New to these reports? Start here
- Dotted-underlined words have a plain-English definition — hover or tap them. Every term is also on the glossary page.
- "Null" means we found nothing, not that something broke. Most reports here are negative results, on purpose — knowing an idea doesn't work is the point.
- Two questions get asked separately. First, is the effect real? Second, is it already priced into the betting odds? An effect can be completely real and still useless to bet on.
- A "calibration" row is a self-check. It runs the same method on something already known to be true. If that fails, the whole report is unreliable — so it's reported alongside the findings.
- If a confidence interval includes zero, the real effect might be nothing at all, so no claim gets made.
MLB consensus-stacking pilot — INSUFFICIENT DATA, leaning REJECT (2026-07-04)
Question: does blending the pitcher-adjusted EloElo ratingA rating system, originally from chess, that moves a team up or down based on results and the strength of the opponent. v2 prob into the devigged closing consensus (NHL v3 recipe) beat consensus-only?
Data
unified_odds MLB capture starts 2026-05-22: 408 games, closing h2h on 389, 12 books. Matched gradeable games (odds × results, ties dropped): 293 (2026-05-22 → 2026-07-03). corr(consensus, v2) = 0.873 — v2 carries almost no information orthogonal to the close.
Results
| setup | n | cons LL | stack LL | ΔLL | paired 95% CI95% confidence intervalThe range the true value is plausibly in. If this range includes zero, we cannot rule out that the real effect is nothing at all. |
|---|---|---|---|---|---|
| Expanding walk-forwardwalk-forwardEvaluating week by week using only what was knowable before each week, mimicking how the model would actually have been used at the time. (train 60%) | 118 | 0.6769 | 0.6834 | +0.0064 worse | [−0.0116, +0.0240] |
| Leave-one-week-out CV | 293 | 0.6736 | 0.6872 | +0.0136 worse | [+0.0025, +0.0245] |
Both fail the gate; LOWO significantly worse.
Power analysis (the real deliverable)
Per-game paired ΔLL sd = 0.0981 → detecting a 0.005 LL improvement at 80% power needs ~3,025 gradeable games. At ~15/day the 2026 season ends with only ~1,900 banked, so the gate is realistically not decidable before mid-2027.
Verdict
INSUFFICIENT DATA, leaning REJECT. In-samplein-sampleMeasured on the same data used to build or tune the idea. Nearly always looks better than reality. the stack is already worse and the candidate signal is 0.87-redundant with the market. Keep v2 informational-only, keep daily capture running, revisit no earlier than mid-2027. Do not wire anything now.
Reproduce: python3 backtests/stacking_eval/mlb/run_mlb_stacking.py