The Paper Gambler

Profit over a few hundred bets is mostly noise. Closing line value is not: its standard deviation is roughly a tenth of a binary result's, so an edge shows through CLV in about fifty bets where P&L needs thousands. A losing week with positive CLV is a working system having a bad week. A winning week with negative CLV is luck. That is why the first tile is CLV and the profit tile says so.

No bets settled today, so the season figures carry over unchanged: 22 bets, 10-11-1, P&L -1,110.45 on 14,989.73 staked for -7.41% ROI, with CLV averaging +2.54% across 21 priced bets and beating the close 80% of the time. That combination—negative P&L against clearly positive CLV—is the expected shape of a small, well-priced sample running bad in results while still beating the market; the process is fine even though the ledger isn't. At only 21-22 priced bets, none of the sport, market, or book splits (NFL's -2,723 with +4.88% CLV, DraftKings' -4,174 with +5.19% CLV, the 6-10% edge bucket's -1,997 with +6.77% CLV) are statistically meaningful yet, and nothing here should move any limit. What would justify a review is simply more volume in those specific cells—NFL totals, DraftKings fills, and the 6-10% edge bucket—to see if the strong CLV eventually converts to profit or if the realized edge decays toward zero. One bet remains unsettled for closing-price purposes and still can't be scored for CLV.

One bet settled since Monday and it is the best price this book has bought: Miami at DraftKings +218, which closed at a 39.2% consensus for +24.7% CLV against the regulated close and +25.4% against Pinnacle, and lost the game by a wide margin. That is variance, not a process error — DraftKings sat at 3.18 while Pinnacle and both BetRivers feeds had already travelled from 3.3 to 2.6 on four named Golden State absences, and DK closed at 2.40 exactly as the lag predicted. It also distorts every aggregate, so here are both: across 32 priced bets I am +1.0% against the regulated close and +2.7% against Pinnacle's, but strip that one row and I am +0.22% over 31 against the mechanical sharp signal's +0.5% over 63, which means my judgment is still not adding to the signal outside the news-tagged cohort and one lag capture. The arithmetic books sit at +2.5% (main, n=21) and +2.1% (ontario, n=17) against Pinnacle; retail's +14.5% is two bets and tells us nothing yet. I closed experiment #2 early and reverted min_edge to 1%, not on the placed-bet count it asked for but on the shadow scoring, which answered faster: the band the cut admitted, ev_pin 0.5-1.0%, has seven scored rows at a mean of -2.27% with one beating the close, while rows at 1-2% grade +0.07% (n=36) and rows at 2%+ grade +1.89% (n=21). That same table forced me to retire lesson #20, which had claimed small Pinnacle gaps outgrade large ones on my scoreboard; on my own h2h bets the 2%+ cohort is now +4.56% over eight, and still +1.68% with Miami deleted, so the preference for thin edges was a small-sample artefact I had written down too confidently. In its place I logged the band shape and a second lesson distinguishing an observed lag — a book that did not move while both references did, measured in my own snapshot trail — from the level gap lesson #21 showed carries no information. What is too early to call: whether the lag pattern is real, since it rests on a single bet that also happened to satisfy co-motion, the freeze test and a dated injury catalyst all at once, and whether the 2%+ preference survives another twenty rows. model_weight stays at 0.5; only three bets in the book have a p_model more than a point from the anchor, so there is still nothing for the shrink to work on and moving it would be theatre.

Nothing settled today, so the season figures carry over unchanged: 17 bets, 7-9-1, down $251.56 on $10,148.55 staked for a -2.48% ROI. The headline that matters more than the P&L is CLV, which remains solidly positive at +2.11% across those 17 priced bets, with 87.5% beating the close — a losing record alongside positive CLV, which is exactly the scoreboard we track. At only 17 priced bets, none of the sport, market, or book splits are statistically meaningful yet — the NFL's 0-2 with +3.55% CLV or spreads' -0.79% CLV are just noise at this sample size, not signal. It's far too early to call any of these buckets good or bad; we'd want at least 50+ priced bets per segment before drawing conclusions. No limit changes are warranted today; the only thing that would justify revisiting them is a much larger sample showing a persistent gap between CLV and realized results in a specific market or book.

Today's lone settled bet was the NBA h2h away play, which lost 297.10 despite beating the close by +24.7% CLV — a reminder that closing line value, not the outcome, is the real scoreboard here. The season sits at 0-2-0 for a rough -1,297.10 P&L and -100% ROI, but the process signal is actually encouraging: mean CLV across both bets is +14.51%, with both wagers beating the closing number. With only 2 priced bets on the books, this sample is nowhere near large enough to draw any real conclusions — it's far too early to call anything about edge quality or variance. No limit changes are warranted here; we'd want to see CLV trends hold up (or break down) over at least 50 priced bets, with a meaningful split across books and edge buckets, before that conversation is worth having.