Testing Manager Change Collapse: The Solver Handles Chaos Better Than Expected
Skip bets within 10 matches of a mid-season manager change? Marginal ROI = -0.1% — filter hurts. 877 changes loaded via Fotmob (529 mid-season). The solver reads Pinnacle odds which already price manager changes. Combined with the xG window test, manager changes are definitively priced. Case closed.
Testing Manager Change Collapse: The Solver Handles Chaos Better Than Expected
When a club sacks its manager mid-season, the prevailing wisdom says everything goes haywire. New tactical system, new roles, new pressing triggers. Our solver's ratings — built on the old manager's matches — should be stale. We tested whether skipping bets on these teams improves the portfolio.
The Question
The hypothesis: mid-season manager changes cause temporary chaos that the MI Bivariate Poisson solver can't account for. For 5-10 matches after a change, the solver's team ratings reflect the old system, creating systematic overconfidence. Skipping these bets should improve ROI.
This is the companion to manager-change-xg-window (which tried to window the xG data instead of filtering bets). That signal found zero marginal ROI. This one asks the simpler question: should we just avoid post-change matches entirely?
What We Built
We loaded 877 manager changes across 26 leagues from Fotmob's coaching history — 529 of those mid-season. For each mid-season change, we estimated the match number where the new manager took over from the previous coach's games-played count.
The filter: for any match where either team is within 10 matches of a mid-season manager change, skip it entirely. One config flag, one variable isolated.
This is a significant improvement over the prior test (2026-03-18) which only achieved 56% team name match rate using FootyStats coach IDs. The Fotmob approach covers all teams in the coaching history — no name matching needed.
What We Found
| Metric | Without Filter | With Filter | Delta |
|---|---|---|---|
| Bets | 6,606 | 6,424 | -182 |
| CLV | +11.2% | +11.2% | +0.0pp |
| ROI | -3.0% | -3.1% | -0.1pp |
| P&L | -195.9u | -198.6u | -2.7u |
Marginal ROI: -0.1%. The filter makes things slightly worse.
The 182 removed bets were performing at or above average. Skipping them removes good bets from the portfolio, not bad ones.
Why It Doesn't Work
Two reinforcing explanations:
1. The solver reads Pinnacle odds, not historical form. Our MI Bivariate Poisson model derives team attack/defense lambdas from devigged Pinnacle closing odds. When a team changes managers, Pinnacle adjusts its odds immediately — the market prices the chaos in real time. Our solver inherits that adjustment for free.
2. Manager changes create variance, not bias. A new manager doesn't systematically make the team worse or better — it makes outcomes less predictable. Higher variance means some bets win bigger and some lose bigger, but the expected value remains similar. The variance regression filter already handles this: it detects when actual results diverge from expectations and bets on regression.
What This Means
Combined with the manager-change-xg-window result (also zero marginal), we can close the book on manager changes as a signal source. Two different approaches — data windowing and bet filtering — both found nothing. The infrastructure we built (lib/backtest/manager-window.ts) stays available for future use, but the manager change hypothesis is dead from two angles.
What's Next
Nothing. Manager changes are priced. Move on.