A New Way to Think About "Buying Low" and "Selling High&#3 (1 Viewer)

[Aabye]

I don't know that tiers should be invited into the conversation. If we are hypothesizing a calculation, draft status is far too subjective to include. I don't know why would want a variable that suggests that we rank Felix Jones ahead of Arian Foster. Tiers only hamper the search for a perfect calculation. We know that LeSean McCoy is younger than Jamaal Charles. Tiering them together removes that pertinent information from our calculation. We need to use the same variables we do for re-draft, then invite a new set, to account for dynasty settings. The difference between re-draft and dynasty is the number of years. Bringing us to projected points scored and years of production.It gets tricky projecting years out. I understand that and personally believe that is the ONLY reason we can't have a VBD calculator as relative and valuable as the re-draft version. But, that doesn't mean adding more subjective variables will fix that.

I would love for the discussion to proceed this way: debating the merits of various ways of modeling average remaining dynasty value.Hell, if we get to the point that everyone agrees that we should be figuring out models that assign dynasty value by projecting remaining career VBD numbers, I'd consider it quite a success. We could then get onto the business of figuring out how to actually construct them.What I've seen is that we get lots of threads about rankings ("is LeSean McCoy a top 5 dynasty RB?") or comparison ("Who's better, Mikel LeShoure or Ryan Williams?") and almost nothing about how much the players are actually worth in terms of VBD. Those threads certainly have their value and I don't mean to disparage them, but I'd like to figure out how we can assign sensible values to these guys rather than just ranking them.Good modeling would let us do cross-positional analysis much more easily, it would let us compare rookies to vets much more easily, and it would give us a much better sense of how to actually improve.

 

[Concept Coop]

I have $ to invest in this project if a small group of people are serious about getting this off the ground. I think it's worth the money/time just to find out if a set of dynasty rankings (driven solely by measurable factors) is really more accurate than what people or generating on the fly currently.

I would be very interested in at least having the conversation.

 

[Concept Coop]

I would love for the discussion to proceed this way: debating the merits of various ways of modeling average remaining dynasty value.

 

[Biabreakable]

I liked your list of factors that go into each players value. The question is how to implement them?

While I agree having a VBD # for every player is the most useful way to compare players, especially of different positions, some of those factors on your list cannot be applied to a VBD #. Well maybe they could in some cases give a % bump up or down, but if you bump a VBD # 15-20 times slightly because of differing factors I am not sure how comfortable I would be with that number at the end of doing so. I have fooled around with this in the past but it seemed very subjective to me.

The other way to do it as was suggested would be to tier VBD groups, then rank the players within that tier based on the other factors. Still subjective but maybe not as bad as fudging the numbers multiple times. This way was more comfortable for me than +/- 5% VBD for each and every factor.

Often in these discussions there is an issue of different league rules that change VBD #s and disagreements because of these different perspectives. So I suggest using FBGs projections just so everyone would be on the same page as a starting point.

The question still remains how best to make each of those factors tangible or if those factors are already baked into the original projection.

 

[FantasyTrader]

I liked your list of factors that go into each players value. The question is how to implement them...This way was more comfortable for me than +/- 5% VBD for each and every factor...The question still remains how best to make each of those factors tangible or if those factors are already baked into the original projection.

You let history be your guide. You initially use a best guess when weighting the values in v1.0 and adjust from there. "Injury concerns" for instance. If 48% of a group have little/no injury history, 28% have moderate injury of the nagging variety and 6% have a history of catastrophic injury, you would weight catastrophic injury 8 times heavier than little/no injury history within the injury concerns factor.

 

[Lab]

I liked your list of factors that go into each players value. The question is how to implement them...This way was more comfortable for me than +/- 5% VBD for each and every factor...The question still remains how best to make each of those factors tangible or if those factors are already baked into the original projection.

You let history be your guide. You initially use a best guess when weighting the values in v1.0 and adjust from there. "Injury concerns" for instance. If 48% of a group have little/no injury history, 28% have moderate injury of the nagging variety and 6% have a history of catastrophic injury, you would weight catastrophic injury 8 times heavier than little/no injury history within the injury concerns factor.

I really like this thread, and I agree with this approach. For all the concern, what we're discussing here (at least my conception of it) seems very doable. It's complex to discuss in words, but it's not too hard too express in a spreadsheet. I do not have time right this moment, but I will try to start a simple Google spreadsheet soon that puts it into practice. Each row is a player, and each of 10+ columns is a value assigned to that player in a category (years remaining, points per year, etc). Then each column is assigned a relative value weighting that acts as a multiplier on its base value. An additional multiplier might be used for confidence interval.That's what I'm thinking at least.

 

[Concept Coop]

I liked your list of factors that go into each players value. The question is how to implement them...This way was more comfortable for me than +/- 5% VBD for each and every factor...The question still remains how best to make each of those factors tangible or if those factors are already baked into the original projection.

You let history be your guide. You initially use a best guess when weighting the values in v1.0 and adjust from there. "Injury concerns" for instance. If 48% of a group have little/no injury history, 28% have moderate injury of the nagging variety and 6% have a history of catastrophic injury, you would weight catastrophic injury 8 times heavier than little/no injury history within the injury concerns factor.

I'd be curious to look into % of games missed as an indicator. I.E. RBs 24-25 years of age who have missed 10% of their eligible games go on to mis an average of x/games a season. 10% of these players retire early due to injury. So on and so on. Taking it a step further, once could also start to look at the individual injuries as well.

 

[FantasyTrader]

Here's an idea. What if, instead of trying to figuring out Ryan Mathews value, then Felix Jones' value, etc. we work it from the other direction? Define the historic characteristics of a player with 1175-1200 fantasy points remaining in his career, the characteristics of a player with 1150-1175 FP remaining in his career, etc.

Prolem is, you'd need to go back 15 years before you're able to do that even in the ballpark of a meaningful sample size. Every year when the season is completed, you answer a set of questions for every player within their tab: FP scored, injury occured, games missed, misdemeanor arrest, felony arrest, etc. It could be as easy as clicking a radio button next to anything that applies. Instead of building a database to fetch remaining value based on player characteristics we build it to find players with characteristics consistant with 900-925 carer FP remaining, etc.

 

[msommer]

Here's an idea. What if, instead of trying to figuring out Ryan Mathews value, then Felix Jones' value, etc. we work it from the other direction? Define the historic characteristics of a player with 1175-1200 fantasy points remaining in his career, the characteristics of a player with 1150-1175 FP remaining in his career, etc. Prolem is, you'd need to go back 15 years before you're able to do that even in the ballpark of a meaningful sample size. Every year when the season is completed, you answer a set of questions for every player within their tab: FP scored, injury occured, games missed, misdemeanor arrest, felony arrest, etc. It could be as easy as clicking a radio button next to anything that applies. Instead of building a database to fetch remaining value based on player characteristics we build it to find players with characteristics consistant with 900-925 carer FP remaining, etc.

How do you determine the remaining FPs though - don't you need the Aabye method for that?

 

[FantasyTrader]

I liked your list of factors that go into each players value. The question is how to implement them...This way was more comfortable for me than +/- 5% VBD for each and every factor...The question still remains how best to make each of those factors tangible or if those factors are already baked into the original projection.

You let history be your guide. You initially use a best guess when weighting the values in v1.0 and adjust from there. "Injury concerns" for instance. If 48% of a group have little/no injury history, 28% have moderate injury of the nagging variety and 6% have a history of catastrophic injury, you would weight catastrophic injury 8 times heavier than little/no injury history within the injury concerns factor.

I'd be curious to look into % of games missed as an indicator. I.E. RBs 24-25 years of age who have missed 10% of their eligible games go on to mis an average of x/games a season. 10% of these players retire early due to injury. So on and so on. Taking it a step further, once could also start to look at the individual injuries as well.

 

[FantasyTrader]

Here's an idea. What if, instead of trying to figuring out Ryan Mathews value, then Felix Jones' value, etc. we work it from the other direction? Define the historic characteristics of a player with 1175-1200 fantasy points remaining in his career, the characteristics of a player with 1150-1175 FP remaining in his career, etc. Prolem is, you'd need to go back 15 years before you're able to do that even in the ballpark of a meaningful sample size. Every year when the season is completed, you answer a set of questions for every player within their tab: FP scored, injury occured, games missed, misdemeanor arrest, felony arrest, etc. It could be as easy as clicking a radio button next to anything that applies. Instead of building a database to fetch remaining value based on player characteristics we build it to find players with characteristics consistant with 900-925 carer FP remaining, etc.

How do you determine the remaining FPs though - don't you need the Aabye method for that?

No. You do it by sampling players whose careers have ended.

 

[Aabye]

I liked your list of factors that go into each players value. The question is how to implement them...This way was more comfortable for me than +/- 5% VBD for each and every factor...The question still remains how best to make each of those factors tangible or if those factors are already baked into the original projection.

You let history be your guide. You initially use a best guess when weighting the values in v1.0 and adjust from there. "Injury concerns" for instance. If 48% of a group have little/no injury history, 28% have moderate injury of the nagging variety and 6% have a history of catastrophic injury, you would weight catastrophic injury 8 times heavier than little/no injury history within the injury concerns factor.

I'd be curious to look into % of games missed as an indicator. I.E. RBs 24-25 years of age who have missed 10% of their eligible games go on to mis an average of x/games a season. 10% of these players retire early due to injury. So on and so on. Taking it a step further, once could also start to look at the individual injuries as well.

In my attempts to model, there is sometimes a problem finding a good "fit" in terms of sample size. So, for instance, if we want to compare some particular 25 year old RB to every other 25 year old RB, there will be a large sample size and fairly reliable data, but it's diluted and more or less unhelpful in projecting for a particular player. If we go in the other direction and look at (e.g.) RBs 24-25 year of age who have missed 10% of their eligible games, you might run into a sample size that's too small to help us out.This is both good and bad news. It's bad news because it means that plenty of relevant factors are going to be disqualified because they are either too broad or too narrow. But it's good news in that this gives us less to work with, which in turn makes the models easier to work with.I did a decent amount of data-mining where I had to just throw things out because there just weren't enough players to draw any sorts of conclusions whatsoever.This is not meant to discourage you from looking at this factor, but my gut sense is that you're going to get a data set that's too small to really work with.

 

[Aabye]

Here's an idea. What if, instead of trying to figuring out Ryan Mathews value, then Felix Jones' value, etc. we work it from the other direction? Define the historic characteristics of a player with 1175-1200 fantasy points remaining in his career, the characteristics of a player with 1150-1175 FP remaining in his career, etc. Prolem is, you'd need to go back 15 years before you're able to do that even in the ballpark of a meaningful sample size. Every year when the season is completed, you answer a set of questions for every player within their tab: FP scored, injury occured, games missed, misdemeanor arrest, felony arrest, etc. It could be as easy as clicking a radio button next to anything that applies. Instead of building a database to fetch remaining value based on player characteristics we build it to find players with characteristics consistant with 900-925 carer FP remaining, etc.

How do you determine the remaining FPs though - don't you need the Aabye method for that?

No. You do it by sampling players whose careers have ended.

When I was doing my own modeling, I looked exclusively at players who entered the league between 1985 and 2000. The reason for this was to get a sampling of players who play in the "modern era" (hence the cutoff at 1985) and players who have mostly retired or are close to retiring (hence the cutoff at 2000). So I think I was trying to do what FantasyTrader is saying.So the idea would be to get as large a batch of retired players as we can without screwing up the sample (e.g. we can't use Y.A. Tittle), then use those players to try to build a set of correlations by which we can measure current players.That idea of ticking off various boxes for a player (arrested, 25 years old, scored between 25 and 49 VBD points last year, etc) seems like a sensible one to me. Assuming that we could get the model up and running first, it should be fairly easy to incorporate new data, tweak the model, etc. I'm fairly low tech so I did all of my work in excel, which isn't a terribly great way to do it.

 

[ZWK]

The closest thing I know of to a complete statistical model of fantasy value is Football Outsiders' KUBIAK projection system for redraft leagues. It's designed to predict next year's fantasy points only, and is based heavily on projections for the team's performance (they also have to use their judgment to assign the values for some variables, like players' roles). A complete dynasty model might need to be a mixture of one projection system like KUBIAK for short-term value (the next year or two) and another projection system for long-term value.

Another approach, instead of trying to make a complete statistical model from scratch, is to start with conventional wisdom (based on ADP in startup drafts, or other people's rankings, or trade values) and then to analyze data to find ways to improve on those consensus rankings. You could start with just a small number of other variables (in addition to a player's consensus ranking), like their position, years of experience, and where they were picked in the NFL draft, and see if you can predict career VBD better by including those variables than by just going with the consensus rankings. Every time you find a variable that helps predict career VBD (better than you could with consensus rankings alone), you've found an error in the consensus rankings that you can take advantage of - certain types of players are being undervalued or overvalued by other owners.

That way you wouldn't need to build a full model from scratch all at once - with the consensus rankings as your skeleton you could gradually add more variables and build a better and better model. Over time more and more of your model will be based on objective variables and it will depend less on the consensus rankings, and maybe eventually you'll eventually get to the point where your objective variables allow you to leave out the consensus rankings altogether, in which case you'll have found a complete statistical model. But even if you don't ever get there, the model can still do a great job of estimating value and of identifying undervalued and overvalued players, which is what you can actually use as a fantasy owner.

One possible downside of this approach is that you'd need the data from old fantasy drafts in order to be able to make it work, so you'd have to get that data and you could only go back as far as you have ADP data. Another potential problem is that the fantasy conventional wisdom may have changed over time, so you might just find flaws in fantasy owners' previous views (e.g. overvaluing RBs relative to WRs) which are no longer present in their current ratings. But the advantages could outweigh that - you'd be able to start out with rankings that already incorporate lots of hard-to-quantify information, so instead of having to build everything from scratch you'd just be looking for specific variables that let you improve on what other people are already doing. Plus, you'd have a great chance of doing better than other owners at valuing players, since the approach is to start with other owners' valuations and then find ways to make them more accurate.

 

[msommer]

Here's an idea. What if, instead of trying to figuring out Ryan Mathews value, then Felix Jones' value, etc. we work it from the other direction? Define the historic characteristics of a player with 1175-1200 fantasy points remaining in his career, the characteristics of a player with 1150-1175 FP remaining in his career, etc. Prolem is, you'd need to go back 15 years before you're able to do that even in the ballpark of a meaningful sample size. Every year when the season is completed, you answer a set of questions for every player within their tab: FP scored, injury occured, games missed, misdemeanor arrest, felony arrest, etc. It could be as easy as clicking a radio button next to anything that applies. Instead of building a database to fetch remaining value based on player characteristics we build it to find players with characteristics consistant with 900-925 carer FP remaining, etc.

How do you determine the remaining FPs though - don't you need the Aabye method for that?

No. You do it by sampling players whose careers have ended.

how do you determine whether a 25 year old back is one with 900 or 1200 FP left?

 

[Cato]

Here's an idea. What if, instead of trying to figuring out Ryan Mathews value, then Felix Jones' value, etc. we work it from the other direction? Define the historic characteristics of a player with 1175-1200 fantasy points remaining in his career, the characteristics of a player with 1150-1175 FP remaining in his career, etc. Prolem is, you'd need to go back 15 years before you're able to do that even in the ballpark of a meaningful sample size. Every year when the season is completed, you answer a set of questions for every player within their tab: FP scored, injury occured, games missed, misdemeanor arrest, felony arrest, etc. It could be as easy as clicking a radio button next to anything that applies. Instead of building a database to fetch remaining value based on player characteristics we build it to find players with characteristics consistant with 900-925 carer FP remaining, etc.

How do you determine the remaining FPs though - don't you need the Aabye method for that?

No. You do it by sampling players whose careers have ended.

how do you determine whether a 25 year old back is one with 900 or 1200 FP left?

I agree with this point, I do like the idea of building a model and working it to see the best buy low, sell high moments. But if you can't predict future value, your success will vary more based on the predictiveness of future value, rather than modelling based on historical results. I suppose building empirical trends based on attributes may help assess future value, but that seems too high level clinical to sort current individual players. There's too many variables to control to extrapolate, unless it's done in a very generic sense... (unless I'm thinking about this wrong.)

 

[Aabye]

Here's an idea. What if, instead of trying to figuring out Ryan Mathews value, then Felix Jones' value, etc. we work it from the other direction? Define the historic characteristics of a player with 1175-1200 fantasy points remaining in his career, the characteristics of a player with 1150-1175 FP remaining in his career, etc. Prolem is, you'd need to go back 15 years before you're able to do that even in the ballpark of a meaningful sample size. Every year when the season is completed, you answer a set of questions for every player within their tab: FP scored, injury occured, games missed, misdemeanor arrest, felony arrest, etc. It could be as easy as clicking a radio button next to anything that applies. Instead of building a database to fetch remaining value based on player characteristics we build it to find players with characteristics consistant with 900-925 carer FP remaining, etc.

How do you determine the remaining FPs though - don't you need the Aabye method for that?

No. You do it by sampling players whose careers have ended.

how do you determine whether a 25 year old back is one with 900 or 1200 FP left?

I agree with this point, I do like the idea of building a model and working it to see the best buy low, sell high moments. But if you can't predict future value, your success will vary more based on the predictiveness of future value, rather than modelling based on historical results. I suppose building empirical trends based on attributes may help assess future value, but that seems too high level clinical to sort current individual players. There's too many variables to control to extrapolate, unless it's done in a very generic sense... (unless I'm thinking about this wrong.)

I'm not sure if this speaks to your concern, but I'm somewhat uneasy with the emphasis on "predicting" a player's value.I think the goal of predicting a player's future value is pretty much doomed. So much is random: injuries, situation, trades, retirement, coaching changes, etc. What I think we can do is generate pretty good data on average remaining fantasy values for groups. So, for instance, we can say something about the average remaining fantasy value of 25-year-old RBs who scored more than 100 VBD points the previous year.My idea is that we can analyze all sorts of groups, then apply these group averages to specific players. So, for example, a model might take into account age, recent production, benchmark production (how early a player started producing), draft status, etc. and based on weighted averages, it could give us a value of 350 VBD for Adrian Peterson. That's not a prediction for what Adrian Peterson is actually going to do. Rather, it's a way of pricing Adrian Peterson.Like in the dice game example from post #2, $3.50 isn't a prediction of what a specific roll of the die will pay out. Rather, it's a way of pricing the game properly. In the same way, 350 VBD isn't a prediction of what Adrian Peterson will produce. Rather, it's a way of pricing him properly.If we had a perfect model, we could price every player accurately and nobody could get the better of us in a trade (at least from a remaining VBD standpoint). Obviously we don't have such a model. But, as I tried to argue in posts 2 and 4, we don't need a perfect model. We only need to use a better model for pricing risk than the other guy is using.If life insurance companies A and B are selling policies, the one with the better actuaries will be able to set prices more accurately and will kill off the other company in short order. They'll exploit inefficiencies in the other company's model and win out in the long run. It's the same idea here. The idea isn't to accurately predict when some specific person is going to die. Rather, the idea is to set a price that accurately expresses the aggregate life expectancy for large groups of people and then assess a specific person based on those broad trends.

 

[Aabye]

The closest thing I know of to a complete statistical model of fantasy value is Football Outsiders' KUBIAK projection system for redraft leagues. It's designed to predict next year's fantasy points only, and is based heavily on projections for the team's performance (they also have to use their judgment to assign the values for some variables, like players' roles). A complete dynasty model might need to be a mixture of one projection system like KUBIAK for short-term value (the next year or two) and another projection system for long-term value.

Another approach, instead of trying to make a complete statistical model from scratch, is to start with conventional wisdom (based on ADP in startup drafts, or other people's rankings, or trade values) and then to analyze data to find ways to improve on those consensus rankings. You could start with just a small number of other variables (in addition to a player's consensus ranking), like their position, years of experience, and where they were picked in the NFL draft, and see if you can predict career VBD better by including those variables than by just going with the consensus rankings. Every time you find a variable that helps predict career VBD (better than you could with consensus rankings alone), you've found an error in the consensus rankings that you can take advantage of - certain types of players are being undervalued or overvalued by other owners.

That way you wouldn't need to build a full model from scratch all at once - with the consensus rankings as your skeleton you could gradually add more variables and build a better and better model. Over time more and more of your model will be based on objective variables and it will depend less on the consensus rankings, and maybe eventually you'll eventually get to the point where your objective variables allow you to leave out the consensus rankings altogether, in which case you'll have found a complete statistical model. But even if you don't ever get there, the model can still do a great job of estimating value and of identifying undervalued and overvalued players, which is what you can actually use as a fantasy owner.

One possible downside of this approach is that you'd need the data from old fantasy drafts in order to be able to make it work, so you'd have to get that data and you could only go back as far as you have ADP data. Another potential problem is that the fantasy conventional wisdom may have changed over time, so you might just find flaws in fantasy owners' previous views (e.g. overvaluing RBs relative to WRs) which are no longer present in their current ratings. But the advantages could outweigh that - you'd be able to start out with rankings that already incorporate lots of hard-to-quantify information, so instead of having to build everything from scratch you'd just be looking for specific variables that let you improve on what other people are already doing. Plus, you'd have a great chance of doing better than other owners at valuing players, since the approach is to start with other owners' valuations and then find ways to make them more accurate.

This is a really interesting suggestion. One additional worry:We don't know that the goal of conventional dynasty rankings is to rank players according to remaining VBD value. So, for instance, my rankings might indicate that Percy Harvin has a fair price of 150 VBD points and Reggie Wayne has a fair price of 75 VBD points, so a Harvin is worth 2 Waynes. Well, that may be a fair way of pricing the two players in terms of objective VBD points remaining, but it leaves out factors that have gone into the conventional dynasty rankings (e.g. Harvin's production will likely be spread out over a larger number of years, Wayne's more of a sure thing, etc.) So even if we point out "inefficiencies" in terms of how conventional dynasty rankings rank certain players, that still may not really be a point in favor of the values that the model assigns.

To put the point in terms that I used earlier:

The model would be able to assign an objective value to a player. But dynasty rankings might be based on both objective value and certain subjective factors (it's assumed that you want production now rather than 4 years from now, it's assumed that you want reliability over high risk/reward, etc.), and so it might be unfair to critique dynasty rankings for not mapping as well onto objective player values.

 

[Cato]

Here's an idea. What if, instead of trying to figuring out Ryan Mathews value, then Felix Jones' value, etc. we work it from the other direction? Define the historic characteristics of a player with 1175-1200 fantasy points remaining in his career, the characteristics of a player with 1150-1175 FP remaining in his career, etc. Prolem is, you'd need to go back 15 years before you're able to do that even in the ballpark of a meaningful sample size. Every year when the season is completed, you answer a set of questions for every player within their tab: FP scored, injury occured, games missed, misdemeanor arrest, felony arrest, etc. It could be as easy as clicking a radio button next to anything that applies. Instead of building a database to fetch remaining value based on player characteristics we build it to find players with characteristics consistant with 900-925 carer FP remaining, etc.

How do you determine the remaining FPs though - don't you need the Aabye method for that?

No. You do it by sampling players whose careers have ended.

how do you determine whether a 25 year old back is one with 900 or 1200 FP left?

I agree with this point, I do like the idea of building a model and working it to see the best buy low, sell high moments. But if you can't predict future value, your success will vary more based on the predictiveness of future value, rather than modelling based on historical results. I suppose building empirical trends based on attributes may help assess future value, but that seems too high level clinical to sort current individual players. There's too many variables to control to extrapolate, unless it's done in a very generic sense... (unless I'm thinking about this wrong.)

I'm not sure if this speaks to your concern, but I'm somewhat uneasy with the emphasis on "predicting" a player's value.I think the goal of predicting a player's future value is pretty much doomed. So much is random: injuries, situation, trades, retirement, coaching changes, etc. What I think we can do is generate pretty good data on average remaining fantasy values for groups. So, for instance, we can say something about the average remaining fantasy value of 25-year-old RBs who scored more than 100 VBD points the previous year.My idea is that we can analyze all sorts of groups, then apply these group averages to specific players. So, for example, a model might take into account age, recent production, benchmark production (how early a player started producing), draft status, etc. and based on weighted averages, it could give us a value of 350 VBD for Adrian Peterson. That's not a prediction for what Adrian Peterson is actually going to do. Rather, it's a way of pricing Adrian Peterson.Like in the dice game example from post #2, $3.50 isn't a prediction of what a specific roll of the die will pay out. Rather, it's a way of pricing the game properly. In the same way, 350 VBD isn't a prediction of what Adrian Peterson will produce. Rather, it's a way of pricing him properly.If we had a perfect model, we could price every player accurately and nobody could get the better of us in a trade (at least from a remaining VBD standpoint). Obviously we don't have such a model. But, as I tried to argue in posts 2 and 4, we don't need a perfect model. We only need to use a better model for pricing risk than the other guy is using.If life insurance companies A and B are selling policies, the one with the better actuaries will be able to set prices more accurately and will kill off the other company in short order. They'll exploit inefficiencies in the other company's model and win out in the long run. It's the same idea here. The idea isn't to accurately predict when some specific person is going to die. Rather, the idea is to set a price that accurately expresses the aggregate life expectancy for large groups of people and then assess a specific person based on those broad trends.

Appreciate the response, to be fair, I may need to re-read the entire thread and think on it some more. I know that last year before the season, I saw Peyton Hillis as a good buy low. Cutler had been chased out of Denver, like Marshall, like Tony Scheffler, I felt or knew it was only a matter of time before Hillis was also sent packing. (You're not a McDaniel's guy get out of here.) From his performance in 2008, I knew or felt given the shot he could produce. So I traded Snelling, for Hillis and a 6th rounder. (Of course I then flipped Hillis in a trade 3 weeks later, which sent Jared Allen my way.) Regardless, given how individual the set of circumstances were on this buy-low scenario, would one have any prior data points that would assist in modeling this? Of course maybe the model would be conceived for more average scenarios where the sample size would be significant enough to extrapolate meaning. Though given the amount of turn around in the rankings, (Hi Arian Foster!). I don't know if we're actually dealing with a great group of exceptions rather than groups which can be generalized. In other words the players are so individuals one can't model them or their current values from prior data sets. I don't know let me re-read the thread, as I may not be approaching the exercise from the right frame of mind. I'm seeing extreme individuals rather than general groups. I'm going to start from the top again.

 

[Aabye]

Appreciate the response, to be fair, I may need to re-read the entire thread and think on it some more. I know that last year before the season, I saw Peyton Hillis as a good buy low. Cutler had been chased out of Denver, like Marshall, like Tony Scheffler, I felt or knew it was only a matter of time before Hillis was also sent packing. (You're not a McDaniel's guy get out of here.) From his performance in 2008, I knew or felt given the shot he could produce. So I traded Snelling, for Hillis and a 6th rounder. (Of course I then flipped Hillis in a trade 3 weeks later, which sent Jared Allen my way.) Regardless, given how individual the set of circumstances were on this buy-low scenario, would one have any prior data points that would assist in modeling this? Of course maybe the model would be conceived for more average scenarios where the sample size would be significant enough to extrapolate meaning. Though given the amount of turn around in the rankings, (Hi Arian Foster!). I don't know if we're actually dealing with a great group of exceptions rather than groups which can be generalized. In other words the players are so individuals one can't model them or their current values from prior data sets. I don't know let me re-read the thread, as I may not be approaching the exercise from the right frame of mind. I'm seeing extreme individuals rather than general groups. I'm going to start from the top again.

My sense is that a model based on group expectations will probably be a bad way of finding breakout candidates. I'd imagine that the model would say something like "6th round RBs have an X% chance of achieving fantasy relevance", "backup RBs who average >4.5 ypc over X number of carries have an X% chance of achieving fantasy relevance", etc. It will be pretty good at telling you how likely a guy is to break out, but will be no good at all at predicting which of those guys will break out. It seems like what you're talking about with Hillis is much more of an individual scouting profile than the information that can be gleaned from a model.As for the other worry - extreme players vs. general groups: I think your intuition is pretty much right on. For instance, the group of 1st round fantasy RBs doesn't have a lot of guys that all scored about 200 career VBD points. Rather, some guys score a huge number of VBD points (Emmitt Smith scored more than 1,300), a lot of guys never score a single VBD point, and plenty of guys fall somewhere in the middle. But it's not as if the data is distributed in some normal way. What you get are a few huge successes, some moderately good players, and a whole lot of busts. The problem is that when you just look at some 1st round RB (Mark Ingram, for instance) you really don't know which group he'll fall into. Maybe if you have a great scout's eye you can figure it out but my general belief is that we armchair aficionados aren't all that great at figuring out which players will bust and which will boom.When you are confronted with Mark Ingram and you want to know how much you should be willing to pay for him, a good model can tell you the average career VBD value of a number of prospects that were similar in terms of quantifiable factors that are correlated with fantasy value (here the obvious correlation is draft position). That number will take into account the possibilites that Ingram booms, busts, or is a moderate success and will spit out a number that factors in each of those eventualities to come up with a sensible expectation.So think of it like this: If I sell 100 lottery tickets and there are three prizes ($100, $60, and $40) then what should you be willing to pay for a ticket? The answer is obviously $2 but the actual values of the tickets are extremes. Three tickets are worth way more than that and 97 tickets are completely worthless. Still, the model which assigns a per-ticket value of $2 is the best way to decide how much you should be willing to pay for a ticket, even though the actual values of the tickets (like the actual values of fantasy players) are all over the place.

 

[ZWK]

Here's another idea: use a hybrid model, which has one model for predicting short-term value (one or two years) and another model for predicting long-term value (the rest of the player's career).

Short-term value is heavily dependent on role and situation, identifying which players have an opportunity to break out, and there are already lots of efforts (including KUBIAK) to predict players' values for the upcoming season. You don't want to completely ignore that information about next season which redraft models use, since leaving out that information will make the model worse at predicting next year's performances, which makes it worse overall. But a lot of those temporary short-term factors aren't relevant once you look out past the next one or two years, so there will basically need to be a separate model to predict long-term value.

 

[FantasyTrader]

Whew... :whoosh: I'm beginning to see why this has never been done. The number of directions you can take this in is staggering. It leads to paralysis by analysis. That's why I don't think there's a shot in he** of doing this with Excel. You need a framework where you can start out very simple and VERY scalable. Then down the line you introduce one layer (ranking factor) at a time and build from there.

 

[Cato]

Appreciate the response, to be fair, I may need to re-read the entire thread and think on it some more. I know that last year before the season, I saw Peyton Hillis as a good buy low. Cutler had been chased out of Denver, like Marshall, like Tony Scheffler, I felt or knew it was only a matter of time before Hillis was also sent packing. (You're not a McDaniel's guy get out of here.) From his performance in 2008, I knew or felt given the shot he could produce. So I traded Snelling, for Hillis and a 6th rounder. (Of course I then flipped Hillis in a trade 3 weeks later, which sent Jared Allen my way.) Regardless, given how individual the set of circumstances were on this buy-low scenario, would one have any prior data points that would assist in modeling this? Of course maybe the model would be conceived for more average scenarios where the sample size would be significant enough to extrapolate meaning. Though given the amount of turn around in the rankings, (Hi Arian Foster!). I don't know if we're actually dealing with a great group of exceptions rather than groups which can be generalized. In other words the players are so individuals one can't model them or their current values from prior data sets. I don't know let me re-read the thread, as I may not be approaching the exercise from the right frame of mind. I'm seeing extreme individuals rather than general groups. I'm going to start from the top again.

My sense is that a model based on group expectations will probably be a bad way of finding breakout candidates. I'd imagine that the model would say something like "6th round RBs have an X% chance of achieving fantasy relevance", "backup RBs who average >4.5 ypc over X number of carries have an X% chance of achieving fantasy relevance", etc. It will be pretty good at telling you how likely a guy is to break out, but will be no good at all at predicting which of those guys will break out. It seems like what you're talking about with Hillis is much more of an individual scouting profile than the information that can be gleaned from a model.As for the other worry - extreme players vs. general groups: I think your intuition is pretty much right on. For instance, the group of 1st round fantasy RBs doesn't have a lot of guys that all scored about 200 career VBD points. Rather, some guys score a huge number of VBD points (Emmitt Smith scored more than 1,300), a lot of guys never score a single VBD point, and plenty of guys fall somewhere in the middle. But it's not as if the data is distributed in some normal way. What you get are a few huge successes, some moderately good players, and a whole lot of busts. The problem is that when you just look at some 1st round RB (Mark Ingram, for instance) you really don't know which group he'll fall into. Maybe if you have a great scout's eye you can figure it out but my general belief is that we armchair aficionados aren't all that great at figuring out which players will bust and which will boom.When you are confronted with Mark Ingram and you want to know how much you should be willing to pay for him, a good model can tell you the average career VBD value of a number of prospects that were similar in terms of quantifiable factors that are correlated with fantasy value (here the obvious correlation is draft position). That number will take into account the possibilites that Ingram booms, busts, or is a moderate success and will spit out a number that factors in each of those eventualities to come up with a sensible expectation.So think of it like this: If I sell 100 lottery tickets and there are three prizes ($100, $60, and $40) then what should you be willing to pay for a ticket? The answer is obviously $2 but the actual values of the tickets are extremes. Three tickets are worth way more than that and 97 tickets are completely worthless. Still, the model which assigns a per-ticket value of $2 is the best way to decide how much you should be willing to pay for a ticket, even though the actual values of the tickets (like the actual values of fantasy players) are all over the place.

This makes sense, we're pricing lottery tickets. Not saying which will hit. I still think we may see disagreement on the price of the lottery tickets though. (Given the expectations of pay out, or confidence in the ticket paying off.)Last year I flipped Reggie Bush for Randy Moss. I know buying 33 year old dynasty assets is not the wise course of action. But still I felt I would rather have 2 seasons, of 100 VBD, 50 VBD from Randy Moss, than 5 years of 25 VBD from Reggie Bush. ie Actually having a WR asset I could use albeit for a shorter amount of time. This didn't work out obviously, though I don't feel I really lost anything. So we're dealing with the process/pricing side of buying low, selling high, not actually the results. I suppose if one keeps buying stock at the right price, eventually you'll hit enough times. Good food for thought, just trying to wrap my head around buying players without being married to the results. I guess I've already done this, sent Boldin+Reggie Bush for CJ Spiller. I have no idea, if the Spiller lottery ticket will pay off, but felt ok as I don't value Reggie Bush and feel Boldin, who I love, is more of an RB than a WR (loosely.) and now is the time to sell, if you can find a buyer.Ok, I'll stop trying to embrace things from a Heisenberg-ian uncertainty perspective. Trends must exist, despite outliers.Hats off to the new ideas, and thought processes, we need more of this.

 

[jmo87usc]

I can't say I understand all of this, but one thing I took from it was to try and buy Eddie Royal low. Got him for a late 3rd rounder, so thanks!

 

[Aabye]

Here's another idea: use a hybrid model, which has one model for predicting short-term value (one or two years) and another model for predicting long-term value (the rest of the player's career).Short-term value is heavily dependent on role and situation, identifying which players have an opportunity to break out, and there are already lots of efforts (including KUBIAK) to predict players' values for the upcoming season. You don't want to completely ignore that information about next season which redraft models use, since leaving out that information will make the model worse at predicting next year's performances, which makes it worse overall. But a lot of those temporary short-term factors aren't relevant once you look out past the next one or two years, so there will basically need to be a separate model to predict long-term value.

This is a really good idea too.

 

[Aabye]

Whew... :whoosh: I'm beginning to see why this has never been done. The number of directions you can take this in is staggering. It leads to paralysis by analysis. That's why I don't think there's a shot in he** of doing this with Excel. You need a framework where you can start out very simple and VERY scalable. Then down the line you introduce one layer (ranking factor) at a time and build from there.

Yeah, my thoughts exactly. I've done my stuff in excel, which is probably not the best or fastest way to proceed. But I don't know anything about database software, so I'm sorta stuck.The cool thing is that it does look like it could be a collaborative project. We probably need to agree on some parameters (e.g. how to restrict the set of retired players we'll look at, what sort of a model we ultimately want to construct, etc.) but a lot of the work could be done independently and then compiled.Sounds pretty cool.