Week 4 Review and Week 5 NFL Picks from the Machine
I don’t have a lot of time to pontificate this week so I’ll get straight to the analysis.
Below are the Machine’s week 4 results:
Last week was mediocre at best from the Machine. It went 7 wins and 6 losses against the spread, which might be considered an average week for many handicappers (53.8% makes you almost break even depending on VIG and operating costs). But I was definitely disconcerted by the way the losses came in. The performance curve was inverted with respect to the absolute size of the factor – the Machine was 5 of 7 for picks under the critical 3.0 threshold and a putrid 2 of 6 for games above the threshold.
The LVH Supercontests cards took a hit with TheSanction going 2/5 and Rotoquant going 1/5. They currently stand as follows:
TheSanction (tied for 259th out of 1400): 12 wins, 8 losses
Rotoquant (tied for 453rd out of 1400): 11 wins, 9 losses
So what happened? Above the 3.0 threshold, based on backtesting and last years performance, we expect picks to be correct 70% of the time. So a reasonable question to ask, one we should be asking every week as we compile more data, is this: how likely is it that we got these results if the edge is really 70%? To answer questions like these we need something called the binomial probability formula, which computes for us the probability of achieving exactly k successes in n independent trials with probability of success for each trial of p. This sounds simple enough but the formula is fairly complicated looking – if you’re interested you can find out more about the binomial probability formula here, and it’s buddy the binomial coefficient formula here.
The answer is that if your edge is truly 70% there is a 7.01% chance you will see 2 or fewer correct picks in any set of 6 picks – which amounts to about once every 14 weeks assuming 6 games above 3.0 each week. This number is not yet in the realm of the impossible with only a 6 trial sample size, but it certainly isn’t encouraging. Fortunately we can put this performance in the context of the performance so far this season to get an idea of whether or not we should start to worry about our method. Through four weeks the algorithm has 16 wins and 10 losses ATS above the 3.0 mark, including last week. The probability of this performance assuming a 70% edge is still fairly low at 12.5% – if we flip the equation on its head and ask ‘what’s the edge at which this performance becomes more likely than not?’ we backdoor calculate the likely edge at a still healthy 59.5%.
This isn’t terrible – you could make a living picking games 59.5% ATS… But, as conscious as I am of falling into the trap of making excuses for a mathematical method instead of honing up to it not quite being performant, I think there are valid excuses for why it’s not tracking exactly to the 70% mark as expected (and, like I said, 59.5% is still good but I’m expecting great). The first and most obvious is sample size, but let’s skip over this for now because it’s always a concern when testing any method of prediction and there’s no point in beating a dead horse. The second, more interesting possible reason for underperformance has tangible ties to the real world… Although the method is highly sophisticated, math is still math and numbers are still numbers – they don’t know a teams strategy or understand key personnel moves the way that a head-coach or NFL insider might. There’s no way to get around using moving averages, for instance, to characterize a team’s performance, and moving averages beyond a 4-week window will necessarily include some of last year’s data. As the season goes on – and as was reflected in the backtesting – these moving averages will converge on the true values for each team for this season. Last-year’s less-than-perfect figures will be relied on less and less to determine the picks, and as a result we might expect algorithmically driven picks that use these measures to improve as the season goes along… or so the theory goes.
Anyway – every week is interesting and gives us more data to comb through. This week the Machine took it on the chin for games over the 3.0 factor. The two weeks prior it made a killing. We will keep checking week by week, but so far I’m staying the course and pressing ahead.
In a side note about last week – my life-long battle with Tom Brady and Bill Belicheck continues. I hate them, they hate me, it’s well understood and documented in the Rotoquant archives – they threw the game to piss me off … and it worked.
Moving along. Below are the week 5 picks:
My card picks will be as follows:
TheSanction – DAL, SD, DEN, CLE, SEA
Rotoquant – DAL, SD, DEN, SEA, CHI
Good luck in all your contests in week 5. We’ve increased our payouts at Victiv for the 5th week in a row, running $38,500 in guaranteed payouts and free contests this weekend… People keep getting money and having a blast playing DFS at Victiv, and you continue to not be one of them – what are you waiting for?!
To help you grab all the money the MegaTool 2.0 has been released this week and is available for download here (it requires the free Wolfram CDF player, available here, to run). This fantasy sports weapon of mass statistical destruction gives you data-driven visualizations and dynamic access to advanced statistics to help you win your daily or season long leagues. Analyze utilization, player participation, target breakdowns, strength of schedule and next week’s matchups in a single high-powered app … This tool is still evolving, but so far the it has been met with rave reviews. Teaser for next week – I hope to integrate a user-driven fantasy point projection mechanism in version 3.0 … stay tuned… Finally, for sports, fantasy sports and sports betting analytics updates throughout the year Follow me on Twitter @TheRotoquant.
Good luck in all your contests – see you next week.