Testing Edge Floors: Why Raising the Minimum Doesn't Help
Bucketed 138K bets into 5 edge brackets and swept 25 floor thresholds. Higher edge does NOT predict higher ROI — the 3-5% bracket (-4.5%) outperforms 12%+ (-8.7%). DSR=0.000. The CLV→ROI gap is structural (calibration + market structure), not driven by edge size. AH near breakeven at all edges; 1X2/OU25 lose everywhere.
Testing Edge Floors: Why Raising the Minimum Doesn't Help
From 55 live paper-trade bets, 7 of 17 losses had edges under 8%. The wins with thin edges were often barely covered or half-wins. Natural hypothesis: raise the minimum edge threshold to filter out marginal bets.
The Question
Should we raise the minimum edge from 7% to 8%, 9%, or higher? The intuition says thin edges convert less reliably — are we wasting bankroll on bets where the model barely sees value?
What We Found
Every edge bracket has negative ROI. Not just the thin ones — all of them.
| Bracket | N (test) | ROI | CLV |
|---|---|---|---|
| 3-5% | 8,922 | **-4.5%** | +3.9% |
| 5-7% | 6,243 | -6.4% | +5.9% |
| 7-9% | 4,359 | -7.3% | +7.9% |
| 9-12% | 3,927 | -7.4% | +10.3% |
| 12%+ | 4,167 | **-8.7%** | +16.3% |
The lowest edge bracket has the *best* ROI. The highest edge bracket has the *worst*. This is the opposite of what we expected.
CLV increases monotonically with edge (by definition — it's the same variable). But ROI gets worse as edge increases. The gap between CLV and ROI grows from 8.4pp at 3-5% edge to 25.0pp at 12%+ edge. Higher model confidence → bigger mismatch with reality.
This is the "edge-band death spiral" we identified in the March 17 CLV-ROI diagnostic. It affects 1X2 and OU25 catastrophically. In AH, thin edges (3-5%) show -0.2% ROI — essentially breakeven — while fat edges (9-12%) show -9.1%.
The Sweep
We tested 25 edge floor thresholds (3.0% to 14.5% in 0.5pp steps) across train/val/test:
The "optimal" threshold on training data was 14.5% (Sharpe -0.087 — the least negative). On the test set, that same threshold produced Sharpe -0.271 and ROI -11.0%. The current 7% floor produced Sharpe -0.295 and ROI -7.8%. The "improvement" is within noise: Deflated Sharpe Ratio = 0.000, confirming the observed difference is explained entirely by multiple testing.
The Nuance
The market-type breakdown explains everything:
- AH bets at thin edge (3-5%): -0.2% ROI. Near breakeven.
- AH bets at fat edge (12%+): -2.5% ROI. Slightly worse.
- 1X2 bets at thin edge: -4.1% ROI. Bad.
- 1X2 bets at fat edge: -18.6% ROI. Catastrophic.
- OU25 follows the same pattern as 1X2.
The edge floor is irrelevant because the problem is market structure, not edge size. In AH (2-way outcome), the model's calibration error causes minor ROI loss. In 1X2 (3-way outcome), the same calibration error amplifies into massive loss — and high-edge 1X2 bets are where the model is most overconfident.
Since we already disabled 1X2 and run OU25 at 0.25x in production, the live portfolio is mostly AH bets where edge floors don't matter much.
What This Means
No production change. The 7% minimum edge stays. The problem we thought we saw in 55 live bets (thin edges losing more) was just variance — the backtest on 138K bets shows no relationship between edge size and ROI direction.
The real lever for improving ROI is the CLV→ROI conversion gap, which is driven by model calibration (especially the home advantage parameter and totals distribution), not by filtering which bets to take.
What's Next
The CLV→ROI gap remains the central challenge. The existing CLV-ROI diagnostic identified three root causes: (1) home advantage inflation, (2) OU25 totals miscalibration, (3) 1X2 3-way outcome amplification. The first two are active areas of solver tuning. Edge filtering is a dead end.