet resource I have set the statistical significance at a confidence limit of 95%, which is standard practice in most scientific research studies. This is important – a statistically significant result does not prove that a trend is meaningful; it just provides strong support that it is.