The T-Score Model

Introduction

I stumbled upon a random article on Reddit, where the author was using WR stats from a previous season to try and predict fantasy success for Wide Receivers (WR) the following season. He had created his own “trinity” score utilizing Target (Tgt) Share, Air Yards (AY) Share, and Yards After Catch (YAC) per reception, and normalized them. Normalizing means taking the stats of a particular metric, and putting it all on a 0-100 (or 0-1) scale, in order to make more fair comparisons. He also looked into alternative combinations of metrics, and the one where he had found the most correlation used total receiving yards, YAC, AY Share, and Yards per Team Pass Attempt, and normalized each stat.

At the time I read this particular article, he didn’t have anything using the stats from 2024 to try and predict 2025 fantasy success, so I decided to take it on myself. I then decided to adjust it, because some of those stats hurt WRs who got injured. I know that decreases sample sizes in some cases, but I decided nonetheless to adjust each stat for "per game". I then combined those normalized stats into a sum to give each WR a score, sorted the data by that score, and ranked them. I found the results for the top-12 interesting.

I proceeded to do some manual testing (with assistance from AI), and my model that I landed on for awhile that I dubbed “T-Score” (which I will dub “v1.4”) used the following metrics (in order of weight in v1.4):

• Target Rate/game
• Yards/game
• YAC/game
• Yards per Team Pass Attempts/game
• Air Yards Share/game

Note: Coincidentally, I found out that “t-scoring” is a common method used in data analytics to standardize raw data points. It is completely by accident that I decided to call my model “T-Score” based on my last name. Fun accident, though.

Since then, though, I’ve done a ton of learning and reading on data analysis, and have greatly adjusted my model. I also learned that I needed to apply qualifications in order for WR to be ranked; I eventually landed on a minimum of both 8 games played in a season, and 15 receptions in a season (this might seem low, but if a WR played only 8 games, that comes out to just under 2 receptions a game; surprisingly, if I raised the minimum much more, I was missing what I felt to be somewhat relevant players in deep leagues).

The Goal: Predicting N+1 PPG

My primary objective has been to build a robust predictive model for wide receiver performance in the subsequent season (N+1), measured by Points Per Game (PPG). I utilized T-scoring (the data analytics method) to standardize different metrics, allowing for fair comparison and combination, and Multiple Linear Regression (MLR) to determine the optimal weights for each metric in predicting N+1 PPG.

MLR is a statistical technique used to model the relationship between a single dependent variable and two or more independent variables. The core idea of MLR is to find a linear equation that best describes how the independent variables influence or predict the value of the dependent variable. For my math nerds out there, the general form of a Multiple Linear Regression equation is:

Now, the problem with MLR and football stats is that MLR makes the assumption that the relationship between the dependent variable and each independent variable (and their combination) is linear. Well, we all know that’s just not the case. Too many things such as QB play, offensive scheme, O-Line strength, and so much more causes football to be a highly variable sport. The good news is, of the main offensive positions used in fantasy football, WR is probably the closest to have a more linear relationship than the other positions.

This makes some sense; for example, while a WR is impacted by a poor O-Line, WRs are not as directly impacted as a Running Back. Once the ball is in a WR’s hands, they are more than likely well beyond the O-Line. That’s just one example, but the point I’m trying to make is that the variables that impact a WR’s play tend to be more straightforward than the other positions. If a WR is targeted a lot, more than likely they will produce a lot.

So, while MLR is not the end-all be-all in figuring out how closely particular metrics relate to PPG, it is helpful (and interesting) for WR data. Once I started applying Multiple Linear Regression analysis with the metrics I had landed on, I learned some very interesting information. Reminder, I first standardized the metrics before evaluating them with N+1 PPG.

Metrics Explored and Their Relationship to N+1 PPG:

• Target Rate Per game (Tgt Rt/G): “Volume is king.” A mantra repeated often in fantasy football circles. Well, this is true; when you complete a Single Linear Regression analysis of Tgt Rt/G and N+1 PPG, the relationship is strong, producing a strong positive coefficient (0.303).
  
  Note: A Coefficient is the number that tells us how important a particular metric is in predicting N+1 PPG. The bigger the number, the stronger the relationship.
  
  However, in terms of my model, Tgt Rt/G actually has a negative Coefficient when completing MLR alongside the other four metrics in v1.4 of T-Score. And its P-Value was larger than any of the others in T-Score v1.4!
  
  Note: A P-Value is a number that tells us how likely a Y-variable’s (i.e., Tgt Rt/G) relationship to the X-variable (N+1 PPG) is just “lucky”. The higher the P-value, the more likely a metric looks good because of luck. I’d like to be clear that I do not have a degree in Mathematics or Statistics; so if I say something wrong…
  
  So, with v2.0, I gave Tgt Rt/G the boot.


• Yards Per Game (Yds/G): Relationship to N+1 PPG: This metric consistently proved to be a highly significant predictor of future PPG. It received a strong positive coefficient (around 0.183 in my latest version, v2.1). This indicates that a receiver's yards per game in the current season is a fundamental driver of their fantasy production in the next, signifying their overall receiving volume and role in the offense.
  
  You might be thinking, “Duh, yards are what make up a good chunk of a WR’s points in fantasy.” Well, you’ll be surprised to learn later that just because it is a part of fantasy points, it does not necessarily have a strong positive coefficient in predicting future fantasy success. Anyways, Yds/G is the production metric anchor in T-Score v2.1.


• Yards After Contact Per Game (YAC/G) and YAC Per Reception (YAC/R): You might be thinking “Why did you use YAC/G instead of YAC/R?” My answer is: I don’t know. When I was initially putting this together, in my head it made a lot of sense. It wasn’t until later that I realized YAC/R makes more sense, as it still works for players who missed games due to injury, but is a closer look at a player’s efficiency once they catch the ball, as opposed to another way at expressing the player’s opportunity/total production (essentially, a good YAC/G can be explained by a good Yds/G).
  
  Anyways, I eventually phased YAC/G out and subbed YAC/R into my model (around v2.0, which is when I approached all of my analysis in a pretty much new way than before). Well, whether by itself, or in MLR, YAC/R does not have a good coefficient to N+1 PPG. This might explain why Deebo Samuel has been a hard player to predict, injuries notwithstanding (and has me a little worried about Rashee Rice…). By itself, YAC/R has a coefficient of 0.036 (ideally we want a coefficient above 0.05) and a p-value of 0.165 (we want this to be below 0.05, or 5%). Within MLR, its coefficient was worse (negative, in fact). So, along with Tgt Rt/G, YAC/R got the boot in v2.0.


• Yards Per Team Pass Attempt Per Game (Y/Tm PA/G): Relationship to N+1 PPG: This metric, like Yds/G, also showed statistical significance and a positive coefficient (around 0.112 in v2.1). It accounts for a player's efficiency and target share relative to their team's overall passing volume. Its significance suggests that players who are efficient within their team's passing scheme are more likely to sustain or improve their PPG.
  
  While not the same, this essentially takes the place of Tgt Rt/G as the opportunity metric anchor in T-Score v2.1.


• Air Yards Share Per Game (AY Sh/G): Now, this is interesting (or at least, it is to me): AY Sh/G has an independently strong relationship to N+1 PPG. The reason I find it interesting is because AY Sh is, in a way, the talent-independent aspect of a WR’s actual production. Now of course, a WR had to be open in the first place to get the target, so don’t hear (or read?) what I’m not saying; I just mean, when compared directly to YAC, AY Sh says less about a WR’s playmaking ability with the ball in their hands. Who are the players you tend to think of when you hear “High AY Sh”? Now, who are the players you tend to think of when you year “High YAC”?
  
  Sorry, I tend to rabbit-trail a lot (I blame the ADHD).
  
  AY Sh/G’s independent coefficient is 0.312, with a 0 (or near-0) P-value (which is good!). However, when combined with the other metrics in MLR, it loses its value. Basically, Yds/G and Y/Tm PA/G provide the same (and better) insight than AY Sh/G.

So, of my original five metrics in T-Score v1.4, only two showed a strong positive coefficient to N+1 PPG. Basically, the other three metrics were just noise. To test this, I applied my T-Score v1.4 using data from 2021-2023 to “predict” success the following season (N+1), and calculated its Recall success rate. I then created a new model, T-Score v2.0, using only Yds/G and Y/Tm PA/G, and calculated its Recall success rate for those same seasons. I did this for several tiers of WRs to give myself the best glimpse as to how well these two scores did at predicting fantasy success. First tier was top-6 WRs, next was top-12, and so on until top-36 (I initially did top-48, but players beyond the top 36 had too much variance and/or inconsistency, so I found it was hurting my research more than helping).

What I found was that T-Score v1.4 and v2.0 performed almost identically to one another. V1.4 averaged a 78.04% success rate in the top-36 players, while v2.0 averaged 76.89%. On one hand, this was great - I had simplified my model! On the other hand, I wished I hadn’t wasted so much time setting up formulas for those other metrics…

However, I felt like relying solely on two metrics didn’t sit quite right. Call it curiosity, obsessiveness, or whatever you’d like; but I wanted at least one more metric to include in my T-Score to help it be more well-rounded.

I started with different versions of my original metrics; namely, Target Share (total) and Air Yard Share (total) as opposed to their per-game variants. More or less, these had the same results as their per-game variants. I then shifted to some new metrics, starting with Yards per Route Run (Y/RR). This proved to be a strong independent metric, but ended up having overlap with Yds/G and Y/Tm PA/G, so did not contribute meaningfully. I also tried Receptions per Game, but it held similar results.

I finally decided to try Red Zone Target Share (RZ Tgt Sh). I had, up until this point, avoided TD-related metrics, because they tend to have so much variance from season-to-season. But, I figured it was worth seeing if a player was targeted in the Red Zone often in one season, how much that related to their fantasy success the following season. In other words, if a player is involved in the Red Zone, they likely will be a part of a team’s Red Zone schemes in the future (players such as Mike Evans and DK Metcalf).

Well, I ran into a wrinkle; my data source (Pro Football Reference, or PFR) did not provide RZ Tgt Sh for a player who switched teams mid-season. The primary example was Davante Adams in 2024, who started with the Raiders and ended with the Jets. I tried finding another source that did provide this data, but I discovered that each source had its own subjectivity in what it considered “Red Zone Target Share”. I won’t bore you with the details here, but I ended up deciding to evaluate total Red Zone Targets. Not my preference, but my alternative was to manually calculate the RZ Tgt Sh myself, or to switch sources for all of my data (which I didn’t want to do).

• Red Zone Targets (RZ Tgt): Relationship to N+1 PPG: This was found to be statistically significant (with a P-value of 0.0051, or 0.51%) even after accounting for Yds/G and Y/Tm PA/G. It carried a positive coefficient (around 0.056). This confirms that a player's red zone usage is not fully captured by raw yardage or team-adjusted yardage, and it adds unique predictive value, likely signaling touchdown upside and high-leverage scoring opportunities.
  
  I should note here that the coefficient is not as big as it is for Yds/G and Y/Tm PA/G, but it still contributed positively. I actually like that, of the three metrics, RZ Tgt has the least weight.

Model Performance and Optimal Choice:

Now that I knew which metrics were best, I applied weights to each standardized metric to combine them into one Composite T-Score. I took the coefficient of each metric, summed them up, and then divided each metric by the sum to determine the weight.

The weights for v2.1 are as follows:

• Yds/G = 52.1%
• Y/Tm PA/G = 31.9%
• RZ Tgt = 16%

I proceeded to compare the predictive success rates of my initial 5-metric T-Score model, v1.4 (as the baseline), a 2-metric model v2.0 (Yds/G and Y/Tm PA/G), and a 3-metric model v2.1 (Yds/G, Y/Tm PA/G, and RZ Tgt).

Key Findings from Model Comparisons:

• Overall Strong Performance: All models demonstrated reasonable success rates in identifying players within various fantasy production tiers (top-6, top-12, top-18, etc.) across different prediction years (2022, 2023, 2024; I chose my data cutoff as 2021, since that was the first season the NFL went to an 18-game season).
• v2.0's Edge: While the 3-metric model often performed similarly to the 2-metric model (v.2.1), it showed significant improvements in key areas: Notably, in 2022 Top-6% success rate, it jumped from 66.7% to an impressive 83.3%. This indicates a superior ability to identify true elite-level performers in some seasons. It also showed modest improvements in some deeper tiers (e.g., 2024 Top-36%).
• Simpler is Often Better: The fact that v2.1 consistently performed on par with or better than v1.4 suggests that focusing on the most impactful and independently predictive metrics is more effective than simply adding more.

Conclusion:

My analysis strongly supports the 3-metric T-Score model (using Yds/G, Y/Tm PA/G, and RZ Tgt), v2.1 as the most effective and balanced approach for predicting N+1 PPG for WRs. Each of these three metrics has a statistically significant and positive relationship with future PPG, and their combination provides robust predictive power, particularly for identifying high-value players.

Once I found the metrics I liked, I placed the results on a 0-100 scale for easier cross-season comparison (ya’ll, 2021 Cooper Kupp was insane). I then grouped scores together into buckets based on percentile:

And thus, the T-Score WR Model v2.1 was complete.

Thanks for reading!