The Cause of 99% of Betting System Failures

Using betting systems is probably the most popular form of finding selections. Also, amongst horse racing fans, the actual building of the systems is very popular.

The only problem is that most betting systems fail to make the profits that they’ve promised to deliver.

And the primary reason for this is… backfitting.

Well, that’s the name that most people use. If we were going to be completely technically perfect every betting system is backfitted by definition.

The reason for systems failing is overfitting.

So let’s start by determining exactly what overfitting is.

The Wikipedia definition is:

In statistics and machine learning, overfitting occurs when a statistical model describes random error or noise instead of the underlying relationship. Overfitting generally occurs when a model is excessively complex, such as having too many parameters relative to the number of observations.

I would say that’s a pretty good definition, but let’s change it in relation to betting:

In betting, overfitting occurs when a betting system only finds profitable selections based on the sample of data used to build the system. Overfitting generally occurs when the betting system has rules created which are specific to the data set being used and not generic enough when used on unseen data.

In normal English, we are overfitting when we have rules in our betting systems that only work on the data we are building the system on. And this most often occurs when we create rules that make no logical sense regarding the conditions of the races and the runners in them or by creating too many rules.

You see less rules means that you are less likely to be overfitting.

The problem is that less rules also means you are less likely to be able to make a profit.

In order to make a profit without overfitting we need to find the balance between rule creation and logic.

And there is one way that we can do this when we build betting systems without much difficulty.

It will increase the amount of time it takes to build the betting systems, but it also gives your systems a far greater chance of being successful when you start placing your money on the selections.

Which is what we want.

The process that I’m about to describe is a simplified version of bootstrapping.

This is a statistical process which ultimately refers to taking random samples of data throughout the development of a statistical model.

And that’s exactly what we want to do when we build betting systems.

Most betting system builders have been designed in such a way that you choose a date range and then you build your system on it.

And that’s the problem.

To be confident that your betting systems aren’t being backfitted, here’s what you want to do…

  1. Get all the data that meets your race criteria (i.e. all the data available to build your betting system on)
  2. Take 60% of this data to build your system on and keep the other 40% aside for testing
  3. From the 60% you have taken to build the system, take a random 60%
  4. Make your first system rule
  5. Put the 60% you took in step three back into all the data you took out in step two to build your system on
  6. Take a new random 60% of this data
  7. Make your second system rule
  8. Repeat steps five, six and seven until you have built all your system rules
  9. Test your system on the 40% of the data you set aside in step two for testing
  10. If you’re happy with the test from step nine, start paper trading your system before moving to live betting

This is the correct way to build a betting system that’s not going to be overfitted and has the best chance of continuing to make profits for a long time into the future.

But when you start to use this you’re going to notice something… it’s a lot harder to build a profitable betting system.

Rather than thinking this approach is so much harder to build a profitable betting system, you want to think about why it is so much harder.

Most bettors want to build betting systems quickly and get betting. Commercial system builders allow you to build systems that make potential profits very quickly and the figures look great.

Until you start betting them.

And then the bettor thinks that something has changed in racing which is why the system no longer works. After all, they can see it worked historically.

So they go back to the system builder and build another one quickly and the same thing happens.

Of course, occasionally there’s a good run of a few weeks before the system fails. But this is more due to luck than anything.

The priority of building a betting system and placing bets as quickly as possible is a mistake.

If you want to make long term profits from betting then where you must spend your time is in the research. Placing bets is quick and simple, anybody can do that.

Spending the required time on the research to build a profitable betting system, using the correct approach, takes time and patience. If it didn’t then everybody would be building profitable betting systems.

Understand that the best place to spend your time is on the building of your betting systems using the correct approach, even if it takes longer, and you will reap the rewards when you are ready to take your betting system live.

Michael Wilding

Michael started the Race Advisor in 2009 to help bettors become long-term profitable. After writing hundreds of articles I started to build software that contained my personal ratings. The Race Advisor has more factors for UK horse racing than any other site, and we pride ourselves on creating tools and strategies that are unique, and allow you to make a long-term profit without the need for tipsters. You can also check out my personal blog or my personal Instagram account.


  1. You’ve made a simple subject extraordinarily complex, therefore I could not agree less.
    When researching systems for, say a particular trainer over the past 5 years, you first of all have to eliminate all the oustanding negatives. This could be any number from 3-8+ initially. You then go on to see what you’re left with.
    An example could be David O’Meara with whom I profited greatly last year. His grey horses had 28 runs withonly 1 win. Ilogicall of course! But would you take the risk?
    Richard Fahey is rubbish at Beverly.
    Backfitting or even extreme backfitting is derided when under the heading of a “system”.
    Call it anything else … Trends, Stats, Form Reading and they assume a credibility that systems never will.
    Just take a look at all the content you receive when there’s a big festival meet and each race is broken down into as many or perhaps even more than 20 Trends, Stats. Now tell this isn’t backfitting gone mad.
    No Michael, I for one amongst many, will stick with systems.
    There’s many more observations I could add and will do so should you wish.

    All the best, Chris

    1. Thanks for the comment Chris, it’s great to hear your thoughts. For me, the difference between a system and trends/stats/form reading etc… is that the first should be completely systematic without any human intervention, every single person would get the same selections. Using trends/stats/form reading etc… will usually require human intervention to analyse the data and make assessments. In those situations then this extreme backfitting can certainly work to your benefit in some scenarios because we are filtering it manually.

  2. I believe you are involved with Mike Cruickshank. I enrolled for Bonus Bagging on 9mar 15, despite several emails to him and Clickbank for a refund I haven’t heard anything.
    Regards, Ken Pain.

  3. Perhaps we can view systems and backfitting as being part of the same process. Backfitting could be finding that top weight in handicaps make a profit and the subsequent system might therefore be to back only the top weight in handicaps.These are obviously not my findings or my suggestion.

    As I think Chris hints at, trend analysis and backfitting are very closely related and it’s the degree to which you adhere to the rules i.e. the how strictly you employ the system, that is important.

    For example, I use historical data to identify common characteristics of past winners (there will be different characteristics in different races); I then apply them to today’s runners until I have a short list or, in some circumstances, a very short list – of one. Would I be foolish to back this horse?

    If I told you that one such ‘qualifier’ was Chatez at Doncaster, another Naddir and that Gabrial was on a short list of two, would you still have the same view. It’s clearly better in better class races with greater depth of history but it also pointed to Warden Bond and Yourartisonfire yesterday.

    Rock solid system or just coincidental aligning of backfitting – who knows?

    1. Thank you for the comment Keith and the valuable insight into how you analyse races. I totally agree that this kind of hyper-backfitting when used in combination with human analysis can be very powerful. So much so in fact that I am building a tool not only to find the six major trends but also to find profiles of the most likely winners in a race using this information. However I have found in my experience that over fitting when it comes purely to system betting and where the user of the system will follow the rules exactly with no deviation based on their own knowledge when looking at the horses leads to losses the majority of the time.

  4. That’s what I call the pendulum quandry – where a very strict adherence to system rules results in losses encouraging the bettor to adopt a more flexible approach which then also results in losses.

    An interesting phenomenon is called ‘thin slicing’ (ref Malcolm Gladwell I think). This is the theory whereby after sufficient study over time, one has absorbed as much information as is necessary to make informed decisions and further study has little or no effect. We refer to it as intuition – perhaps rightly so.

    It can manifest itself as simply looking superficially at the race information with no study at all of the detail and simply ‘seeing’ the winner.

    Ultimately the simple fact that these myriad challenges exist only emphasises how difficult it is to make profit over time. That’s why your other article is so relevant

  5. Far too many punters consider statistics and trends to be twins, or, at least, treat them in a similar manner. While they may be more than distant cousins, they can’t be treated in the same way. It is true that both trends and statistics should point to all recent Derby winners as being 3yos, but what about favs at a mythical racecourse. The last 5 seasons have seen (in chronological order) 26, 28, 30, 32, and 34% winners. The 5 year statistics will point to an average of 30% while trends show a steady increase. Maybe one can be used to corroborate the other, but they are certainly not mix’n’match.

  6. Would the above be recommended for dealing with large data sets also? Or how would you go about splitting data in large data sets? I’m talking about splitting it for the purposes of regression,NN’s that type of thing.

    1. Splitting data sets for NN or regression is a bit different. What you get from these processes correlates very strongly to what you put in. You should separate data into specific race conditions, and then remove any correlated factors (or merge them into one). Building a good model, that has an edge, using this approach is very difficult and time consuming. There’s a reason it’s the way big teams work and most pro-punters don’t. Unless you are a statistician you’re going to find it very difficult to, and in the UK even harder because the conditions are so varied in comparison to other countries.

  7. Hi Michael, Must agree with Chris here. Am a major believer in the KIS school of thought. Sys1 started Feb 2011.Trainer based, certain minimum Win and Place %, never bet under 4/1 bookies SP and MUST be able to handle the going. Average PPD to level Stakes as of end May 2017 are running at 7. Sys2, into 3rd year, for Flat and AW and Sys 3, into 2nd year, for NH both OR based and are a very simplified version of your own 5 Step system taken from the first chapter of your book and as of end May 2017 are both averaging 8 PPD to level stakes. None of these systems has ever had a losing month and all 3 average between 5-8 Points Per Day profit to level stakes although I personally use a 2% staking system. I have just started Sys 4 based on your Daily Trainer/Jockey stats although i must admit i don’t even look at the jockey stats and I start with a far higher % than 17, which I find is way way to low. I the apply my own version of your Ultimate Horse Racing Conditions System and starting on June 5th 2017 as of yesterday using a 100pt bank and 2% win only, my other 3 systems being Win and Place, I have as of yesterday amassed 1460pts in only 55 days using same same minimum 4/1 SP as my other systems; At 75 must finally have it cracked I hope. Hope this comes out ok as my keyboard has decided that it has a mind of its own and goes haywire. Cheers Stuart

Back to top button