CY Elite Lab

Confirmed / Projected — whether both teams have ALREADY published their official lineup for the day or not yet. With the lineup confirmed, the Model projects with the nine hitters who will actually play.

The odds next to the team — what the market pays right now: the consensus and, if there isn't one yet, the reference market's price.

% win and the odds in parentheses — the win probability the Model gives each team and, below it, the price that would be fair if the Model were right, with no commission. If the price above pays more than the one below, the Model sees value on that side.

EXP RUNS — the runs the Model projects for each team. It is an average, not a line. It is checked every morning against the runs actually scored.

o2.5 · o3.5 · o4.5% — the probability that this team goes over its runs line for the day, the same one the Market shows (the one closest to even). It can't be deduced from the EXP RUNS: a team scores 4.41 on average but 4.0 at the median, and gets shut out 7.2% of the time. That's why the break-even point of an over 4.5 is not at 4.5, it's near 5.2: with the team projected at 5.00, the over still only comes in 48% of the time. Calibrated on Oct 6, 2026 over 980 team games.

Expected total — the sum of the two projections; below it, the market line with the over price and the under price. The expected total is not compared to the line, and it's the easiest mistake to make: it is a mean, and the line is set where half the games go over it (the median). Over 368 games, at each line teams scored on average between 0.2 and 0.8 runs MORE than the line, so a well-calibrated model always ends up above it. For the over/under, look at the chip below, which already compares probability against probability.

Details ▾ — opens up, for each team: its profile in five categories, the starter with his last 5 starts, the lineup with each hitter's OPS and the bullpen's actual workload over the last 48 hours.

Profile · CON POW EYE BSR DEF — how that team plays, not just how much it scores: CONtact (how often it puts the ball in play), POWer (how hard it hits when it connects), EYE (how much it walks), BSR (how much it gains on the bases) and DEFense. Two teams with the same OPS can play in opposite ways. It goes in Details, not on the card, because it's context: the Model projects with OPS against today's pitcher's hand, not with these letters.

How defense is measured — with defensive efficiency: the percentage of balls in play it turns into outs. Not fielding percentage, which only counts errors — and a slow outfielder who never reaches the ball never commits one. Across the 30 teams, efficiency ranges from .684 to .737; fielding percentage is bunched between .983 and .991.

Card order and freezing — the cards are ordered by first pitch time, from the earliest of the day to the latest. When a game starts, its card freezes just as it was: a record of what the Model said BEFORE, and refreshes only touch games that have not yet begun. There is no live score here: the Model shows what it projects, not how the game is going.

ERA · FIP · K-BB — the starter's three numbers, always in that order and with their grade beside them. The ERA is the runs charged to him; the FIP is what truly depends on him; the K-BB%, his strikeouts minus the walks he gives away. They're the numbers everyone recognizes, but the Model doesn't project with any of them (see “What the Model projects with”).

The two FIPs — the card shows the one from this season, the one you recognize at a glance. Details shows the one from three seasons, which is the pitcher's baseline quality. When the two are far apart, that itself is information: that pitcher is having an unusual year compared with his track record. The grade (A/B/C) always comes from the three-season one.

What the Model projects with — since Sep 17, with a composite: two-thirds of the Statcast xERA (the ERA he deserved for the quality of contact he allows) and one-third FIP. With it, the ~240 starters in the league are ranked every day and the ranking is saved, the same way the reference source does it. If a starter has no Statcast sample, the Model uses his three-season FIP. ERA doesn't go in: splitting the season of 80 starters into even and odd starts, the correlation between the two halves is −0.06 in ERA y +0.32 in FIP. Today's ERA does not predict tomorrow's.

A starter not yet announced (TBD) — is not treated as an average starter: he goes in as his own team's bullpen, which is who will actually pitch. Measured across 2,206 team-games: teams with no announced starter allowed 5.13 runs versus 4.48 for the rest.

A FIP with no grade — means that pitcher doesn't have a large enough sample (fewer than 300 batters faced over three seasons). It's not a bug: it's how we avoid hanging a letter on him that he hasn't earned. Internally, the Model does the same: it shrinks him toward the league average.

F (FIP) — measures only what the pitcher controls (strikeouts, walks, home runs); removes luck and defense, which is why it predicts better. Same scale as ERA. FIP lower than his ERA = improvement coming; higher = overvalued.

BP — that team's bullpen: the ERA and FIP of its relievers, weighted equally. Next to it, how fresh it is: rested, the innings it has thrown in the last 48 hours, or how many relievers are burned out (threw on two straight days). Since Sep 17, the Model measures it with a ranking of the 30 bullpens that mixes each reliever's FIP and xERA, weighted by their innings: a bullpen's ERA is one of baseball's noisiest numbers.

OF — that team's offense against today's starter's throwing hand, not its overall index: 100 = the league. Hitting .700 against lefties is not the same as against righties, and the Model matches each offense against the opposing starter's hand.

PF · WX · UMP — the three factors that multiply the projected total, each with the % it moves. They are shown separately so you can see where the number comes from.

PF (park factor) — how many runs this ballpark produces relative to the league average: the run index from Statcast, over three seasons and by venue (not by team, which breaks when a club moves). It's used in full, without trimming: Coors +25%, Globe Life −12%.

WX (weather) — how far today's conditions are from what is normal in THAT park at THAT time of year, not from the league average: Kansas City's typical August heat is already built into its park factor, and counting it twice used to inflate the total. It uses temperature, wind speed and wind direction. Measured over 1,513 games: +0.5% per degree above normal and +0.4% per mph of wind blowing toward center field. Closed = the roof is closed and weather does not apply.

UMP (umpire) — the home plate umpire's effect on runs, measured over 6,934 games across three seasons. Each game is compared with that ballpark's average that year (so the park doesn't skew it) and the result is shrunk according to how many games he has worked: at 80 games it keeps less than a third of its number. The real spread between umpires is 0.31 runs. If it doesn't show, the umpire hasn't been announced yet.

Fair line — the market consensus with the commission already removed: what the market really believes. It tells you how far the Model disagrees with the consensus. The chips' edge, on the other hand, is measured against what you're paid (see below), because that difference is what determines whether there is an edge.

CHC ML +4.2% — the Model's edge on the winner, in percentage points: what the Model gives that team minus what the price you're paid requires, commission included. It shows up whenever there's an edge, big or small; if neither side has one, the chip says "No ML edge".

Over 9.0 +2.1% — the same for the total, also in percentage points. The Model converts its expected runs into over and under probabilities, and each side is compared with what that side pays; the side with the larger margin is the one shown. For now it's context, not a signal: on Sep 17, across 227 games, the totals model favored the right side 50.9% of the time, and at −110 a 52.4% hit rate is needed to break even. Now it's measured each morning: the side it favors is right 50.5% of the time, requiring 52.1% (105 games); it will turn green when it beats the close by 3 points across 150 games. It will turn green at 3 points, like the winner's chip, once the side it favors beats what its prices required by at least 3 points over 150 games; and if it later drops, it goes back to white on its own.

The chip's color — green = 3 points or more (on the winner; on the total, only when it has earned it, see above), white = less than 3, gray = no edge. Heads up: the color does NOT mean "stronger signal". It only shows how far the Model is from the price. Read the entries below before using it.

The Model's edge is NOT a signal, and this was corrected on Sep 9, 2026 — until that date this said the 3-to-6-point band was the only one with a positive balance (50.3%, +4.8%). That number came from comparing the Model with the analysis history. Measured the right way, against the ACTUAL results of the games, the result was the opposite (with the Model and the measurement method in use before Sep 17):

3 to 6 points — 40 games: hits 35.0% needing 47.4%. Balance −31.2%, with a confidence interval entirely below zero. It was the worst of the three bands, not the best. · 6 to 10 points — 41.2%, balance −12.4% · More than 10 — 28.6%, balance −18.9%.

Why it failed — over 507 games with results, the Model had a Brier of 0.2488 versus 0.2386 for the market, worse than always predicting the same thing (0.2480). It squeezed everything toward 50%: when the market gave the home team 68%, the Model said 58%. And when it disagreed with the market it was wrong in the direction of its disagreement.

What was fixed on Sep 17 — it was recalibrated with 372 games split in two: the slope, which measures how far the Model commits compared with the market (1.00 = the same), went from 0.47 to 0.95, and the Brier fell from 0.2433 to 0.2378 in the half the Model had not seen. That same day the starter composite and the bullpen ranking went in, so that measurement has to be redone: it's being done with the data saved every day since then.

What to do with this — since Sep 9 the Model doesn't reach the brain: analyses are decided without it, with pitching, bullpen, offense, form and price. The way back is written in advance: its Brier has to match or beat the market's over 150 or more games, and the morning report checks it automatically. Even then it would come back gradually: first in shadow mode, without deciding, and then with a small weight. Meanwhile, use it to understand the game — which offense is hitting, which bullpen is holding up, how much the park and weather weigh in — not to choose a side.