4th-10th March Bookings Model Results
As promised, I’ve got analytics ready this week, all the data so far has been parsed and grouped up – it’s time to see what factors are best at predicting the total number of bookings in a game. The results have been incredibly informative – this write-up will be a lot more data-heavy and objective than usual, so if numbers are your thing you might enjoy this one (I did).
I’m going to break up the results into four sections:
- By Unders / Overs
- By Model Difference
- By League
- By Market Line
First off, let’s start off by going over the results this week in isolation. Also, it might be worth reminding you that I have been posting the predictions to a Telegram Group, which is free to join using this link.
Weekend Analysis
Full results available to view in the PDF – Bookings Model Tracking – Week 6
- 87 Bets
- 62 Unders (71.2%)
- 56 Winners (64.3%)
Overall, the results we achieved this week were in line with what we have become accustomed to. Hovering around the 70% mark for unders, and a 60% win-rate.
However, the model excelled this week at identifying where the true value lay, which was absolutely ideal.
If you had joined the group after my last post and tailed the highlighted main lines you would have won 9/10 bets, and that’s without even considering the ladders – a few games finished on 0 cards!
- 4/4 Main Lines from Tuesday to Thursday
- 1/2 on Friday
- 3/3 on Saturday
- 1/1 on Sunday
Over the last two weeks, this is how the results are looking on main lines. We’ve breached the +20u mark already.
Despite the good weekend, there is still a lot of work to be done, and I do believe the process of selecting main lines can be further improved. With that in mind, let’s look into a more comprehensive breakdown of the numbers so far.
Lifetime Analysis
Overall, the model has predicted 1799 cards, there has been 1684 shown in total, which is an accuracy of 93.7%. It’s not been a secret that the model has been adamant on taking unders, but it has actually overestimated the total number of cards shown, which may come as a surprise. However, it just goes to show that unders are the play. If you were to blindly back the unders on cards for every game every week, it would be a profitable strategy.
Results by Unders / Overs
411 games covered in total, 254 winners – and it’s the unders market to thank really for the decent win-rate up until now.
The numbers tell the story here – going forward, I’ll have to be more wary with taking overs.
Results by Model Difference
But it’s not quite just as simple as unders = good, overs = bad. Or is it? Let’s break it down a bit further.
The unders show a nice, smooth positive correlation between win-rate and model prediction. When the model is way lower than the market line, in most cases the model prevails, and win-rate correlates nicely with a bigger difference. This is exactly what we want. The optimal range seems to be the 10 to 12 or greater, but anything greater than a difference of 6 booking points should be taken, and that’s off a pretty decent sample size. Of course, those numbers need to be a lot larger to draw any concrete conclusions, but that data is seriously promising.
Overs data is much less convincing, although it is worth mentioning that the dataset is a quarter of the size. Because of that, we have a pretty large anomaly with 2/9 winners in the 8-10 band, which is absolutely not what is expected. Without drawing too many conclusions again, more data is required on the overs predictions, but for now, taking caution is sensible.
Results by League
Breaking the results down by league also yields some interesting trends – Serie A, Championship and Premier League are the standout domestic league,s with cup competitions also proving lucrative.
La Liga and Bundesliga have provided poor results thus far, and again require a bit more monitoring. Ideally, we need to be hovering around the 100-200 game sample size here before reacting, but I’ll probably have to err on the side of caution once more for the poorly performing leagues.
It’s surprising to see the Championship and Premier League on such a high win-rate. The expectation would be for these games to be getting the highest volume of bets, and therefore have the sharpest lines. However, it’s the overs that seem to be inflated more, which is probably actually working to our advantage by consistently creating value on the unders,.
Results by Market Line
Finally, I speculated last week that it could be a play to simply take the under when lines are much higher than usual (examples would be in derby games), and take the over when lines are much lower than usual (cup competitions where there is a large supremacy, for example).
I’ll be honest, these results aren’t particularly useful given the extremely low sample sizes at the extreme ends, which is what we are investigation.
However, there seems to be a slight correlation between win-rate and taking the unders on higher lines. Whereas taking overs on higher lines haven’t been as successful (which was what I thought might happen).
However at the other end, that assumption falls apart. Taking the over on extremely low lines wasn’t particularly great, less than a 50% win-rate.
The lesson to learn here, is that it’s viable to taken a under for a high line, but not the same for taking the over on a low line. But again, more data is required.
So much to learn from the data so far, and I’ll be using those aforementioned lessons going forward. Results have been so much better than expected up until now, but that doesn’t mean the project is as good as it could be. I’ve enjoyed putting in the work so far, and will continue to do so irrespective of results.
Now, we have a quiet midweek period (no predictions on European games for now – I still need to get my head around those), but it will be business as usual over the weekend. After that, we have the final international break for 2 weeks or so. Again, very doubtful for any predictions, but not completely off the table.
I’m also debating whether to cap the Telegram Group. Cards lines can tend to drop quite fast, and a large group tailing will contribute to that, which would not be ideal. A preliminary figure would be 100 members in the group, but again, ideas are welcome.
Thanks for the support as always!