Natasha Cloud points — every estimate
3 of 4 estimates available for points; they disagree by 7.00 (narrow). Inside the 7.2 cut on points, where nothing is measured either way. Width is NOT a direction signal: signed error and de-vigged calibration are flat across width on every cut (all |t| <= 1.10). Never flip a side on this number.
Tonight
THU · SEP 24Out She is Out tonight (Left Knee). Every rate below comes from games she played — tonight starts with whether she plays.
Season to date, tonight's game excluded. Rates, not projections.
The ladder, on one axis
points3 estimates span 7.00 (l10 14.00 down to h2h 7.00) — narrow. Inside the 7.2 cut on points, where nothing is measured either way. Width is NOT a direction signal: signed error and de-vigged calibration are flat across width on every cut (all |t| <= 1.10). Never flip a side on this number.
Computed on 3 of 4 rungs. A partial ladder spans less than a full one for reasons that have nothing to do with the prop, so the width below is not comparable to the wide/narrow cut for points.
This question named no opponent, so the head-to-head rung is read against her next scheduled game — the Washington Mystics. Ask it against another team to move that rung.
Every rung, and what it is worth
3 of 4 available · missing ones say whyplain mean of the last 5 games
plain mean of the last 10 games
No system projection is attached to this question, so the rung is empty. It is measurably worse than the closing line in any case — the finding is beside it.
The system projection is unreliable — measurably worse than the closing market line and worse than a plain L10 mean.
points MAE 5.428 vs the closing line's 4.764 (13.9% worse, t = +7.94); loses to an L10 mean on 8 of 8 stats; its side hits 53.19% against 53.18% for "always bet the under" on identical rows (+0.01pp)
8 games, -4.7, t = -2.14, measured (this sample resolves gaps above 4.3). Head-to-head is surfaced, never applied. It was deleted as a projection adjustment; it remains a field to read.
Head-to-head is surfaced, never applied. It was deleted as a projection adjustment; it remains a field to read.
the adjustment cost +0.087 MAE overall and +0.357 where it fired (24.5% of rows), with 62.7% of large shifts moving away from the actual; the coherent rebuild's MAE-minimising weight is zero (monotone sweep, w=0 -> 1.9604, w=1.0 -> 2.2869)
What the spread itself is worth
measured3 estimates span 7.00 (l10 14.00 down to h2h 7.00) — narrow. Inside the 7.2 cut on points, where nothing is measured either way. Width is NOT a direction signal: signed error and de-vigged calibration are flat across width on every cut (all |t| <= 1.10). Never flip a side on this number.
Measured, and NULL against its pre-registered bar: ladder width does not earn a stake rule. It is definitively NOT a direction signal — never flip a side on it. On POINTS ONLY, a very wide ladder means the outcome is harder to predict (a bigger miss in EITHER direction), which is a reason to size down or skip. On rebounds, assists and threes it is not measurable at all.
the bar required dispersion to rise monotonically with width AND market error to stay flat. Leg (b) passes on every cut — signed error and de-vigged calibration are flat across width, all |t| <= 1.10 (t = +0.17 at the points threshold). Leg (a) fails: the monotone rise is the line level, and controlling for the line the slope flips to -0.155 (n.s.); under (stat, line) fixed effects the pooled slope is +0.147, t = +1.91, p = 0.057, and the quintiles are not monotone. What narrowly survives is points-only and a TAIL STEP, not a gradient: the top ~17-20% of widths carry +0.61 more market-error MAGNITUDE (t = +3.28, p = 0.0010 on the pre-specified top-quintile rule; quintiles -0.037, -0.115, -0.216, -0.114, +0.486), robust in both season halves (+0.67/+0.68), leave-one-out stable across 141 players (+0.59..+0.72) and not a line artefact (t = +3.43 with the line controlled) — but it FAILS out-of-sample threshold selection (fit on H1, tested on H2: +0.51, t = +1.71, p = 0.088), so it is suggestive, not established. Non-points top quintiles are null: rebounds +0.036 (t = +0.44, p = 0.658), assists about -0.007 (n.s.), threes n.s. Median ladder width is 4.31 on points against 1.69 rebounds, 1.40 assists and 0.95 threes, which is why no single global cut can serve them
wide above a ladder width of 7.2 on points: the top quintile, 20.1% of rows, where market-error MAGNITUDE runs +0.61 higher (t = +3.28, p = 0.0010) with signed error flat. Suggestive, not established — this threshold failed out-of-sample selection (fit on the first half of the season, tested on the second: +0.51, t = +1.71, p = 0.088). Points is the only stat with a measured cut. A wide spread is at most a reason to stake less, and never a reason to pick a side — the ladder disagreeing tells you the estimates are uncertain, not which of them is right.