Tweet
Full Disclosure
The mathematical model used in The Limper’s calculations is firmly based on past performance data, and it uses this data to compare with future opponent matchups, and thus project outcomes of success or failure. Basically, if conditions remain the same (or, at least similar), what has happened before (like passing yards gained, 3rd down conversions, time of possession, and so forth), will happen again; and based on this, the winner and margin of victory may be known. Since 2009 I’ve deployed this model every NFL season; and, except 2010 and 2018, it’s been semi-successful, hitting close to 70% straight-up and between 53% and 54% against the spread. In fact, including 2010, it hasn’t failed to reliably project specific variable outcomes; nevertheless, I spend every off-season going under the hood, looking for bugs and ways to improve the model’s performance – and then came 2018.
The model’s weak performance last season - 62.6% SU and 43.5% ATS – was a shocker, and I spent weeks looking for holes – looking for some logical reason for it to have gone off a cliff – something I could fix and move on. What I found, however, was that nothing was broken – there was nothing to fix; rather, it was the fact that the foundation of the model – past performance data – had simply become less reliable as indicators of future performance; and this had been going on, albeit, slowly, over the past decade. It wasn’t a matter of re-weighting the variables used in the calculation, or tinkering with the algorithm itself – it was a matter of rethinking the assumptions on which the model was based, and I don’t know how to do that.
Essentially, a mathematic model is constructed in a static universe which is logical, unchangeable, and certain, ie. where 2+2 always equals 4. Moreover, we live in a universe of semi-static construction; where conditions are such as to be rational, logical, and pretty much unchanging. However, our universe is, nevertheless, dynamic – far less mathematically predictable, especially so when speaking of human behavior. Of course, this is no surprise, but despite the fact of reality’s dynamic nature, 2+2 still equals 4, and as long as this dynamism doesn’t intrude – as long as the conditions of reality remain more or less static - other mathematical calculations still work.
The problem is that the conditions of football reality have changed considerably over the years, and have made mathematical calculations of behavior within the football context far less reliable. Specifically, what has changed is that game plans for many teams became, week to week, a lot more dynamic, highly fluid and ultimately disposable - IN-GAME. Pass-crazy teams became, suddenly, conservative, while lesser pass-reliant teams would switch up and become pass-first-run-averse teams; and such dynamic, in-game, game-plan reversals made the model’s static, past performance data projections fairly unreliable – so, what to do?
Such changes are the result of technologically-enabled, enhanced game management, where in-game opposing team tendencies are observed, recorded, and instantly analyzed for immediate response. Of course, everything depends on who’s calling the shots and not all coaching staffs are equal; a great many head coaches and assistants are still bone-headed ex-players, who despite in-game data telling them to go for it on 4th, rely on past-performance – static interpretations of what worked last week or last year, and punt instead. Many, on the other hand, can read and do the math, and make dynamic judgments based on what’s happening there and then; and it is their numbers who will inevitably increase, making static-based projections even less reliable - so, again, what to do?
I’ve been building new variables which attempt to account for coaching talent and tendencies, but that data is not easily available and I just may give that up. Thing is, although static-based, past-performance projections may be less reliable than previous, they are still useful as a point to begin; so, bottom line, I’m not changing anything. This means, of course, that my weekly projections are more than ever just a guide, and not predictions to take to the bank.
The model needs 3 weeks of data to run so, for Weeks 1-3, it must use performance data from the prior season (which I manually (ie. subjectively) update based on upgrades and downgrades made on each team by position), making early projections even less reliable. The last 2 weeks of the NFL season are also tainted with starters resting and teams tanking, so that only a handful of games can be projected with any confidence at all.
Every week, I’ll be posting first projections for the following week each Tuesday (except this week), which will be based on performance data gathered the week just ended, as well as prior weeks. The Tuesday projections, however, won’t include players injured in the prior week, and sure to be OUT the following week. The second posting on Wednesday nights or early Thursdays will include these injury factors, as will the Saturday and Monday morning postings. So, if you’re following, check the latest post prior to kick-off for the model’s best projections. The model grades itself based on the last MOVs posted before each game and the Vegas Insider closing lines.
Also, although the model does project “SCORES” it has never been successful at over/under picks. Its calculations are intended to result in a MOV, and the projected SCORES are only a by-product of that process, so take them with that in mind.
The NFL is a hard thing to beat, and there are no shortcuts to winning, but if you wager, you can do a lot worse than use the numbers as a guide.
GLTA
The mathematical model used in The Limper’s calculations is firmly based on past performance data, and it uses this data to compare with future opponent matchups, and thus project outcomes of success or failure. Basically, if conditions remain the same (or, at least similar), what has happened before (like passing yards gained, 3rd down conversions, time of possession, and so forth), will happen again; and based on this, the winner and margin of victory may be known. Since 2009 I’ve deployed this model every NFL season; and, except 2010 and 2018, it’s been semi-successful, hitting close to 70% straight-up and between 53% and 54% against the spread. In fact, including 2010, it hasn’t failed to reliably project specific variable outcomes; nevertheless, I spend every off-season going under the hood, looking for bugs and ways to improve the model’s performance – and then came 2018.
The model’s weak performance last season - 62.6% SU and 43.5% ATS – was a shocker, and I spent weeks looking for holes – looking for some logical reason for it to have gone off a cliff – something I could fix and move on. What I found, however, was that nothing was broken – there was nothing to fix; rather, it was the fact that the foundation of the model – past performance data – had simply become less reliable as indicators of future performance; and this had been going on, albeit, slowly, over the past decade. It wasn’t a matter of re-weighting the variables used in the calculation, or tinkering with the algorithm itself – it was a matter of rethinking the assumptions on which the model was based, and I don’t know how to do that.
Essentially, a mathematic model is constructed in a static universe which is logical, unchangeable, and certain, ie. where 2+2 always equals 4. Moreover, we live in a universe of semi-static construction; where conditions are such as to be rational, logical, and pretty much unchanging. However, our universe is, nevertheless, dynamic – far less mathematically predictable, especially so when speaking of human behavior. Of course, this is no surprise, but despite the fact of reality’s dynamic nature, 2+2 still equals 4, and as long as this dynamism doesn’t intrude – as long as the conditions of reality remain more or less static - other mathematical calculations still work.
The problem is that the conditions of football reality have changed considerably over the years, and have made mathematical calculations of behavior within the football context far less reliable. Specifically, what has changed is that game plans for many teams became, week to week, a lot more dynamic, highly fluid and ultimately disposable - IN-GAME. Pass-crazy teams became, suddenly, conservative, while lesser pass-reliant teams would switch up and become pass-first-run-averse teams; and such dynamic, in-game, game-plan reversals made the model’s static, past performance data projections fairly unreliable – so, what to do?
Such changes are the result of technologically-enabled, enhanced game management, where in-game opposing team tendencies are observed, recorded, and instantly analyzed for immediate response. Of course, everything depends on who’s calling the shots and not all coaching staffs are equal; a great many head coaches and assistants are still bone-headed ex-players, who despite in-game data telling them to go for it on 4th, rely on past-performance – static interpretations of what worked last week or last year, and punt instead. Many, on the other hand, can read and do the math, and make dynamic judgments based on what’s happening there and then; and it is their numbers who will inevitably increase, making static-based projections even less reliable - so, again, what to do?
I’ve been building new variables which attempt to account for coaching talent and tendencies, but that data is not easily available and I just may give that up. Thing is, although static-based, past-performance projections may be less reliable than previous, they are still useful as a point to begin; so, bottom line, I’m not changing anything. This means, of course, that my weekly projections are more than ever just a guide, and not predictions to take to the bank.
The model needs 3 weeks of data to run so, for Weeks 1-3, it must use performance data from the prior season (which I manually (ie. subjectively) update based on upgrades and downgrades made on each team by position), making early projections even less reliable. The last 2 weeks of the NFL season are also tainted with starters resting and teams tanking, so that only a handful of games can be projected with any confidence at all.
Every week, I’ll be posting first projections for the following week each Tuesday (except this week), which will be based on performance data gathered the week just ended, as well as prior weeks. The Tuesday projections, however, won’t include players injured in the prior week, and sure to be OUT the following week. The second posting on Wednesday nights or early Thursdays will include these injury factors, as will the Saturday and Monday morning postings. So, if you’re following, check the latest post prior to kick-off for the model’s best projections. The model grades itself based on the last MOVs posted before each game and the Vegas Insider closing lines.
Also, although the model does project “SCORES” it has never been successful at over/under picks. Its calculations are intended to result in a MOV, and the projected SCORES are only a by-product of that process, so take them with that in mind.
The NFL is a hard thing to beat, and there are no shortcuts to winning, but if you wager, you can do a lot worse than use the numbers as a guide.
GLTA
Comment