If this is your first visit, be sure to check out the FAQ by clicking the link above. This is an entirely free site so all we ask is that you create a membership in order to view messages and post. Register here to proceed. And welcome to The Powder Blues community of Charger Fans. We look forward to building this community together. Go Chargers.
Also there are clear discontinuities and misorderings in the consensus board. I propose this is due to using "mean" based averaging on players with left skewed distributions.
As a thought experiment, consider Player A who 16 NFL teams have a 3rd round grade on and 16 NFL teams have a 5th round grade on. Such a player has a good chance of going in the 3rd round and about a 5% chance of going in the 5th round.
Contrast with Player B who all 32 teams have a 4th round grade on. There is very little chance such a player goes in the 3rd, most likely the 4th, some chance in the 5th.
Both players have the same "mean" ranking yet very different chances of going before their mean value.
A way to partially correct for this using HS math is to take the median instead of the mean. This is well known even in the general population.
However, a better way to do this is to creat a probability distribution using Monte Carlo simulations. Without knowing actual team boards (except Dallas) we can't do this pefectly but we can create n:1 mappings of public boards onto teams and then randomly reassign to account for draft ordering and team needs. This would create the probability distribution mentioned earlier.
I'm pretty sure one of the analytics sites, maybe NGS, was doing this is the first round.
Now, to order players on the board you would take the median value of the simualtions and order them. This should provide a better ordering than current consensus methods of naive mean and close the gap somewhat.
However, to determine if a player is a reach once you have the probability distributions you would look for a player taken one or two standard deviations before their mean value. In such a case, Player B would likely be considered a reach in r3 but Player A would not as Player A has a much higher stdev.
-‐-----------------‐-‐-------------------------
tl;dr naive means don't work on players on whom the boards disagree so it can't be used to determine reaches. More work and skill is needed. Evidence: NFL draft outperforms consensus big board.
Further, the impact of reaches decreases as the draft goes on as the marginal delta between player values decreases, so this criterion should be scaled up as the draft progresses.
Also, distributions and stdevs will be tight early in the draft and expand as the draft progresses and pick numbers increase in orders of magnitude.
Tried to follow - but got lost around the 3rd or 4th paragraph. Can we just say that evaluating player is (extremely) difficult?? And the further from the upper tiers you get the more difficult it gets.:thumbsup:
Tried to follow - but got lost around the 3rd or 4th paragraph. Can we just say that evaluating player is (extremely) difficult?? And the further from the upper tiers you get the more difficult it gets.:thumbsup:
Instead of taking the average draft position where draftniks have a guy placed,
take the middle position of the collective draftnik grouping for a player.
ex: 1rst, 1rst, 4rth, 4rth, 5th.....the middle position is a 4rth round grade.
They would arrange it by overall player # ranking and not round #: #165th player on the board, for example
I'm not familiar with the Monte Carlo simulation: it was first conceived to handle problems of neutron scattering off of nuclei by atomic scientists and mathematicians. Suffice to say its complicated and can be extended to various deterministic problems involving the evolution of systems in many different fields.
For the draft (an evolving system), it takes into account the degrees of freedom of the system at any point in time (a GM's cluster for each round).
From a quick read, to develop a good predictive model it requires many samples - meaning many actual NFL drafts need to have been conducted to accurately determine where a player should go.....Chaincrusher's nirvana.
Monte Carlo simulations are commonly used for "just run millions of simulations and collate the results". They can be used to create probability distributions when no probability function is known.
Monte Carlo simulations are commonly used for "just run millions of simulations and collate the results". They can be used to create probability distributions when no probability function is known.
Yes but MC Sims are susceptible to the same input errors as any modeling problem. MC inputs are banks of possible outcomes or banks of possible inputs with a known algorithm to then calculate the outcome. Computer randomly selects numbers over and over thousands or millions of times from those data sets (as many times as the modeler may wish to wait for) and then a distribution is fit to the data set that is output.
If you decide every outcome is equally possible, e.g. every possible value has the same frequency in the MC input data banks, then you get the same answer for every player. ANY alternative to this “no information” baseline is based on some input assumptions - where do those come from - the same set of guys that created the consensus board to begin with?
Yes but MC Sims are susceptible to the same input errors as any modeling problem. MC inputs are banks of possible outcomes or banks of possible inputs with a known algorithm to then calculate the outcome. Computer randomly selects numbers over and over thousands or millions of times from those data sets (as many times as the modeler may wish to wait for) and then a distribution is fit to the data set that is output.
If you decide every outcome is equally possible, e.g. every possible value has the same frequency in the MC input data banks, then you get the same answer for every player. ANY alternative to this “no information” baseline is based on some input assumptions - where do those come from - the same set of guys that created the consensus board to begin with?
Holy cow!! Aren't there any construction guys on the board? Or do we need a PHD to participate / translate?:hadworse:
Yes but MC Sims are susceptible to the same input errors as any modeling problem. MC inputs are banks of possible outcomes or banks of possible inputs with a known algorithm to then calculate the outcome. Computer randomly selects numbers over and over thousands or millions of times from those data sets (as many times as the modeler may wish to wait for) and then a distribution is fit to the data set that is output.
If you decide every outcome is equally possible, e.g. every possible value has the same frequency in the MC input data banks, then you get the same answer for every player. ANY alternative to this “no information” baseline is based on some input assumptions - where do those come from - the same set of guys that created the consensus board to begin with?
I thought I was fairly clear on method. Establish team big boards based on averaging (somehow) of sets of public boards. I haven't figured exact methodology but in general, if 20% of boards have a 3rd rd grade on a player, 50% 4th round, 30% 5th round, the team boards should follow the same distribution.
Draft is simulated based on those big boards accounting for ordering and team needs. Then the boards are shuffled and it's run again, until you fill the whole space.
Comment