Track record
Every projection, graded after the games against the real box scores (and, when available, DraftKings' exact contest points). Wins and misses, nothing hidden.
Season ledger
The public GPP portfolio — the lineups built before kickoff every Sunday — graded week by week against the real DraftKings field. Built from the database by jobs/recap.py after each slate is scored; nothing here is typed in.
Methodology — what each column means
Consensus view (the number to watch): every lineup — ours and every real entry in the contest standings — is re-scored in the same simulated games on neutral market projections. Our own opinions are removed, so it measures how well the lineups are put together against what the field actually played. Expected ROI after rake; the field averages roughly −15% to −25%. “Beats X% of field” is where our portfolio's expected ROI sits among a random sample of real entries' expected ROI.
Model view: the same replay on our own projections. Model view minus consensus view, over many weeks, is how over-confident our projections are.
Realized: what actually happened on Sunday — average points vs the field's median entry, the share of real entries each lineup beat, best finish, and lineups that finished in the top 1%. One Sunday is mostly variance; the consensus column is the signal.
Dupes: exact copies of our lineups among the real entries. A prize is split across copies, so “winnings lost to splits” is the share of the payout curve given up to duplication.
Real $: the owner's own entries in the same contest, matched by DraftKings username in the standings export. Dollars are the contest's payout curve applied to the real finishing ranks at the recorded entry fee — a reconstruction from the standings file, not an account statement, and only the contests with an imported standings file are counted.
Members get the full projection-accuracy backtests, replayed the same way, in their account.
Loading…
Season to date
Average miss (MAE, in DraftKings points) for DFS-relevant players each week. Lower is better; r is the correlation between projection and result.
Which projection was closest?
Every projection source scored on the same players. sim = the Monte Carlo game simulation, baseline = a simpler recent-form model. Give it several weeks before trusting the ranking — one Sunday is noisy.
Every player
Proj = projection at kickoff. PIT = where the actual score landed inside the simulated range (0.5 = right on the median; near 0 or 1 = a bust or boom the sim thought unlikely). Own = real ownership when a contest file was imported.
Lineup scoreboard
Every lineup recorded before kickoff — the Sunday task's cash + GPP lineups and your own builder exports — scored with actual points. sim = where the actual landed in the lineup's simulated distribution; field = share of real contest entries it would have beaten (needs a standings file in data/ownership/).
Lineup record — season to date
Contest Flashback
Every real entry from the standings export is scored in the same simulated worlds as our lineups: "how much would these lineups make on average against these opponents". Consensus view puts every player on the market projection and judges construction alone; model view uses our projections. The field's own expected ROI is about −15% (the rake). One week proves nothing.
News agents — did the notes help?
Per-game news readers file sourced role/injury notes Saturday evening and Sunday morning; the lineup task applies the confident ones. helped = the adjusted projection was closer to the actual than the sim's.
Upset picks vs the market
The agents' win probability for every game next to the spread-implied market probability, scored by Brier (lower is better).