← Lineup builder

Tournament model

NFL-factor-1 · experimental assumptions, not a calibrated contest simulator. Projection is always the sum of imported fantasy-point projections, with the Captain multiplier where applicable. It is not relabeled as a median. GPP Score is a separate preference index.

What the score means

GPP Score = 0.30 × Projection + 0.70 × average score in the best 5% of simulated outcomes − ownership weight × geometric mean ownership. Default ownership weight is 5; 0 disables the penalty. These are transparent preference weights, not estimated payout coefficients. Correlation changes the simulated joint outcomes; there are no stack point bonuses.

The ownership product and its geometric mean are popularity/duplication proxies only. They assume independent roster selections, which is unrealistic for an actual field; we do not multiply them by field size or report predicted duplicate counts. Lower popularity supplies the leverage preference. We do not double-count it as a second bonus. Ownership is used as a fraction; zero estimates are floored at 0.1%. Showdown uses Captain ownership for the Captain and Flex ownership for the other players. If any eligible role lacks ownership, the penalty is disabled across the build.

Correlated outcomes

Players share normal latent factors for game pace, passing production, and game script. Each has independent residual variation. Paired teammates have opposite opportunity-allocation factors. This factor construction produces a valid positive semidefinite covariance structure without patching an arbitrary correlation matrix.

PositionGame paceOwn passingGame script*
QB0.300.60−0.10
WR / TE0.300.50−0.10
RB0.10−0.250.40
DST−0.2000.40
K0.150.150.10

*Game-script signs reverse between opponents. DST additionally loads −0.50 on opposing passing production. These are factor loadings, not final player correlations. Teammate WR/TE pairs and RB pairs share negative allocation covariance of 0.08 divided by their number of eligible competitors. Shared offensive success can still leave teammate receivers positively correlated overall. The model avoids assuming every teammate pair is either universally good or universally bad.

This creates positive QB–pass-catcher and game bring-back relationships, positive RB–DST relationships, negative opposing RB game-script effects, and negative DST–opposing-offense relationships. It cannot distinguish receiving backs, mobile QBs, depth-chart roles, weather, injuries, or specific team play styles from the currently available inputs. Coefficients are assumptions and should eventually be fitted to historical residual outcomes.

Cross-team factors require verified opponent pairs. Classic never guesses a matchup from team names. Use the matchup editor or an opponent CSV column. Showdown's two-team pool defines its matchup. Unrelated games share no factors.

Volatility and upper outcomes

Optional imported standard deviation in fantasy points takes precedence. Otherwise an upper-outcome point estimate supplies SD = (upper outcome − projection) / 1.28155, a normal-scale approximation. A provider's unspecified “upside” rating or percentage is not treated as a point ceiling. The simulated P90 is reported separately and is not guaranteed to equal the supplied upper outcome.

When neither input exists, SD is the larger of 2 points (4 for DST) and the following fraction of absolute projection: QB 0.32, RB 0.55, WR/TE 0.65, DST 0.75, K 0.40. These fallback assumptions are always disclosed. Marginals are shifted lognormal distributions matching the input expectation and chosen SD, with lower bounds of −2 (−4 DST), or projection minus 1 if lower. This is an approximate fantasy-score model, not a play-by-play scoring simulation.

Portfolio search and reproducibility

Six candidate portfolios come from the existing constrained optimizer: original projections, a volatility-adjusted objective, and four sampled correlated slate objectives blended with projection. Candidate search retains salary, roster, stacking, ownership, uniqueness, locks, flag exposure and Captain/Flex constraints. Classic candidate search is capped at 24,000–180,000 nodes per lineup depending on portfolio size, targeting at most 3.6 million nodes per seed portfolio; reaching that budget is disclosed.

Starting from the best complete feasible set, two swap passes improve 0.70 × expected best-lineup excess above a common target + 0.30 × average GPP Score. The target is the average simulated P90 of the first feasible seed set. Selection uses up to 1,024 training outcomes. Every swap preserves the full set's exposure and exact captain counts. The search is approximate and can miss a feasible or better portfolio.

Evaluation uses a separate seeded set of 2,048, 4,096 or 8,192 outcomes. Every evaluated lineup shares the same simulated slate. Reported portfolio probabilities mean at least one selected lineup crosses the score threshold; they are not sums of individual probabilities. Final display ranking uses the evaluation GPP Score. Repeating data, settings, model version and seed reproduces the results.

Sampling noise matters: at 4,096 outcomes the worst-case standard error for a probability is about 0.8 percentage points. Zero hits do not prove an event impossible (roughly 3/N is a 95% upper bound). Model misspecification may be much larger. Independent evaluation limits selection overfitting but is not historical validation.

Data currently available

The inspected dfs_projections schema contains id, scraped_at, site_updated_at, season, week, provider, slate, player, pos, team, fpts, salary, fpts_per_dollar, ownership_pct. It contains no ceiling, opponent, game, actual-result or Captain-ownership columns. Its latest view joins the maximum scraped_at per provider, not per slate. The CSV workflow validates one provider, slate, week and snapshot at a time. Select DraftKings and verify the export matches your contest; Thu–Mon data is not automatically a Sunday main slate.

CSV data stays in your browser. There is no new public database endpoint, live feed, database migration or permission change. DraftKings exports still validate player IDs, positions, salaries and slate membership. Assigning fewer entries than your full generated set can change exposure percentages and Captain allocations.

Research context

Positive passing correlations are supported by FantasyLabs' correlation research. Their discussion of correlation anomalies also explains why game script and competition matter. These references inform relationship directions; this implementation's coefficients are not fitted to those articles or represented as current empirical estimates.

No simulated opponents, payout structure or calibrated historical outcomes are available. The site therefore does not report tournament win rates, top-finish rates, expected ROI, or exact field duplication.