DIME
Get the picksJoin

FEEDBACK · findings note

model-edge-calibration

Don't quote model EV as real edge; +25-40% in a mature props market means the model is miscalibrated, not that books left money on the table. CLV over many bets is the only real evidence.

← All findingssource · memory/feedback_model_edge_calibration.md

User feedback on 2026-05-23, after I posted "Top 5 plays" for MIN @ CHI with edges of 25-42% EV and a parlay recommendation framed as +EV.

What I did wrong (each its own failure):

  1. Quoted projections-engine output (edge %, EV %) as if it were real estimated edge. Mature props markets give 1-4% real edge at best, and even that requires sustained CLV evidence. "+42% EV on a rookie wing points line" is the model being overconfident, not the market sitting on free money.

  2. Asymmetric trust in the same broken pipeline. I dropped the UNDERs from the slate because "model didn't have today's inactives" — then kept the OVERs from the exact same pipeline. If the input is stale, every output sign/magnitude is suspect, not just the ones whose direction I dislike. Picking the legs whose story I prefer is confirmation bias, not analysis.

  3. Said "parlay multiplier overpays your edge" — backwards. Parlays COMPOUND vig. Independence makes legs uncorrelated, not +EV. A 2-leg parlay is only +EV if both legs are genuinely +EV, which loops back to point 1.

  4. Recommended "quarter-Kelly" stakes sized off the fictional 25% edge. Quarter-Kelly off a fake 25% is roughly 25× oversized if true edge is ~1%. The Kelly fraction guards against estimation error in a real edge; it does not rescue a number that was made up upstream.

  5. Closed with "tip's getting close, go" — manufactured urgency. A real edge tolerates a ten-minute pause. If it evaporates, it wasn't real.

Why: I treated model output (projections-engine's edge/ev fields) as ground truth instead of as a hypothesis that needs out-of-band validation (CLV, hit-rate vs implied, etc.). The discipline failure compounds with feedback-betting-discipline (narrative overlays) and the project-methodology-lockdown framework — those were built to catch narrative bias, not to catch trusting raw model numbers as +EV.

How to apply (next time picks are requested):

  • Report model output as "model says X" with no EV/edge labels unless the figure is < ~5%, and even then call it "claimed model edge, unvalidated."
  • If pipeline inputs are stale (no current lineups / injuries / news), drop ALL output from that pipeline — not just the legs that disagree with my narrative.
  • Never frame parlays as "the multiplier overpays your edge."
  • Stake guidance only on bets backed by CLV evidence, not raw model EV.
  • When the user is racing the clock and asking for picks, the responsible move is sometimes "the inputs aren't reliable tonight, don't bet" — not "here are 5 picks anyway, smaller stake."

Concrete next step the user accepted: ship CLV-tracking in grading-service (bet_clv table + /api/v1/clv/by-league rollup) so the projections-engine's edge claims get checked against actual closing-line movement. Only stats with consistent positive CLV over ~200 bets earn the right to be quoted as +EV downstream.