Separating Genuine Improvement From Regression Bait

## The Regression Trap Not every outperforming team is a regression target. Some teams genuinely improve — through a major transfer, a tactical innovation, or a manager change. Confusing genuine improvement with regression bait is a costly mistake. ## The Signals of Genuine Improvement **xG rate change, not just results change:** If a team's xG for AND against improves — they are creating better chances and conceding fewer quality chances — this is structural improvement, not just results luck. Regression is less expected. **Process-level improvement:** High pressing metrics, improved defensive shape indicators (passes allowed into the box), forward runs per match — these process metrics indicate genuine tactical improvement that underlies the results. **Consistent vs intermittent performance:** A team that outperforms for 3 matches with inconsistent underlying metrics is more likely regression bait. A team that consistently produces improved xG across 10 matches with clean process data is more likely to have genuinely improved. ## The Manager Change Signal A new manager typically produces a sharp initial performance change in either direction — both genuine (different system) and psychological (players reset effort for new boss). The first 5–8 matches under a new manager are highly volatile and regression-prone in both directions. The 8–20 match range is more informative about whether the change was genuinely positive. ## The Model Update Protocol When a significant genuine improvement is identified (new manager system, key transfer): update your model explicitly. The prior historical rating for this team is now partially obsolete. Bayesian approach: keep 40–60% weight on the old rating, 40–60% on the new performance. As new data accumulates, progressively shift weight toward the new performance. Do not fully discard the historical rating — it remains informative. Do not ignore the improvement — it is real. Blend them proportionally based on sample size of the new data.