25th – 31st March Bookings Model Update

Domestic action returned after a 2-week hiatus for the internationals, and a new league started. The Brazilian Serie A kicked off, and predictions also commenced for the MLS – it was a another busy week for the cards code, and here is how it fared…

Weekend Results

The weekly results are a bit different this week – they are (and will be) a lot more colourful going forward.

Gold = High Value on the Total (unders only)
Grey = Good Value on the Total (unders only)
Red = Value on Away Handicap
Green = Value on Home Handicap

This is just to make it more obvious where the value lies at a glance, just to automate the process.

MLS predictions and Brazil Serie A have also been included at the bottom – more on those later.

Available to view as a pdf here – Bookings Model Tracking – Week Recap

85 games covered this week (15 MLS, 10 Brazil Serie A included)

In terms of the main lines posted to the group, Saturday saw a much higher volume than usual, but 5 wins from 12 main lines led to an overall loss of 2.68 units. Sunday, however, really came through. 5 wins and 1 void from 7 bets – an overall profit of 5.84 units on main lines alone. Accounting for ladders and doubles, that figure was way higher.

March Results

In terms of the tracking, there was a long break throughout March as I went away for a week or so, then we had the internationals. With another weekend in profit, we came out with some really nice profit on the main lines for the month which is pleasing.

Remember, the group is free to join here.

Lifetime Results

Unders continue to dominate, win-rate continues to hover around the 60% mark which is ideal. Serie A, Scottish Premiership and Premier League have been the best performing leagues, with cup competitions also up there (off a decent sample size now).

Still need to be wary of Bundesliga and La Liga though.

The reason I decided to post handicap bets is because of the above data – it is predicting splits to an accuracy just shy of 98% off a sample size of over 550 games. With the performance as good as that of the totals (which I didn’t expect), I have confidence in the predictions.

However, unlike with totals, it’s difficult to determine whether the numbers actually generate winners. For totals, if the model prediction is below the market line, it suggests an under. Then, if the total cards in that game ends up below the market line, it tracks as a winner – simple enough.

With the handicaps it’s a bit more complicated – because there is no link to the actual handicap number on bet365 (at least without manually tracking each game). If I simply go off the team to get most cards, it will be overly simplified as the home team usually gets less cards, the the model usually predicts that.

Therefore, I think the best way to evaluate it’s success (for now at least), is to determine how close the predicted split compared to the actual split – as shown above. Any suggestions on improvements, please do let me know.

Anyway, lets talk a bit about the newly-added leagues.

MLS Plan

The plan for the MLS is to try and determine and optimal sample size of data to start using the current season model. Each week, I will make predictions using data from this season.

As it stands, after six gameweeks, using data from this season is currently not viable – the sample size is just way too low. Also, with the way the American leagues are structured, teams don’t play home-away-home-away. They can play a few away games in a stretch, then a few home etc. So for one of the teams, after 5 weeks they had played one home game, and received 0 cards in that one game, which the model did not respond well to.

Also, the MLS games are pretty much all highlighted, which shows there are quite notable discrepancies between the market lines. Again, the reason is that model predictions are based off small sample sizes, so can be heavily influenced by anomalies.

The above point mostly applies actually to the handicap bets – it predicted totals relatively well with 9 / 15 winners, and 12 / 15 unders. But still, it’s only been one week of testing, so a bigger sample size is needed. My current standpoint is, five weeks of data is not sufficient to base predictions off.

Once predictions become better, I will begin sharing them to the Telegram Group, to further increase the volume.

Brazil Serie A Plan

The Brazilian model is different in the sense it is based off data from last season, and predictions will be done every week in order to determine if it is a viable strategy to base predictions off last seasons’ data. And if so, how long for.

What I would also like to do, is do an additional set of predictions based off current data (after maybe 4 weeks or so), and see where there is a turning point whereby using current season data seems to outperform previous season.

I will then be able to use this information for the start of the European leagues later this year in summer.

There are a few challenges with the current method – the first being that 4 teams were promoted from Serie B. Predictions for these sides (Santos, Mirassol, Ceara and Sport Recife) are based on their data in the Brazilian second division last season, which obviously isn’t ideal. However, it was a toss up between ignoring these teams and having an incomplete dataset, or using a less relevant set of numbers.

Eventually I settled on doing a bit of extra work, getting predictions sorted and having more data to go off. If the predictions are useless, I can use that information going forward, but if they are useful then that is perfect.

Unsurprisingly, all four fixtures involving newly-promoted sides were flagged as value – two on totals, and two on handicaps.

I’ll definitely need more time on this, but we got 5 / 10 winners on totals. Overall, the predictions actually looked relatively reasonable, although promoted sides were a bit low on the handicaps – understandable as data is based off them being at the top of the league, which in reality they will likely be closer to the bottom.

Regardless, it is great to have these leagues setup and added into the model – the predictions right now are not quite at the level I need them to be at, so they won’t be shared.

Also as a side-note, I had a few mentions that images were not displaying properly – hopefully that has been fixed but please do let me know if there are any persistent issues.

Plenty of steps in the right direction, but still a lot of work to be done. As always, I really appreciate your support on this project!