The method, for analysts and hiring teams

I'd like to work on recruitment models and the visual side of how clubs read them. In October 2026 I rebuilt my whole event-data pipeline after finding bugs in the coordinate frame, the match clock and the VAEP formula, and re-ran everything on top of it. My transfer-corridor piece retracted one of its own March claims. In November 2026 I tested every model again: action value is now a sequence model that reads the last 32 actions, expected goals lost its calibration layer, expected passes and the price model changed learner, and the published numbers were re-run. That's how I'd work on your data too.

• Transfer success by league

  Six seasons of moves against an intra-league control group: which leagues' players hold their level, and why most failures are about minutes.

  Read September 2026; the re-run is out in December

• Fair transfer price

  Worth learned from market values, then calibrated on 3,226 real fees: selling league, age, window, football inflation, and the one buyer that changes a fee most, a Premier League club. Two prices per player and their mix. Backtested window by window.

  Read Use the app

• Squad needs

  What a squad is short of, read from its own numbers: a quality that lags what a paired quality implies, checked against teams that play the same way and kept only if it repeats.

  Read October 2026

• Position clustering

  Roles from where and how players act, not from team sheets: the role behind every comparison I make.

  Read May 2025: an earlier version, the method has changed since

• Team style similarity

  Which teams play like yours, from possession and pressing maps, so needs are measured against real peers.

  Read March 2025: an earlier version, the method has changed since

• Dixon-Coles ratings

  Team strength from my own xG model plus the threat a team builds before the shot, comparable across leagues: how I translate a player's numbers from one league to another.

  Read March 2025: an earlier version, the method has changed since

• Defensive value

  What a defender's zone concedes, from convex hulls and VAEP. I tested whether it rates the defender or his team: it mostly follows the team and doesn't carry to a new club, so I read it as context, not as a rating. Event data struggles with defending, and I say where.

• Standstill take-ons

  Two proxies for beating a man from a standing start. The simple one was the more repeatable.

  Read February 2026, before the rebuild

• Explosive carries

  An acceleration proxy from event data, compared against Gradient's tracking data: the La Liga outliers largely match.

  Read February 2026, before the rebuild

• Passing skill

  Two passing skills: security (a Bayesian model of how much more than expected a player completes) and value (what the passes he picks are worth, the one that predicts next season). Shooters, keepers and duels in the same framework.

• Tracking data

  Peak velocity, and why one number isn't enough; off-ball runs three seconds before the pass. Both on Gradient's free World Cup 2022 data.

  Peak velocity Off-ball runs March and April 2026