What CLV Actually Rewards on MLB

Executive summary

1. Model selection beats the universe. Across ~1,100 games, picks the model flagged as +EV averaged +0.38pp CLV. Sides it didn’t flag averaged −0.11pp. On paired same-game comparisons the picked side outperformed the opposite side by +0.62pp on average. The model adds value — but not uniformly (see Finding 9 for the moneyline caveat).
2. Public sentiment is the single strongest CLV predictor we measured. Correlation of public-vs-sharp closing spread with CLV: r = +0.33. When public books are heavily fading our side, mean CLV is +5.3pp; when public is heavily on our side, mean CLV is −2.0pp.
3. Model “edge” doesn’t predict CLV. Price does. Correlation of pick edge with CLV: r = −0.02. Correlation of decimal odds with CLV: r = +0.30.
4. The sharp market keeps agreeing with underdogs, not favorites. Heavy dogs (+150 or longer): +3.0pp CLV. Heavy favorites (<−150): −5.4pp CLV. The pattern moves smoothly across every price bucket in between: the more of a dog we back, the better CLV we book.
5. Away-dog Run Lines are the biggest single inefficiency: n = 96 · CLV +8.9pp · WR 55.2%. The dog effect exists independent of public sentiment.
6. Curated TDL picks improved sharply after the June v1.2 retune. Before (n=38 settled): 39.5% WR, CLV −0.90pp. After (n=42 settled): 50.0% WR, CLV +0.09pp. The retune folded CLV-derived filters (away-dog, sharp drift) into selection.
7. Weekend games grade worst, Wednesdays best. Sunday −0.70pp (highest public-backing of any day), Wednesday +0.26pp.

What is CLV, and why publish it?

Closing Line Value measures the gap between the price you took and the sharp-book price the market settled at, in percentage points of implied probability. It grades process — not outcome. A sharp bettor books CLV whether the game wins or loses; the closing line is the market’s best guess at fair, and beating it means someone on the other side moved money toward your number after we locked it in.

TDL stamps every pick to GitHub + Rekor before first pitch, then joins the sharp-book closing consensus after the game. Every number in this note is auditable against those two feeds.

Finding 1 — Edge is not CLV

Plain English: the model’s own confidence score doesn’t tell you whether the sharp market will agree with the pick. Longer prices do.

The model’s own edge estimate is uncorrelated with the closing line’s verdict. Decimal odds carry roughly 30% of the CLV signal on their own.

The model is more confident on favorites, and confident-favorite picks systematically pay too much juice relative to close. Longer prices are where the sharp market has been agreeing with us.

Finding 2 — By market

Plain English: moneyline picks are roughly breakeven on price, totals and run-line picks give a small amount back to the market.

Moneyline is the only market breakeven on CLV. Spread runs consistently negative in aggregate — but that hides a critical sub-slice.

Finding 3 — The dog effect

Plain English: favorites win more often, but you pay too much for them. Underdogs win less often, but the sharp market agrees with the price you took — that’s value.

Favorites win more often — and pay for the privilege. Dogs win less often but book sharp CLV on average.

Finding 4 — By day of week

Plain English: mid-week baseball is where the sharp closes are. On weekends the public pounds the boards and prices drift.

Wednesday is the only positive-CLV day. Sunday shows the classic public-favorite pattern: highest WR of any day (50.9%) alongside negative CLV — we’re getting into positions the sharp market fades. Sunday also carries the strongest public-book skew of any day (mean public sentiment −1.70pp, vs Wednesday’s −0.97pp) — see Finding 7. Monday breaks the pattern (moderate public backing, worst CLV) so sentiment doesn’t fully explain the weekday effect.

Finding 5 — Team leaders

Plain English: some teams are consistently good bets, some are consistently traps. Braves are the sharpest side we’ve backed. Giants are the biggest trap.

Teams by mean CLV when TDL backed them. Minimum 40 picks.

Braves lead by a wide margin. Giants have been the season’s biggest process leak — the public keeps buying them, sharps fade every time.

Finding 6 — Outcome and process are only loosely coupled

Plain English: a great night doesn’t mean you made great bets, and a bad night doesn’t mean you made bad ones. Judge a service on a month of CLV, not on last night’s scoreboard.

Correlation of daily win rate with daily mean CLV across 90 game-days: r = +0.30.

Winning days trend sharper on average, but the coupling is loose — small samples of games are outcome-dominant. A 3-0 day tells you very little about how sharply the picks were priced. This is exactly why 30-day rolling CLV is the metric a sharp bettor grades a service on, not last night’s record.

Finding 7 — Public sentiment is the strongest signal we found

Plain English: when DraftKings and FanDuel post worse odds on your side than BetOnline and LowVig do, that’s the retail books hiking juice to slow down the public’s money. When they post better odds than the sharps, they’re trying to attract action to that side because nobody’s betting it. The picks that beat the close the most are the ones where the public isn’t on your side.

For every pick we computed a public sentiment value: the gap between the public-book close (DraftKings + FanDuel averaged) and the sharp-book close (lowvig + BetOnline averaged), in pp of implied probability. Negative = public book charges more juice than sharp on our side (public is on us). Positive = public book charges less than sharp on our side (public is fading us). Both numbers come from the same closing snapshot, so the comparison is apples-to-apples.

The pattern moves smoothly from most-negative to most-positive across every tier. Correlation of public sentiment with CLV: r = +0.33 — larger than the decimal-odds correlation (r = +0.30) and the largest single-signal correlation we measured. When the public is heavily fading us, mean CLV is +5.3pp. When public is heavily on us, mean CLV is −2.0pp.

This explains most of what Finding 5 showed: cold teams (SF, HOU) carry heavy public backing (−3.3pp, −2.0pp); the one top-CLV team that public actually fades is ATL (+1.2pp), and it’s the best CLV team we have. It does not explain the away-dog RL slice: heavy dogs carry near-zero public sentiment on average (0.00pp) yet still return +3.25pp CLV. That edge is real, not a public artifact.

Finding 8 — Curated TDL picks: pre vs post retune

Plain English: everything above was every side the model flagged. What we actually publish as a TDL Pick is a smaller, curated list. In late June we rebuilt the model to use these CLV findings in the pick-selection step. Since then, our published picks are winning at a 50% rate and beating the closing line.

Everything above was based on the full population of model-flagged +EV candidates. The curated TDL picks — the subset that gets stamped and published — is a smaller and more recent slice. On 2026-06-24 we shipped a v1.2 model retune that (a) removed leakage from the training candidates and (b) folded CLV-derived filters (away-dog, sharp-drift ≥+1pp) directly into pick selection.

Finding 9 — Does the model actually add value?

Plain English: we graded the model against itself. We measured the CLV of every side of every game — not just the ones we picked. Then we asked: did the sides the model liked beat the sides it didn’t? Yes on spreads and totals. No on moneylines — we’re working on that.

Everything above measured what correlates with CLV within our selected picks. The natural next question is whether the model’s selection itself beats a random side. To test it, we computed hypothetical CLV for every side of every MLB game we had both a morning-open snapshot and a closing snapshot for — 1,137 games × up-to-6 sides each = 6,426 side-pairs across the season. We then split by whether the model flagged that side as +EV.

Paired test: on the 1,494 games where the model picked exactly one side of a market, the picked side’s CLV beat the opposite side by an average of +0.62pp. On a side-by-side coin-flip, the model wins 47.9% and loses 46.8% (ties on the median).

But: the value isn’t uniform across markets.

Selection discipline: the model flags roughly 24% of the universe of sides as +EV (1,534 of 6,426). It isn’t forcing action — on 3 out of 4 sides available, it stays out. The selection itself beating the universe on spreads and totals is what makes this a betting product, not a data hobby.

Methodology

• Population: every model-flagged +EV pick logged to suggestions_log.csv, 2026-03-25 through 2026-07-06. This is our own selection, not all MLB games — findings describe the CLV of picks the model already liked.
• Closing line: sharp-book consensus (lowvig + betonlineag averaged) captured within ±15 min of commence. Median snap-to-commence gap: 5.4 min; max 14.4 min.
• Universe comparison (Finding 9): open price = sharp consensus from morning_lines snapshot (~7:25 AM ET), close price = sharp consensus at commence ±15 min. CLV computed per side; line moves handled with the same half-run tax + cap as the primary flavor. 91 dates of morning data joined against the closing cache (2026-03-25 – 2026-06-29). Sides are matched to suggestions_log by (date, home, away, market, side) to identify selection status.
• Public sentiment: (1/sharp_close − 1/public_close) × 100 in pp. Public close = average of DraftKings + FanDuel at the same snapshot. Same source doc as the sharp close, no additional fetches. On 311 picks (10%) the two book families closed at different point-values (spread/OU line move); those cases compare each book’s own line and are flagged as books_same_line=false in the joined dataset.
• CLV formula: (1/bet_decimal − 1/closing_decimal) × 100 — pp of implied probability.
• Line-move handling: ~16% of OU/spread picks were bet at a line the market later moved off. For these we apply a half-run-tax adjustment: 16pp per full run of OU move, 36pp per full run of RL move, capped at ±35pp per pick to prevent extreme cases (e.g., market flipping which team is favored) from dominating the aggregate. A separate strict flavor excludes these picks entirely; every finding above is qualitatively identical in strict, confirming the tax assumption doesn’t drive the story.
• Coverage: 99.5% of settled picks joined a closing line in the primary flavor; 83.1% in the strict flavor.
• Audit: 300 stratified random picks were independently re-computed against the raw closing snapshots. 100% match on the numeric join. A secondary historical-odds check was also run directly on 5 line-move cases and 103 missing-CLV rows.

What this means for a bettor reading this

1. Don’t grade a service (or yourself) on last night. 30-day rolling CLV is the signal.
2. Be skeptical of favorite-heavy picks. Even when a model likes them, the market usually likes them more.
3. Weekday closes are sharper than weekend closes. Wednesday–Friday hold the best value.
4. The model’s own edge score is not enough. We’re using this analysis to reweight future selection toward price zones where the sharp market has been validating us.
5. The moneyline market needs work. Selection there is a negative signal in this window. Spreads and totals are where the model earns its keep.