Note · 2026-05-15

How a Chiefs/Rams spread gets computed end-to-end

We walk a single NFL spread pick through all seven pipeline layers — de-vig, Elo, prediction markets, AI modifier, EV, Kelly, CLV. Concrete numbers at every step. No hand-waving.

A reader asked: when Consensus Edge says "B-grade pick, +5.7pp EV on Chiefs -3.5 at FanDuel -105," what's actually happening? Here's the full pipeline with real numbers from a typical NFL Sunday matchup.

The setup

Late-season Kansas City Chiefs hosting Los Angeles Rams. Pinnacle posts the spread at Chiefs -3.5 (-110/-110). FanDuel reduced juice is -3.5 (-105). DraftKings is -3.5 (-115). The best price for the bettor is -3.5 at -105.

Layer 1: de-vig the market

Pinnacle's -110/-110 reflects ~4.5% combined juice. Strip it: implied prob of each side = 52.4%, but combined that's 104.8% (the juice). Normalize so they sum to 1.0 → home cover prob drops to 50.0%, away cover prob rises to 50.0%. We do this across every book and weight Pinnacle higher because it's the sharpest anchor.
Square consensus: 54.0% Chiefs cover

Layer 2: run the Sharp model

Sport-specific Elo: KC at +145 rating vs LAR at +65 → 80-point differential → baseline 57.5% on a neutral field. Layer the situational signals from the NFL pipeline:
Sharp consensus: 58.7% Chiefs cover

Layer 3: read the prediction markets

Polymarket's NFL spread liquidity is thin but improving. Kalshi has cleaner regulated event contracts. On this matchup, Polymarket prices Chiefs cover at 58 cents, Kalshi at 57. Average (volume-weighted) ≈ 57.5%. PM signals tend to lead retail spreads by 30-90 minutes when there's news flow; here there's nothing new, so PM and Sharp roughly agree.
PM consensus: 57.5% Chiefs cover

Layer 4: blend with dynamic weights

The blend weights adjust based on signal agreement. All three (Square 54.0, Sharp 58.7, PM 57.5) point the same direction → confidence high → weights stay near baseline (Square 45%, Sharp 35%, PM 20%). Weighted average lands at 56.6%, rounded to 56.8% after the dynamic-confidence shaping.
Consensus blend: 56.8% Chiefs cover

Layer 5: compute EV

Best price for Chiefs -3.5 is FanDuel -105 → implied prob 51.2% (after de-vig assuming a 4.5% book hold). Our consensus is 56.8%. The EV in percentage points is the difference: +5.6pp. That's the raw edge before AI modifier.
Raw EV: +5.6pp at FanDuel -105

Layer 6: apply the AI modifier

The modifier scales raw EV based on three factors: Composite modifier score: +0.02 → effective EV: 5.6 × 1.02 = +5.7pp. Grade ladder for spread: ≥4pp = B, ≥7pp = A. Lands in B-grade territory.
Grade: B · Effective EV +5.7pp

Layer 7: capture for CLV audit

We log the offered line (-105 FD) and the time at pick. At kickoff we re-pull the closing line at the SAME book and at Pinnacle. The model's CLV = pick-time price minus no-vig closing price, in pp.

What CLV looks like for this pick later

Suppose the line closes at -4 -110 at FanDuel. The de-vigged closing implied prob for the Chiefs side is ~55.2%. Our pick-time implied prob (from -105) was 51.2%. CLV = 51.2 - 55.2 = -4pp from the FanDuel side; or against Pinnacle's closing prob if FD's close drifted, the math may differ. We publish the model's aggregate CLV at /trust/clv — if we're not beating the close on average, you'll see it there before you trust the next pick.

Why this matters

Most retail products show you a pick and a confidence level. We're not asking you to trust the pick — we're showing you the math so you can verify whether each layer makes sense, and publishing the calibration evidence at /trust/calibration so you can see whether our predicted probabilities actually hold. Sharp users want both. Sharp users get both.

See the live version of this for any current pick at /pick/<id> — every layer above renders on the permalink with the actual numbers.