Lies, damned lies, and NRL statistics

In 2014 only 62% of pre-game favourites actually went on to win. In 2015 this fell to just 61%.

This is far lower than other sports (AFL is over 70% for instance) and tells us two things:

  1. NRL is a tough game to predict the result

  2. There’s some tasty odds if you can pick the right underdogs

So could you use statistics to predict results? Or more importantly pick results better than the bookies? If you could, what statistics would be useful? What factors actually impact the result?

There are obvious things, or at least factors you might think are obvious, such as whether the team is doing well (ladder position), have they won their last few games (form), are they playing at home (ground advantage), have they just signed Roger Tuivasa-Sheck (and hope his skills are instantly transferable)!!

Could the same factors that influence the outcome in other team sports be transferable to NRL predictions?

These are some of the questions we set out to establish, and like many punters before us, have turned to statistics to help us see into the future.

For example, the raw stats for 2015 show the following win percentages when each factor is considered in isolation:

Bizarrely where there was a difference in ladder position last year it had only a marginally positive impact on the result, and form wasn’t much better. Player experience, measured by NRL appearances was the biggest determinant of the result, followed by player ratings.

So what is the TipBetPro approach?

Based on the above, and our experience from other sports, we wanted to create a “player based methodology” which essentially provides less weighting to more traditional “team based” tipping features (eg form and ladder position) and a higher weighting to the 17 players that actually run on to the ground.

So unlike other punters, we put a heavier focus on the main factor which determines the result…….who is named in the 17 that week! Our experience across other sports has shown this is consistently underestimated by most “models” and that includes those used by the bookies.

It sounds obvious, and in commentary on every forum, you’ll see supporters predicting outcomes and linking it to the latest injuries, or who’s returned from suspension, or how the halves smashed the same opponent’s last time out. And to a certain degree others do factor these in, but how much, what is the impact, typically it’s not enough.

To figure what is enough (or try to) we took statistics from every game over a fiveyear period and determined which features have the most correlation on winning games of rugby. By fitting the data we were able to conclude that a certain weighting of factors, when combined could improve our tipping percentage to over 64%. This is not a massive improvement than the other factors in isolation, but it is more, and importantly more than the bookies.

But of course tips is one thing, betting on games that will return value is another thing again.

For each of the factures we assign two things, firstly the margin we expect given the teams picked, including standings etc, and then the confidence we have in that result. The higher the margin the more confident we are. For example, in round 2 of this year we predicted Brisbane would beat New Zealand by 16 points. Brisbane at home, and having won the previous week were quite important, but their experience and player rankings were more important still (see the blue line from the 8th to 15th players where Brisbane start to dominate).

Of course the result was pretty close to our prediction with Brisbane winning by 16 points, but critically we didn’t recommend a bet. We didn’t recommend the Warriors either, even though the value was higher (ie the odds offered by the bookies was higher than we expected, but the probability of a win was lower than our cut-off).

For betting recommendations, we use a simple rule of thumb, of the tip confidence probability multiplied by odds and treat as a bet recommendation if the expected value is greater than 1.10. However, there are exceptions under certain conditions - for instance, we never bet on a team with less than a 35% probability of winning (40% early in the season where we have less current data) because, at the tails of the distribution, there is less confidence in the accuracy of the model due to there not being a statistically significant data set on upset results with less than 35% probability of winning.

We also classify our tips and bets, including novelty bets and our famous bets of the week (see below for fame!). So, for round 2 this week we had:

Does it work?

Well its very early days, but the early signs are encouraging, certainly helped by some generous specials.

From our experience with the AFL we have seen the following results:

We are cautiously optimistic we can do the same for NRL and look forward to our cricket modelling starting soon as well.

We would be delighted to hear if you have any thoughts on predicting results and any suggestions to improve our modelling further. As always, we only recommend betting if you can afford to lose, as there is no guarantee. Just slightly more of a chance than not following us!

Hope you enjoy , Facebook (TipBetPro) and Twitter (#TipBetPro_NRL).

Featured Posts
Recent Posts
Search By Tags