FSO utilizes the most sophisticated processes available to analyze, manage, and visualize fantasy football related data based on customized preferences and league specific rules. FSO provides five types of statistical discourse.
Exploratory Data Mining
Statistical analyses produce trends whether or not they are relevant, relational, useful, or trivial. Exploratory data mining is the process of sorting through large amounts of information and analyzing it from several different angles. The challenge, and the opportunity, is to discover which perspective offers material insight into the underlying themes of any patterns. Consider different time intervals for instance. It is possible to analyze player statistics per season (as do most fantasy sports web sites) or by first and second halves of one or more seasons. By looking at more than just one interval, FSO captures trends that might otherwise be missed such as knowing who is a slow starter (and, therefore, might be a good candidate for a waiver wire pick-up or trade) and who is an effective player early but may fade due to their age or proneness for injury.
Some factors are important but cannot be analyzed due to a lack of history. As such, FSO constructed its processes to encourage you to account for unquantifiable influences like:
- New coaches and coordinators
- Ownership’s influence
- Contract negotiations
Predictive Data Modeling
Many patterns that are discernable, whether identified by data mining or any other statistical technique, are due to serial events that may or may not repeat themselves. Often cited examples are found in a variety of sport myths (such as the Sports Illustrated cover-page jinx) that seem to be predictive indicators because patterns do indeed exist when evaluating them after the fact. So, in order to be correctly identified as predictive, data must have collaborating evidence of ancillary trends as substantiated through a number of statistical techniques. The key to successful predictive data modeling is to have an open architecture system, like FSO, that is predicated on dynamic (automatic) model selection. FSO’s system may consider up to 13 separate models to select the most predictive ones.
Some examples of less obvious trends that FSO data mining may capture include:
- Learning curves by position (QB, RB, and WR)
- Changes in coaching philosophy
- Strategic utilization of positional players (e.g., RB-by-Committee)
- Injury proneness and rehabilitation times
- Player age and its impact on performance
Linear Programming (LP)
FSO uses a mathematical concept called linear programming to account for constrained variables. In order to do so, FSO converts one or more objectives into mathematical form. These algorithms function to seek a minimum or maximum optimal solution(s) based on a series of linear equations set for equality or inequality. The geometrical application of these equations results in a feasible range of solutions from which an optimal solution can be sought. For example, FSO utilizes LP to account for various league-specific rules such as the number of players allowed by position as well as certain proprietary imperatives. An example of the later is that since FSO assumes to win your fantasy league, you’ll likely need to take more risk than the median team (i.e., it’s not about being average - it’s about winning).
Large Scale Optimization and Sequencing
The NFL is comprised of 32 teams, approximately fifteen hundred players, and millions of statistics. In fantasy sports, each league imposes a variety of constraints by player, position, or team. Without going into the mathematics, a resulting optimization using a traditional mean-variance approach may easily include more than 20 million calculations, qualifying it as a “large scale” optimization problem.
Even five years ago, it was very difficult to solve large scale optimization problems because of technology constraints and because of the potential for estimation error due to issues such as insignificant alphas, asymmetric distributions, multi-collinearity, and homogeneity of inputs. Now, however, many disciplines benefit from recent advancements in large scale processing such as finance, biomedical, aerospace, and engineering. One such advancement is sequential quadratic programming. FSO utilizes a similar technique based on the unique characteristics of fantasy sports so as to calculate successive algorithms in such a way so as to always improve towards an optimal solution.
Dynamic Stochastic programming
A live draft is subject to two types of variables: known and unknown. Prior draft picks are known variables. Future draft picks are unknown variables. This is a partially observable framework. There are two (primary) ways to determine possible future outcomes: 1) not using prior information and only looking forward at each step (often described as a Markov Decision Process) or, 2) using prior information to determine probabilities of future actions based on changing properties and preferences (Dynamic Stochastic Programming). As part of its Live Cheat Sheet, FSO uses the later to determine the best position to draft at a given point in a sequence of decisions.
FSO makes every effort to monitor categories that might be used in your league’s scoring system. In total, FSO currently tracks 26 statistics by team, squad, player and/or opponent (see list below). Many of these include custom ranges based on yardage, points, etc.
- 1st downs (by type)
- 2 pointers (allowed, by type, tried, made)
- 3rd downs (attempts, conversions)
- 4th downs (attempts, conversions)
- Assisted tackles
- Defense (passes defensed, QB hurries)
- Field goals (allowed, attempted, made, missed, blocked, return yds, distance, total percentage)
- Game (won, lost, margin of score)
- Fumbles (forced, lost, yds lost, recoveries, TDs)
- QB Rating (player, team)
- Interceptions (made, yds lost, TDs)
- Kickoff returns (return yds, TDs, fair catch, end zone, touchback)
- Passing (allowed, attempts, completions, yds, total percentage, TDs)
- PAT (allowed, attempted, made, missed, blocked)
- Penalties (yds)
- Points (by quarter, defense, special teams, offense, opponent, allowed, scored, out/underscored)
- Punt returns (yds, TDs, fair catch)
- Punts (made, return yds, blocked, inside 20, touchbacks)
- Receiving (allowed, attempts, made, yds, percentages, TDs)
- Rotisserie (avg rushing yds (per carry), avg receiving yds (per reception), avg passing yds (per attempt)
- Rushing (allowed, carries, yds, percentages, TDs)
- Sacks (made, yds lost)
- Tackles (made, yds loss)
- TDs (allowed, made, distance)
- Time of possession
- Yards (total (net), total all purpose, allowed)
Bold indicates fully customizable features as currently available in The Optimizer.
FSO will attempt to adjust for any categories that might be missing from its database but that are part of your scoring system. If the missing information negatively compromises expected results, as determined through statistical discourse, FSO will inform you of its inability to conduct the analysis and will refund 100% of your money.
Prior to running its models, FSO filters the data to account for sparse data, invariant variables, outliers, and missing or otherwise spurious or invalid data. FSO uses only statistically significant data, meaning the results of its analyses are likely valid. There are a number of ways to measure statistical significance but all depend on having a sufficient number of observations. Common areas where there is a lack of information are rookies and those players who have been chronically injured. As such, FSO does not include rookies as part of its feasible data set. That being said, for those utilizing The Optimizer, you may still select rookies (or any other players) that you designate. FSO will simply analyze your team after adjusting budget and position constraints in accordance with specified player designations.