Reggie Mckiver The Architect Of Modern Sports Analytics And Strategic Player Evaluation

What This Framework Actually Does

Most people encounter Reggie Mckiver The Architect Of Modern Sports Analytics And Strategic Player Evaluation when they're trying to move beyond traditional box score stats and actually evaluate what players contribute in ways that impact winning. The approach is fundamentally about deconstructing player performance into quantifiable components that separate noise from signal. It works across basketball, football, and increasingly in soccer analytics, but the core methodology stays the same regardless of sport.

The basic workflow starts with data collection. You pull play-by-play data or tracking data depending on what your sport offers. Basketball people usually work with NBA play-by-play files or Second Spectrum tracking. Football folks deal with NFL Next Gen Stats or Hudl tracking feeds. Soccer analysts often use Opta or StatsBomb raw feeds. The format differences between these matter more than most people admit. I once spent three full days converting an Opta event file because the coordinate system used 105x68 meter grids while my visualization pipeline expected yards. The workaround was writing a quick normalization script that mapped the metric coordinates to the nearest yard equivalent using a 1.09361 conversion factor. Nothing complicated, just something nobody warns you about. The framework itself is organized around three layers: raw event processing, context adjustment, and projection modeling. The raw event layer takes discrete actions like a pass, shot, or tackle and assigns them initial values based on historical outcomes. A shot attempt inside the paint isn't just "a shot" — it gets weighted by location, defender proximity, and game state. The context adjustment layer then normalizes those raw values for league average pace, strength of competition, and situational factors. This is where most amateur models fail because they skip straight to projection without properly normalizing. I've seen analysts produce rankings that looked impressive until you adjusted for the fact that the player's team averaged 100 possessions while the league average was 92. The projection layer is where the actual evaluation happens. You build models that estimate what a player's future performance will look like based on their adjusted production, aging curves, and contextual durability factors. The counter-intuitive part most beginners miss is that shot quality metrics tend to regress faster than passing or defensive positioning metrics. If your model overweights shooting efficiency in year one, you'll systematically overvalue role players who catch and shoot rather than create their own shots. The workaround is applying a shrinkage factor to shot-based components — typically around 0.6 to 0.7 of the raw value — while keeping creation and positioning metrics closer to 0.85. This isn't arbitrary. It comes from observing that shot quality has higher year-over-year variance than either playmaking or off-ball defensive positioning.

Setting Up The Infrastructure

You don't need a PhD or a data science team to run this. A decent laptop, Python, and patience will get you most of the way there. The packages you actually need are pandas for data manipulation, numpy for numerical operations, and scikit-learn for the modeling piece. If you're working with tracking data specifically, you'll also want scipy for interpolation functions and matplotlib or seaborn for visualization. I use postgres for storing processed player seasons because querying flat files becomes painful once you have more than three years of data. A simple table with columns for player_id, season, league_context, adjusted_metrics, and projection_output handles everything you need. The data pipelines themselves are the hardest part. Every league structures their raw feeds differently. The NBA gives you clean play-by-play with player IDs already attached. The NFL requires you to match jersey numbers to player rosters manually for older seasons before 2016. Soccer data is the most fragmented because different providers use completely different event taxonomies. I maintain a mapping spreadsheet that translates Opta codes, StatsBomb codes, and Wyscout codes into a unified schema. It's been a living document for four years and it saves me roughly two days of work every off-season.

The Modeling Process Step By Step

Start by calculating raw rate stats per 100 possessions or per game depending on the sport. For basketball, per-100-possession metrics are standard because pace varies wildly between eras and teams. For football, per-game or per-touch metrics work better since possessions aren't as clearly defined. Next, apply league-average normalization. Subtract the league mean from each player's raw stat and divide by the league standard deviation. This gives you z-scores that are directly comparable across seasons. A 2.1 z-score in 2023 means the same thing as a 2.1 z-score in 2019. After normalization, you apply the context adjustment. This involves computing teammate and opponent filters. If a player's surrounding roster has an above-average shooting percentage, their own shooting numbers get slightly deflated because floor spacing effects are partially shared. Conversely, if a quarterback throws to receivers who consistently run routes against coverage, those passing yards deserve a small positive adjustment. The exact formulas vary by position. For basketball forwards, I use a simplified version of cleaning the glass metrics that weights rebounding opportunities by offensive glass rate and defensive board availability. For soccer wingers, I weight progressive carries and passes by the defensive line height of the opposing team. Projection modeling uses a combination of Bayesian hierarchical priors and regularized regression. The key insight is that you don't predict raw stats — you predict z-score movements. Historical data shows that a player's z-score in year N correlates with their z-score in year N+1 at roughly 0.45 for scoring metrics and 0.60 for playmaking metrics. The lower correlation for scoring reflects the higher variance and regression tendency I mentioned earlier. Your model should output a posterior distribution rather than a single point estimate. This means instead of saying a player will score 18.3 points per game, you say there's a 70% probability they'll score between 15.1 and 21.7. That range is what actually matters for decision-making.

Where This Framework Breaks Down

There are clear scenarios where this approach produces unreliable results. The biggest limitation is small sample sizes for position players in sports like baseball or soccer where individual touches are sparse. A soccer goalkeeper facing 15 shots in a season has noisy data that looks meaningful if you don't apply sufficient shrinkage. I learned this the hard way when I rated a goalkeeper as a top-5 performer globally based on six games of data. The model hadn't converged yet and I was reading too much into variance. The fix is applying a minimum appearance threshold — usually 20 games for soccer goalkeepers, 40 for position players, and 50 for basketball players — below which you collapse estimates toward the league mean. Another failure mode is positional misclassification. When a player switches positions mid-season, the model attributes their new position's context adjustments to their old profile unless you explicitly handle the switch. I encountered this with a basketball player who moved from small forward to center due to injuries. His rim protection numbers looked poor because the model was still comparing him against small forward norms for the first half of the season. The workaround is tracking position changes as separate observation periods and recalibrating context adjustments at the transition point. It adds complexity but prevents systematic bias. The final limitation is that this framework measures what happened, not what will happen under changed conditions. Trade scenarios, coaching changes, and system fit adjustments aren't captured by any statistical model. I've seen front offices waste millions projecting a player's production in a new system based purely on historical z-scores without accounting for tactical mismatch. The model can tell you what a player did in environment A. It cannot reliably tell you what they'll do in environment B without manual adjustment from someone who understands the sport tactically. That human layer is non-negotiable.

Getting Started

If you want to implement this, start with one league and one position group. Don't try to build a universal model. Pick basketball power forwards or soccer center backs and work through the full pipeline end to end. The first iteration will be slow and probably wrong. The second will be better. By the third season of data, you'll have something usable. The entire process from raw data to ranked player list typically takes 45 to 90 minutes depending on your hardware and whether you've automated the feed parsing. The first time through it took me about three hours because I was still debugging the normalization script. Now it runs overnight while I sleep and the rankings are on my desk by morning. Open source implementations exist for the underlying statistical pieces but none of them bundle the full Reggie Mckiver framework as a ready-to-use tool. Most of the methodology lives in custom scripts written by individual analysts. I share mine publicly on GitHub and update it after each season as the leagues release new data formats. The repo includes the data pipeline, the normalization functions, the position-specific context adjustment tables, and the projection model. It's not polished but it's functional and it's free.