How It Works
Five steps, compressed from the plain-English build log: simulate, forecast, model the field, optimize, and choose a portfolio — each with one true sentence and one honest caveat.
1. Simulate
Praedictum simulates each game play by play, not each player alone — when a quarterback and his top receiver connect, it happens inside the same simulated game, together, thousands of times over.
The simulator doesn't model downs, distance, sacks, or kicks. Every defense is graded degraded, on the record, because of it.
On any simulated Sunday, one play call decides who gets the target — if the WR1 catches it, the tight end didn't, the same real trade-off a defense forces. A quarterback's yards and touchdowns are literally the sum of what his receivers caught; there's no separate quarterback distribution that could disagree with them. Praedictum runs thousands of these Sundays, and the same seed reproduces the same Sunday, bit for bit, on any machine.
2. Forecast
Every player's projection is opportunity times efficiency — expected touches, and what he does with each one — shrunk toward what similar players usually do, and reported with a spread, never a single number.
This hasn't been checked against a real, played-out result yet. No full slate has run end to end here, and no accuracy claim is made.
Praedictum builds projections for running backs, receivers and tight ends; quarterbacks and defenses come out of the game simulation itself, not a standalone projection. Kickers aren't part of this format at all.
3. Model the field
Ownership — the share of the contest expected to roster each player — builds a simulated crowd of thousands of opposing lineups, which is then checked for how many would have built the exact same lineup you did.
Ownership here is a model, not a measurement. It isn't fit to real ownership data, and by design it can never read the optimizer's own output — no feedback loop.
The crowd is built one entry at a time — pick a quarterback by ownership, decide whether this entry stacks with him, fill the rest — then checked against DraftKings' own roster rules before it counts. Two of the inputs that control how the crowd behaves are supplied by an analyst with a stated reason, not a hidden default.
4. Optimize
An exact solver searches every legal lineup under the salary cap and the roster rules — rules checked field by field against DraftKings' own published rules page — and returns the best one it can find, honoring your locks and excludes.
“Best” only means the search finished. If it runs out of time first, the result is labelled feasible, not best, and the two are never shown as the same thing.
Locks and excludes are hard rules, not preferences the model can override, and they're kept separate from DraftKings' own game locks (a game that has already kicked off). Unverified rule sets exist for experiments, but a lineup built under one carries that label all the way to the export file.
5. Portfolio
A whole slate of lineups is chosen as a set — rewarding upside across the group and limiting how much of it leans on any one player — then compared against entering your single best lineup that many times over.
Every result today is scored against exactly one real, human-observed DraftKings contest. Nothing here says how it would score in a contest nobody has watched.
The comparison runs on the same simulated Sundays for every lineup in the set — a fix that closed a real gap where a group-level result could have quietly used a different Sunday for each lineup without anyone noticing.
None of this has been checked yet against a real result. No full slate has been run end to end here, and there is no accuracy claim anywhere on this page — that evaluation is next.