Vegas is the floor. Everything else has to earn its place.
We start from the prop market, the sharpest signal in sports, and turn the full alternate-line ladder into a probability distribution. One rule governs every proprietary layer on top of it: it ships only after it beats a market-only baseline on data it has never seen. That discipline is the moat, not a pile of secret sauce.
The engine, step by step
Read the whole betting market
Not just the headline number. We read the full range of outcomes the sharpest room in sports is already pricing for each player, along with how the lines open and move. That market is the most accurate public signal there is, and it is our starting point.
Strip out the bookmaker's margin
Sportsbooks bake a margin into their prices. We remove it so what is left is the market's true implied probability, with extra care on the longshot outcomes where that margin hides and a naive approach would inflate the ceilings.
Rebuild the full range of outcomes
We turn those cleaned odds into a complete outcome distribution for each player and simulate the season many thousands of times, with yards, receptions, and touchdowns moving together the way they actually do, never independently. The output is a full picture in fantasy points: floor, median, ceiling, boom%, bust%.
Classify why the line moved
Sharp steam, injury, weather, public money, book balancing, hype. A move driven by public sentiment before a primetime game is noise, and we pull the distribution back toward the opening line instead of chasing it. A sharp or injury move is already correct, so we leave it alone.
Add only what beats the market on the residual
Season point projections, week-to-week volatility, role-change detection, situational signals. Every signal is measured against what the market already implied, so nothing Vegas priced gets counted twice, and only the effect the market missed survives.
Rank by value, not raw points
Value over replacement, computed from your league config, drives the rankings, with tiers at the real value cliffs and a projected finish range on every player. The Divergence Score is how far Nova's independent projection sits from the Vegas number, scaled by how much uncertainty surrounds it, so thin markets and volatile players self-discount.
The market is the baseline every layer has to beat
The market consensus, with no proprietary layers at all, runs as a permanent benchmark. Every layer we add ships turned off and earns its way on by beating that baseline out of sample, walk-forward by week, never a random split. Most fantasy narratives fail this test, because the market already priced them. A layer that only underperforms a season average is not alpha. A layer that beats the prop-implied projection is. That is why signals adjust the residual, not raw points, and why a layer that stops clearing the bar gets switched back off.
What we measure, and why those scores
This is a distribution product, so we grade the distribution. MAE is published too because the industry speaks MAE, but it only scores the median and would reward shading everything toward the middle. So it is never the headline.
The proper score for a full predictive distribution, not just the median. MAE grades only the middle point. CRPS grades the whole shape, which is the product. When CRPS comes in below MAE, that difference is exactly the value the distribution adds over a point guess.
How often the actual landed inside our P10 to P90 band, against an 80% target. If actuals fall inside far more or far less than 80% of the time, the bands are mis-calibrated and we say so. It is the most intuitive calibration stat there is.
Every boom% and bust% claim is scored with a Brier score, plus a skill score against simply predicting the base rate. Skill above zero means the probabilities beat guessing. A probability you can't score is just a number.
Start/sit verdicts carry a win probability, and we bin the claimed probability against the realized record. A 56% lean that loses is recorded as exactly that. The point is calibration, not a spotless win column.
Every projection is locked before kickoff
Each week's projections and start/sit verdicts are written to a hash-chained ledger before any game starts. The lock records a timestamp and the input snapshot, and each entry chains to the one before it, so a projection cannot be quietly edited after the fact. When the actuals come in, the scored report references the lock it grades and nothing in the lock is ever modified. The accuracy record is tamper-evident by construction, which is the whole point of an accountability-branded tool.
For every player we publish where the actual landed inside our distribution (the percentile), classified plainly as in range, too high, or too low, and the misses are diagnosed, not buried.
The live accuracy record starts filling in Week 1
We publish no accuracy numbers we have not earned. There is no back-tested hit rate on the homepage, because historical player props are effectively unobtainable, so the core distribution thesis has to be proven live rather than fitted after the fact. The scoreboard starts empty on purpose. Week 1 is where the CRPS, coverage, Brier, and calibration numbers begin, in public, including the weeks we get wrong.