In my last article where I’ve interviewed Erik from SoccerBetBlog on how he gets to figure out Estonian football games I promised to give you an article outlining a simple method to evaluate the performance of tipsters, based only on a few simple to obtain metrics. It is all about a statistical method called **the t-test**, but even if math is not your thing don’t be scared to read further. The method is quite straightforward and to grasp it you won’t need more than a basic understanding for numbers and some common sense.

## Tipster Performance Evaluation

What constitutes a good tipster performance? Obviously, one is always aiming at the highest yield possible. However, the yield does not tell the whole story. You can check out the few tricks bad tipsters are using to boost their yield without having much quality in their tips. Mostly it revolves around a low number of bets, high odds and large stakes. Those increase the likelihood that the yield of a tipster is more the result of chance alone than anything else. You will find out all these factors are accounted for in the t-test. So while my article on deceiving tipster tricks was giving some examples from practice, in this article you can see the basic theory behind those.

Our task in evaluating a tipster’s performance is to determine how likely it is that the yield the tipster has achieved is the result of chance, as opposed to the tipster’s skill or anything else. If this likelihood is low enough we might conclude that the tipster is a good one.

## The T-test

We can calculate this likelihood via the t-test. It is assumed that the betting returns follow a t-distribution and that the tipster is using level stakes (I will come back to this in a minute). Then we can compare the achieved return of a tipster with the theoretical return from a bunch of random bets. For this we need to calculate a t-score. A higher t-score translates to a higher likelihood of the yield not being the result of chance. In other words, the higher the t-score, the better. The following article from Pinnacle gives us a neat formula for calculating this t-score:

With t = t-score, n=number of bets, r=betting return, o=average odds;

Again, the higher the t-score is the better. The factors that influence the t-score are the ones that are relevant for evaluating the performance of a tipster. A higher return (r), a higher number of bets (n) and lower average odds (o) all lead to a higher t-score. You can input those three, calculate the t-score and convert it to a probability (p-value) using the Excel TDIST formula or one of the many tools and tables freely available on the Internet.

A higher t-score leads to a lower p-value. The p-value is the likelihood that the yield of the tipster is the result of chance. It is all relative from here on, but in general: you can consider a p-score below 5% as giving moderate evidence against the thesis that the yield is a result of chance alone. A p-score below 1% could be considered a strong evidence against this thesis, and under 0.1% a very strong one.

### Examples:

- Using the formula above I will quickly go through the record of Erik from SoccerBetBlog, which can be obtained from his account on TipsterTube. He has a yield of 18.24% (r=1.1824) with average odds of 1.9 (o=1.9) from 69 bets (n=69). Inserting these figures in the formula gives us a t-score of 1.6448 and a p-value of around 5.2%. So at this point it is somewhat likely that the yield Erik has achieved is a result of his skill and not luck alone. Given his professional approach to tipping and a history of unrecorded profits I personally believe this p-score will only go down further in the future. But I suggest you don’t take my word for granted and do your own research
- I am following an American Football strategy from the Green-All-Over blog, which is about backing away underdogs on the spread. It has performed pretty well this season with a yield of 5.83% with average odds of 1.954 from 177 bets. Placing these inputs in the formula gives me a t-score of 0.7967 and a p-value of 21.33%. So according to the t-test the chance that there is some merit to the system is lower here than in the first example. However, it is still high enough for me to continue to follow it and give it a chance. If anyone is interested, I might report on those numbers later in the season again

### Additional factors

You can repeat the exercise for any system or tipster that you follow or consider following. You can also compare as many of them as you like. Just keep in mind (and that is very important) that the formula above relies on the assumption of level stakes. However, most tipsters use variable stakes. Therefore, the obtained results by this simple formula will only be an approximation of the real results. This must be fine as long as the tipster does not vary stakes too much and achieves similar yields from similar stakes. However, beware of tipsters that have won huge from a small number of large-stake bets as demonstrated in the Cheap Tipster Tricks article. This is **not **accounted for by the formula and is something you will need to check additionally by yourself.

That’s it for the t-test. I hope you will find it useful in evaluating systems and tipsters. If anyone would criticize the piece above as not scientific enough I would totally agree. I am aiming this article at bettors with little to no background in statistics, which I believe to be the case for the majority of people. If you want to engage in a more in-depth discussion on the matter or have any questions feel free to drop a comment below.

Thanks for reading, I wish you all a Happy New Year and a great 2017!

Hi Nemko,

great piece of work! I was wondering if there is some way to tackle the same tipster performance evaluation by dropping the level stakes assumption. Should we look for a different distribution? What about some heuristic corrections instead? What’s more, I still cannot convince myself of, let’s say, the nature of the denominator [Sqrt((r*(o-r))/n)] of the t-score, which should represent a kind of standard deviation term. Am I missing something?

Hi Lorenzo,

if you drop the level stakes assumption I don’t think you can come up with a one-size-fits-all formula. Intuitively, if you have many small bets and one huge one, this one bet will add up the most of the variance (especially if the odds are high) and its outcome will be detrimental for your total yield. So average odds won’t tell you much in this case.

So if you want to be precise, I think you would need to add up the variances of all bets, figure out the variance of the whole portfolio and implement it in a t-test.

The denominator is the sample standard deviation. It is a simplification of the general formula used to avoid adding up all the variances and just be able to work with a simple formula. Sqrt(r*(o-r)) is the standard deviation of a bet with odds o and return r: http://i65.tinypic.com/264nr0o.jpg

I hope that helps.