Signal Methodology 2026 08

How Signal Syndicate Researches Sports Markets — August 2026 Edition

SOURCE NUMBERS (from the append-only PG ledger, 2026-08-21): 427 graded signals, 262-162 record, +13.70 units, +3.21% ROI. Run Lines 116 graded 62-53 +1.84u (backtest +1.77% ROI, n=998, 2024-2026). Hits Props OVER 96 graded 58-38 +7.47u (backtest +13.32%, n=2,235). Hits Props UNDER 182 graded 122-60 -2.79u (shadow). MLB Totals UNDER 17 graded 12-3 +7.91u (shadow). F5 Innings 16 graded 8-8 -0.73u (shadow). CLV: Run Lines n=36, average price-CLV +7.11% (median +2.5%). Weather study: windy-game TB MAE 1.40 vs calm 1.53. Market-alpha study v1.1: pooled 65.1% (block-bootstrap CI 64.1-65.9); hits 57.9%, RBIs 60.3%, total bases 49.9% (coin flip), HR 91.9% (base-rate artifact, published with caveat). Projection audit: live grading bias -0.30, level ratio 0.643, walk-forward directional 49%.

The Ledger Is the Product

Every signal we publish gets graded WIN/LOSS/PUSH at the real market odds we posted at entry. Never flat -110 assumptions. Never retroactive edits. 427 graded signals through August 21, 2026. Net +13.70 units, +3.21% ROI. The ledger is append-only Postgres. We do not edit history. We add to it.

Shadow lanes (Hits Props UNDER, MLB Totals UNDER, F5 Innings) are tracked publicly but never promoted until they clear backtest, holdout, and forward shadow gates. Right now they sit in shadow for a reason.

Five Lanes, One Discipline

Run Lines and Hits Props OVER are live. The other three run in shadow on purpose. A lane earns promotion through validation, not through a hot week. MLB Totals UNDER has never produced a promoted signal — its edge cap has filtered every candidate since the lane began — and the settled ledger shows exactly that. We would rather publish silence than noise. The gate itself: 75 forward shadow bets minimum before any promotion review.

Projections Scored Like a Quant Desk, Not a Tout

We retired directional accuracy as a vanity metric after our own audit showed it flattered a broken engine. Level ratio 0.643, live bias -0.30. Those numbers failed the gate. So we publish them anyway.

Now we score bias near zero, level ratio near one, error versus naive baselines, and alpha versus the market line. The player engine is a 4-layer weighted blend: season 20%, recent form 25%, vs-team 20%, park 12.5%, plus a situational delta. There is no secret fifth layer. The Statcast layer we once advertised has no free data source. So we say so.

Weather and Market Alpha, Published Either Way

Our weather study found wind matters for total bases. Windy-game MAE 1.40 vs 1.53 calm. A candidate real layer, not yet promoted.

Our market-alpha study says our projections beat the book on hits (57.9%) and RBIs (60.3%) but not total bases (49.9%). The 91.9% HR number is a base-rate artifact we flag ourselves. Pooled 65.1%, block-bootstrap CI 64.1-65.9. Honest numbers are the moat.

Closing Line Value, the Long Game

Run Lines CLV is price-based. The spread never moves; the price does. n=36, average +7.11%, median +2.5%. We are beating the closing price more often than not. Per-model CLV for props and F5 begins accumulating now that closing prop snapshots persist. Small n is reported as small n, never stretched.

Validation Is Walk-Forward or It Doesn't Count

Weights are chosen on past data only and scored once on unseen data. Errors are block-bootstrapped by game day. The OOS evening-odds reserve is held out of training entirely. When a number fails a gate, we publish the failure. The August projection audit is our most-read research for exactly that reason.

What We Will Never Do

No cherry-picked screenshots. No flat -110 fairy tales. No player names in public briefs. No editing the ledger. Educational research only, not betting advice.