Promoted Teams Are Overrated — But Not by the Model
We tested whether filtering bets backing promoted teams improves ROI. 339 promoted teams, 26 leagues, 5 seasons. The effect exists (+0.1pp) but is statistically insignificant (p=0.46) and temporally unstable. The MI solver reads Pinnacle odds, and Pinnacle already prices promotion correctly. Signal #42 tested, signal #42 rejected.
Everyone knows promoted teams struggle. They finish bottom half, they concede more, their players aren't good enough. Surely the model overrates them, right?
We tested it. 339 promoted teams across 26 leagues and 5 seasons. The hypothesis was clean: the MI solver builds ratings from devigged Pinnacle odds, but promoted teams' ratings carry residue from their lower-division data. Filter bets that back promoted teams, capture the edge from the market overpricing them.
The result: rejected. Not because the direction was wrong — it wasn't. But because the effect is so small it's indistinguishable from noise.
The Implementation
We built automatic promotion detection into the data loader. A team is "promoted" if it appears in a league season but wasn't present the previous season. UCL was excluded — teams qualify, they don't promote.
The filter was surgical: skip directional bets (1X2 Home/Away, AH) that *back* a promoted team. Bets *against* promoted teams pass through. O/U bets are unaffected — they're not team-directional.
339 promoted teams detected. About 588 bets removed from the 6,606-bet stack.
What the Numbers Say
Marginal Impact
| Metric | With Filter | Without Filter | Delta |
|---|---|---|---|
| N | 6,018 | 6,606 | -588 |
| ROI | -2.8% | -3.0% | **+0.1pp** |
| CLV | +11.2% | +11.2% | 0.0pp |
The filter improves ROI by one-tenth of a percentage point. That's real, but it's nothing.
Gate Results
The signal ran through our 10-gate approval process:
| Gate | Result |
|---|---|
| Pre-registered | PASS |
| True standalone (minEdge=0) | PASS |
| Minimum N >= 1,000 | PASS |
| Marginal ROI > 0 | PASS (+0.1pp) |
| Bootstrap significance (p < 0.10) | **FAIL** (p=0.46) |
| IS/OOS replication (within 3pp) | **FAIL** (3.9pp gap) |
| Regime stratification | PASS |
| Suspicious N | **FAIL** |
| Practical significance (> +0.5pp) | **FAIL** (+0.1pp) |
| Walk-forward (2/3 folds positive) | **FAIL** (2/4) |
Five gates failed. The two most damning:
Bootstrap p = 0.46. You'd get this result by chance almost half the time. The signal is statistically indistinguishable from zero.
Walk-forward collapse. The filter worked in 2022 (+2.3% ROI) and 2023 (-0.2%, basically flat). Then it fell apart: 2024 at -3.9% and 2025 at -9.8%. Whatever small edge existed early didn't persist.
Why It Failed: The Market Already Knows
The hypothesis assumed the MI solver would overrate promoted teams because their ratings carry lower-division residue. But the solver doesn't build ratings from historical match results — it extracts them from devigged Pinnacle odds.
Pinnacle already prices promoted teams correctly. Their odds reflect the market's consensus that Ipswich isn't as good as Liverpool, that Holstein Kiel isn't Bayern. The solver inherits this pricing. There's no systematic overvaluation to exploit because the information source (sharp odds) already incorporates the promotion penalty.
This is a general lesson: signals that rely on information the market already has are dead on arrival. Promotion status is public knowledge. It's priced in before the first ball is kicked.
The Odds Quality Angle
One interesting finding from the informational gate:
| Tier | N | ROI | CLV |
|---|---|---|---|
| Sharp (EPL, La Liga, Bundesliga) | 1,952 | -1.6% | +11.3% |
| Medium (Serie A, Ligue 1, etc.) | 1,730 | -5.5% | +11.0% |
| Soft (lower leagues) | 2,336 | -1.8% | +11.1% |
CLV is +11% everywhere — the model is equally good at finding edge regardless of league tier. But medium-tier leagues have worse ROI despite similar CLV. This isn't about promoted teams — it's about the CLV-to-ROI conversion problem we've documented elsewhere.
What We Learned
Not every plausible hypothesis is a signal. "Promoted teams are overrated" sounds right. The data shows a directional effect (+0.1pp). But directionally correct and deployable are different things. The pipeline exists to catch exactly this — ideas that feel right but don't survive statistical scrutiny.
The solver's information source matters more than its architecture. We assumed the solver would be blind to promotion status. It's not — because it reads odds, and odds already know. If we built ratings from historical xG instead of market prices, this signal might have worked. But we don't, and that's by design.
339 promoted teams, 588 filtered bets, +0.1pp improvement, p=0.46. That's the denominator doing its job. Signal #42 tested, signal #42 rejected. The registry keeps us honest.