May 2025 Update - Sports Betting, Model Calibration & Speed of Iterations
Narrowing the timeline.
Building momentum and speed of iterations is the main concept I’ve been thinking about in May.
How can you make rapid iterations and adapt systems (sports betting models in this case) to get better updated info and results without taking forever to do so.
Code and automation is the obvious answer but also narrowing the timeline is another less discussed one.
In May we had some really good progress whilst only “flow state” working for about 4-5-hours a day on average, around 30-35 hours / week. This is not just daily to-do’s but actual building of models, calibration, coding etc.
The key was just understanding leverage points and narrowing the timeline to a ridiculous degree by aligning tasks and projects properly.
By narrowing timelines I mean running parallel projects (not more than one thing at the same time) but understanding lag-time.
For example, if it takes me 10 hours to build a scraper, 10 hours to develop the scripts within a model and 10 hours to calibrate the model, that’s 30 hours of active flow-state work which could be 3 or 4 days at a push.
Issue lies in the lag time surrounding this and the “waiting for” parts.
You cannot move onto part 2 until part 1 is completed and same for part 3.
So a 30 hour project could balloon to 10 days without proper alignment of multiple projects.
The main 2 I’ve noticed in these examples is;
Scraping entire seasons takes up to 12-15 hours (generally we need 2 per league).
We need the odds manually collected for these games (VA takes around one day per season per league).
So as you can probably see where this is going, instead of working on one model at a time, you can rotate 3 and complete them in the same period of time without any additional work, but this allows you to iterate almost 9x faster as you are doing similar things at once (batching).
Just a fancy way of saying you can become more effective by just planning when lag times will be (not just from business perspectives but also personal and just make sure something is running overnight for example).
Learning iteration or “you just get better”
But the best iteration is the learning-iteration.
This has been incredibly obvious looking back at the early versions of the sports betting models that I have developed.
Mostly they just look trash.
The most clear example of this is the Rugby union model, which took months to get right and profitable (resulting in a RU V5 which actually does work really well on the whole). But this build did take maybe 400 hours to complete + all the millions before that that we just ignore.
Fast forward to making improvements on this, a V6 with similar results(but different underlying logic) took 20 hours and 2-3 days.
But on average improved the combined results of rugby union by about 1.5-2%.
Obviously the Version 6 doesn’t get built without the 400 hours on the V1-5 previously but the point is that the results are not granular or linear nowadays, especially with Ai being integrated into everything.
The results are now exponential.
For example, adding just 2% edge in sports betting over the course of a season/sport/year where you have say 500 bets a year and previously went 56% for: 280 wins and 220 losses (great results by the way) for: 280 - (220*1.1) = 38 units profit & 7.6% ROI per bet.
These are incredible results but if your V5 to V6 takes you 2% better to 58%, you don’t make 2% more, you actually make almost 100% more (when you factor in consistent bankroll adjusted betting throughout the season) but even on the simplified version as per above = 290-210, for: 290 - (210*1.1) = 59 units profit & 11.8% ROI per bet.
So essentially a 50% increase in profit from a 2% increase in edge performance. That’s the core of why iterations are so essential.
In my opinion this is why going into the tank for a couple of week stretch to make sure everything is essential and then taking a few days or even weeks completely off is a way better way to work than just X amount of hours per day where you aren’t all-in on the results.
Which is why the ancient industrial age work week still being a thing in 2025 is absolutely ridiculous but that’s a rant for another time.
Anyway, I had Claude build this fancy interactive chart to show the exponential vs linear difference of edge increases, feel free to play around with it and become excel rich. There’s nuances to this around liquidity as well, getting a £1,000 bet vs £10,000 on a bet is a very different challenge where you will likely get a worse price (and hence ROI decreases).
Interactive version here: https://claude.ai/public/artifacts/cead0d9e-dcb5-45bf-a0cb-345ce12524ba
Sports Betting Situations & Updates
May was a great month for builds and a good month for betting.
Although overall volume goes through the floor over summer due to no hockey and no rugby union, we now have improved a number of models that return good rates on decent volumes.
Rugby league (both super league and NRL) have been completed and perform very well both in back-testing and in May so far, returning about 20% ROI collectively.
Cricket has improved with the new V6 model which (hopefully) has finally solved the issues around that, although it’s still early days.
Aussie Rules Football (AFL) was created in May, with a V5 and V6 that both perform well, although its still early days on live betting for this.
May is likely going to finish around +£10-12k or so depending on how the last week goes, which is a good improvement from the negative 5k in April but nowhere near our March performance.
Based on £1k units we’re probably sitting a bit lower this month around 8% ROI this month.
We’ll see how performance goes through the summer with the bankroll growing in preparation for our November and December £50k/month attempts for the end of the year.