Advanced methods and game theory

Researched 2026-09-10. Format assumed throughout: 12-team, 20 rounds, weekly lineups, 1.5 PPR TE / 1.0 PPR others, 2 FLEX, ~4,500 teams, championship round = weeks 15-17 total points + regular-season PPG, ~20% of the field qualifies, $1M for 1st.

Conventions: "Evidence: strong / moderate / thin" tags reflect how much published, quantified data exists, not how much I agree. Numbers are quoted from the cited source; my own back-of-envelope calculations are labeled "(my estimate)". Several key sources (ETR, RotoViz, fantasymojo deep dives, Yvo Keuter's MCTS post) are paywalled or 403'd; where I could only get a search-engine summary I say so.

0. The contest math that should drive everything

The scoring formula. Official rules (FFPC official rules): Championship score = regular-season per-game average (weeks 1-12) + week 15 + week 16 + week 17. No head-to-head, no waivers in weeks 15-17, 20-man rosters. Roughly 85 top-2%-overall teams qualify automatically, plus every league's 1- and 2-seed and league champion; 849 teams (~20%) qualified in 2024 (fantasymojo 2025 retro).

Where the money is (from the 2026 payout table in research/rules.md; my arithmetic):

So: 1st place is 55% of the championship pool and top-3 is 72%. Conditional on qualifying (~900 teams), the expected championship prize is ~$2,000, and more than half of that is P(finish 1st) x $1M. Anything that raises P(1st | qualified) by 10% relative is worth about as much as raising P(qualify) by 5.5% relative, for the championship pool only. The league pool, by contrast, rewards being a good, consistent team (1-seed by record, 2-seed by points, league champion in weeks 13-14).

Variance decomposition of the championship score (my estimate). PPG over 12 weeks has a standard deviation across qualifiers of maybe 8-12 points (they are all good teams), whereas a 3-week total for a lineup scoring ~150/week has SD on the order of 45-60 points. PPG therefore contributes on the order of 3-5% of the variance in the final ranking; the ranking among qualifiers is ~95% a 3-week sprint. The PPG term matters mainly as a small head start, not as a variance lever.

Implication. This is a two-stage game with different objective functions:

1. Weeks 1-12: maximize mean weekly score and win probability (record and points both open doors). Mean-seeking, floor matters.
2. Weeks 15-17: maximize P(being the single highest 3-week total out of ~900). Variance-seeking, correlation-seeking, schedule-seeking.

ETR's Best Ball Mania manifesto (ETR) reaches the same structure for Underdog: "a 50% change in finals win rate is equivalent to a 1% change in regular-season advance rate" (their conservative alternative is a 4:1 ratio), and "we need to think of upside through the lens of playoff upside, not total points across the season." The FFPC ratio is less extreme than BBM's because FFPC pays a large league pool, but the championship pool itself is more top-heavy than BBM (55% to first vs. BBM's ~2M of a much larger playoff pool).

1. Game theory and optimization in drafts

1.1 MCTS / beam-search lookahead pick optimizers — Evidence: moderate (hobbyist), no peer-reviewed fantasy-football MCTS paper found

• Yvo Keuter, "Using Monte Carlo Tree Search for your Fantasy Football draft" and follow-up "Performance of MCTS for Fantasy Football drafts" (part 1, part 2). Both 403'd for me; per search summaries, the setup is a 10-team PPR snake draft, actions are chosen at the position level (then best-projected player at that position), opponents are simulated, and the reported result is that MCTS beats random and beats a greedy "take the position that adds the most value now" policy. No effect sizes retrievable. Treat as proof-of-concept.
• Ben Jensen, "Simulating the Snake: An AI-Assisted Fantasy Football Draft Strategy" (Medium) — 403'd; same genre.
• Stathole Sports, "We're Off To The Monte Carlo..." (Substack) — 12,000 simulated drafts (2022-24), 7-player lineups, control = realistic ADP-based opponents, success = top 8% of teams by a WAR-style metric. Ranking by top-8% rate: Elite QB/RB > Zero RB > Elite QB/WR > mid QB > Heavy RB > Late-round QB (worst). Second-round RBs averaged 1.06 WAR vs 1.49 for first-round RBs. Caveat: 1-QB, no TE premium, no FLEX x2, no bench, no weeks 15-17 weighting, so the specific strategy ranking does not transfer; the method (simulate the draft, score by tail probability rather than mean) does.
• Academic adjacent: A data-driven framework for team selection in Fantasy Premier League (belief-state MDP + Bayesian Q-learning, top-percentile vs millions of humans) and a Penn State honors thesis, An Optimization Approach to Fantasy Football Draft and Roster Management. Neither is directly usable, but both confirm the standard formulation: state = (available pool, all rosters, pick index); action = player; transition = opponent policy; reward = terminal roster value.

Practical takeaway. Full MCTS is overkill for one night. The bulk of the value comes from a 2-step lookahead with survival probabilities: for each position P, compute E[best available at P at my next pick] using per-player probabilities of surviving the intervening picks (from ADP mean/SD), and choose the pick that maximizes (value now) - (expected value forgone at the next turn). This is "dynamic VORP" and captures 80% of what tree search buys in a snake draft where you control 1 of every 12 picks.

1.2 Opponent modeling — Evidence: strong (one large behavioral study)

Drafting strategies in fantasy football: A study of competitive sequential human decision making (Judgment and Decision Making, 2022; Cambridge Core, open HTML). 1,350 Sleeper leagues, 2017, 188,426 picks (27% autopicks excluded). Findings that matter for an opponent model:

• Drafters are well calibrated to ADP for RB/WR (pick-order vs. performance correlation ~ -0.55) and poorly calibrated for QB/TE.
• Role selection is remarkably uniform across rounds: people draft to mirror the starting-lineup requirements (representativeness heuristic), then fill K/DST last.
• Herding is real and measurable: the probability of taking a QB (early), K, or DST rises sharply immediately after an opponent takes one; the effect is strongest for DST. It did not improve win rate (Bayes factors 4.7-8.8 favoring "no effect"), i.e., it is a bias you can exploit, not a strategy.
• Handcuffing essentially does not happen (<1% of expected pairs) and had no win-rate effect (51.0% vs 50.6%).
• Authors explicitly note that the empirical pick distributions (their Fig. 7) are a better opponent model than fixed-strategy assumptions and that drafters' shallow lookahead is exploitable.

Model to implement: P(opponent j takes player i) proportional to exp(-(ADP_i - pick)/tau) x need_j(pos_i) x run_bump(pos_i), where need_j = 1 + a x (unfilled starters at pos) and run_bump = 1 + b x (number of that position taken in the last k picks). Fit tau to the ADP SD in your ffpc_main_event_adp_2026*.csv; a and b can be estimated from the draft boards in data/boards/ (you have 27 of them) by regressing "was position X taken" on need and recent-run counts.

1.3 Positional-run dynamics — Evidence: thin (no rigorous study located)

Only heuristics were found (e.g., "runs accelerate after 3 consecutive picks at a position", fantasystrategyguide), plus the JDM herding result above, which is the only quantified evidence that the previous pick shifts the next pick's position probability. Estimate the run effect from your own FFPC boards rather than trusting a published number. The exploitable asymmetry is the one JDM identified: runs are strongest at QB/K/DST, where the drop-off in value is smallest, so the correct response to a run is usually to not join it.

1.4 Optimal stopping for QB/TE timing — Evidence: moderate, and format-dependent

No secretary-problem treatment of fantasy drafts exists in the literature I could find; the practical evidence is from best-ball advance-rate data:

• 4for4, "How Winners Draft Quarterbacks in BBM VI" (4for4): QBs with ADP in rounds 1-4 posted a 14.3% advance rate (2.4 pts below the 16.7% base) and 3.0% finals rate; "it was actually optimal to have zero QBs drafted through round 13"; 3-QB builds beat 2-QB in 4 of 6 BBM years; advanced QB metrics (BTT, EPA/play, scramble rate, PFF grade) correlated -0.01 to -0.10 with advance rate (nothing).
• Fantasy Footballers' Underdog construction study (FF): first QB before round 10 advanced 17.4% vs 11.8% for round 11+ (older BBM II data), i.e., the opposite direction. Two stud TEs early: 10.2% (bad).
• ETR manifesto: "investing in elite QBs had a clear edge in regular-season advance rates" but "playoff performance was unaltered by QB structure."
• Stathole simulation: late-round QB was the worst strategy, elite QB the best (1-QB, no TEP).

Reconciliation: elite QB helps the mean (qualification), not the tail (championship), and the sign flips year to year with which QBs hit. For FFPC 1-QB with weekly lineups, the QB decision should be a pure dynamic-VORP calculation (1.1) with the 15-17 weighting from Section 2, not a fixed rule.

TE in TE-premium: 4for4 TE-premium study (2015-23): average TE1-to-TE24 gap in 1.5 PPR = 148.7 pts vs WR 159.1 and RB 215.0; TE12 ADP in FFPC Main Event was 91.5, i.e., WR41/42 territory, and "24 WRs are coming off the board right after the seventh tight end." Draft Sharks FFPC guide: among the top-24 scorers under FFPC scoring in 2021 there were 12 RB, 9 WR, 3 TE; recommends two WRs in the flex and delaying the second TE/QB to ~round 6. The data say elite TEs are priced roughly fairly in FFPC (the premium is already in ADP); the exploitable piece is the second TE as a FLEX starter, which the market prices as a backup.

1.5 Minimax regret across projection sources — Evidence: thin (no fantasy-specific empirical work)

The general theory is in Minimax Regret Optimisation for Robust Planning in Uncertain MDPs and Distributionally Robust Regret Minimization; FFA has a practical piece on projections and uncertainty and an optimization simulation. Nobody has shown minimax-regret pick selection beats expected-value selection in fantasy. My view: minimax regret is a conservative criterion and is the wrong objective for the championship-pool half of this contest; use it only (a) as a tiebreak among near-equal picks or (b) for the floor-oriented weeks 1-12 objective. Cheaper and better-supported: use the spread across sources as the per-player uncertainty input to a simulation (Section 2.4).

1.6 Bayesian shrinkage of projections toward market — Evidence: strong

• Harstad, "Fantasy, in Theory: Bayes and Bob" (Footballguys) 2015 test vs rest-of-season results: ADP alone r = 0.548; early-season performance alone 0.659; 50/50 average 0.697; expert (Bob Henry) 0.703; best fitted blend 43.2% expert / 35.7% ADP / 21.1% early-season, r = 0.721. Conclusion: "we're probably never going to be able to explain significantly more than 50% of the variation ahead of time."
• Fantasy Football Analytics, 12 seasons of projections (FFA 2026): preseason projections explain only ~14-26% of within-position variance (RB most, QB least); the FFA average beat individual sources in 69% of head-to-heads; equal weighting beat accuracy-weighting (47.4 vs 47.8 MAE) because source accuracy does not persist year to year. Position MAEs (season points): QB ~61, RB ~52, WR ~40, TE ~31.
• Nathan Braun, Bayesian fantasy football (writeup): ADP as prior, weekly results as likelihood.
• RotoViz's projection philosophy is explicitly "historical points scored at this ADP" as the prior (RotoViz tools).

Implement: final_mean = 0.6 x equal-weight average of your sources + 0.4 x ADP-implied points (historical mean points by ADP slot and position). The 40% market weight is Harstad's ~36% rounded up because you have no early-season data yet. Use the between-source SD as one input to per-player uncertainty.

2. Tournament theory for a top-heavy season-long contest

2.1 Ceiling vs mean under this payout — Evidence: strong in best ball, transferable by structure

• ETR manifesto (above): playoff-week upside dominates; teams with 14+ "live" players in the final week had dramatically higher EV; optimal drafting window mid-July to mid-August (irrelevant now, but the "live players" point is not: avoid week 15-17 risk such as late byes-adjacent injury returns, teams likely to rest starters, and rookies on short leashes).
• Illustration of why variance wins the sprint (my estimate). If the ~900 qualifiers' 3-week totals are ~N(mu, sigma) and you can raise your own SD by 14% (the measured effect of a QB-WR stack, below) at equal mean, P(exceeding the field's 99.9th percentile) goes from 0.10% to P(Z > 3.09/1.14) = 0.34%, a 3.4x increase in P(1st). Even with a 2-3 point mean penalty, the tail probability still roughly doubles. The same logic argues for concentrating your championship lineup in high-total games and domes.
• Draft-round hit rates are the other half of "ceiling": see Section 4.1 for the round-by-round top-12 rates that should feed the tail of each player's distribution.

2.2 Ownership / leverage in season-long — Evidence: thin, but the structure is decisive

• In DFS, leverage = ceiling x (1 - ownership) (Stokastic, 4for4).
• Structural fact for season-long: each player exists once per 12-team league, so no player can be on more than 1/12 = 8.3% of all rosters. Ownership among qualifiers can be higher (a player who doubles his teams' qualify rate is on ~16% of championship rosters), but nothing approaches DFS-style 40% chalk. Uniqueness is therefore mostly automatic at the player level; the leverage decisions are at the construction level (which stacks, which week 15-17 game environments, which second TE).
• Underdog's own research (Uniqueness in Best Ball Tournaments) concludes uniqueness is "a tertiary thought" behind fundamentals; a 4for4/Underdog summary notes less-unique teams were over-represented in finals. In a total-points sprint the correct framing is: being different only helps if the different player has equal or better ceiling.
• fantasymojo's FFPC Data Warehouse publishes "Player Advance Rates: ownership percentages of championship-round roster players" and a Stack Explorer for the Main Event (2025 deep dive index, Big Gorilla preview). The tables are premium; if you have access, that is the only FFPC-specific ownership-among-qualifiers dataset I found.

2.3 Correlation / stacking in season-long — Evidence: strong

• RotoWire, four seasons (2022-25), 1,300+ team-stack seasons (RotoWire): weekly correlations QB-WR1 +0.31, QB-TE +0.27, QB-RB +0.07, same-team WR-WR -0.02. A QB-WR stack adds ~1.8 pts to the weekly ceiling and costs ~1.6 on the floor; QB-WR-WR: +2.2 / -2.0; stacks are ~14% more boom-or-bust; season total is unchanged (correlation redistributes variance, it does not add points).
• Underdog/Fantasy Footballers (FF Underdog guide summary): QB1-WR1 team stack +4.7% advance rate over expectation; QB1-WR1-WR2 stacks reached finals 2.6x expected (2023); 3 team stacks -> 2x finals rate vs 0 stacks; 3 stacked QBs +23% finals vs zero stacked -26%. Bring-backs: when a QB scores 25+, the opposing QB1 is 13.7% more likely to also boom and the opposing WR1 11.3%.
• 4for4 BBM VI: 3 team stacks -> finals twice as often as 0; the champion game-stacked four week-17 matchups; more game stacks -> higher finals totals.
• ETR: game stacking in the final week "+50% estimated win rate"; 6-9 game-stacked players in week 17 optimal in BBM (that is an 18-man best-ball roster; scale down for an 11-man weekly lineup).

For FFPC: stacking is neutral-to-slightly-negative for weeks 1-12 (weekly H2H floor) and clearly positive for the 3-week sprint. Since you set weekly lineups, you can build the stack into the bench and only start it in weeks 15-17 (e.g., a QB2 who stacks with your WR1 and plays in a dome in weeks 15-17). Prefer QB-WR and QB-TE stacks (r ~ 0.3) over QB-RB (r ~ 0).

2.4 "Championship odds" season simulation — Evidence: methodology mostly undisclosed

None of ETR, RotoViz, 4for4, or Fantasy Points publishes the internals of their season-long championship-odds models; ETR's BBM equity model is the most transparent: EV = sum over stages of P(reach stage) x P(win stage) x payout, with multiplicative adjustments for live-player %, ADP-value percentile, and stack configuration (worked example in the manifesto: ~$133 baseline EV per quarterfinal entry). Public playoff-odds tools (ffwrapped, spreadsheetsolving) all do the same thing: per-team weekly score distributions, 5,000-10,000 Monte Carlo seasons over the real schedule.

A minimal FFPC simulator that captures the structure (2-3 hours):

1. Per player per week: score ~ Normal(mu_w, sd_w) truncated at 0, with mu_w = season mean x matchup/venue multiplier (Section 3) and sd_w from position (RB/WR weekly SD is roughly 60-80% of mean; use your own data); an injury hazard per week (~2-4% for RB, ~1.5-3% WR) with a duration draw.
2. Correlate same-team QB/WR/TE draws with rho ~ 0.3 (Gaussian copula).
3. Weekly lineup = best 11 by projection (not realized), which is what you actually do.
4. Stage 1: simulate your league (11 ADP-typical opponents) to get P(1-seed), P(2-seed), P(league champ), P(top-2% points) -> P(qualify).
5. Stage 2: simulate ~900 qualifier teams as ADP-typical good teams; compute your 3-week total + PPG rank -> P(1st), P(top-3), P(top-200).
6. EV = league pool terms + championship pool terms. Rank draft candidates by marginal EV, not by projected points.

2.5 Roster construction evidence you can lean on (best ball, so weight accordingly)

• RB timing flips year to year. BBM VI (2025): RB in round 1 -> ~24% advance vs ~12% without; RB-RB start ~30% vs ~8% with no RB in rounds 1-2; early RBs 10.1% finals rate (+2%) (4for4 RB). BBM II/2020 data said the opposite: the single best construction was 7 RBs with zero in the first 5 rounds (19.4% advance), and 4-RB builds with no early RB were 9.1% (Fantasy Footballers). ETR's multi-year read: lower RB draft capital + 5-6 RBs = best regular-season advance rates. Lesson: structure results are dominated by which players hit that year; do not hard-code Zero/Hero RB, let the simulator price it.
• WR count: 6+ WRs > 16.5% advance, 5 ~15%, 4 10.9%, 3 4.0% (18-man best ball).
• TE count: 3 TEs 18.0%, 2 16.1%, 1 11.7% (best ball; in FFPC weekly lineups with TE premium, 2-3 TEs is also what the FLEX math suggests).
• FFPC-specific: RotoViz's FFPC Roster Construction Explorer and Best Ball Win Rate Explorer (tools, 2020 extreme Zero-RB FFPC piece) are the only FFPC-format construction datasets; both paywalled.

3. Environmental factors and how much they matter

3.1 Weather — Evidence: strong for direction, moderate for magnitude

Dome vs outdoor (DraftKings Network, 2020-24): team passing fantasy 12.96 (outdoor) vs 14.29 (indoor) per game (+10%); rushing 16.51 vs 17.89 (+8%); team receiving PPR 52.13 vs 55.68 (+7%). Fantasy Footballers: combined scoring 42.4 outdoors vs 46.2 in domes/retractables (+9%); books already price it (avg O/U 44.6 indoor vs 41.7 outdoor). Kickers: no dome advantage (Fantasy Index).

Caveat on all of the above: dome/weather effects are partly confounded with which teams play in domes; use the Vegas total for the actual game (which already embeds venue, weather expectation, and team strength) as the primary signal and weather only as a late (week-of) adjustment. In weeks 15-17 the decision-relevant items are: (a) games in domes/warm sites, (b) probability of 20+ mph wind or sub-25 F cold at BUF/CHI/GB/CLE/PIT/NE/DEN/KC/NYJ/NYG/PHI/BAL, (c) kickers on outdoor cold teams.

3.2 Home/away — Evidence: thin

Home teams win ~57% (FiveThirtyEight); the fantasy-point differential is small and I could not retrieve a clean per-position estimate (the one Medium study, Fantasy Outliers, was blocked). Home/away is already inside the spread; do not double count.

3.3 Opponent defensive strength and preseason SOS — Evidence: strong that it is weak

• 4for4, "Do Defenses Repeat?" (2014-24) (4for4): year-over-year correlation of fantasy points allowed: vs QB 0.26, vs RB 0.23, vs TE 0.16, vs WR "very little." Top-5 defenses vs a position repeat 20-30% of the time with an average next-year finish around DEF12-14; bottom-5 repeat 16-32% (avg DEF19-21).
• Preseason SOS built on last year's points allowed is therefore near-noise for the season (Fantasy Football Blueprint says it is "nearly worthless for the season, useful in two spots: playoff-week planning and streaming"). ESPN's Mike Clay uses roster evaluation rather than last year's results for team-level SOS (ESPN 2026).
• Live proof of the noise: for weeks 15-17, ESPN rates WAS/NO/ARI easiest and PHI/SEA/SF hardest, while FTA's schedule-adjusted model (FTA playoff SOS) ranks WAS 31st (hardest) and JAX 1st. Use the Vegas game totals for the specific weeks instead; ETR's playoff-schedule piece (ETR) explicitly uses posted week 15-17 totals because "the books stand to lose real money if they hang bad numbers."

3.4 Vegas totals / win totals as a predictor — Evidence: strong direction, no published r for season-long

• Team implied total is the standard DFS predictor ("fantasy points are most strongly correlated to team implied totals", Fantasy Footballers); I could not find a published correlation coefficient for season-long output vs preseason win total.
• PFF, Vegas win totals and game script (PFF): over three seasons, QBs on winning teams averaged 18.4 FPG vs 14.2 on losing teams; RBs 26.2 vs 20.6 (team-level); but 50% of QB scoring happens while trailing (46% of plays) and RB targets are 52% while trailing vs 29% while leading. Fantasy Points' game-script pieces (2021, 2022) make the same point: win totals predict team scoring, and trailing game script predicts pass-catcher volume. So for pass catchers, a high team total on a team that will also trail (high total, underdog) is the best environment; for RBs, high total + favorite.
• The best single number for weeks 15-17 is the posted game total/spread for that game; data/vegas_2026.csv carries season win totals (DraftKings) and season implied PPG (Sharp) as the fallback until week 15-17 lines are posted.

3.5 Pace / plays per game — Evidence: moderate

Hayden Winks (NBC): year-over-year changes in plays/game are mostly 0-1 except with coaching changes (+/-2 to 4); neutral-situation pace ranged 1.81-2.40 plays/min; every team ran >= 2.0 plays/min when trailing by 7+. PFF on neutral pass rate: prior-year neutral pass rate beats raw pass rate for projecting volume; league neutral pass rate is at a multi-year low (~57.4%). Use pace x neutral pass rate x team total as the volume prior; the marginal information beyond the Vegas total is modest.

4. Player-level forecasting signals

4.1 Base rates by ADP round — Evidence: strong

Fantasy Squawk, ADP vs reality 2015-2025 (source); hit = top-12 at position, bust = outside top-36:

Other base rates from the same study: preseason top-12 players who finish top-12: RB 54%, WR 48%, QB 38% (of top 6), TE 48% (of top 6). Consensus No. 1 at a position finished top-12 in 31 of 32 seasons when healthy (<= 3 games missed) and 1 of 12 when injured; the No. 1 overall pick finished No. 1 at his position once in 11 years. Takeaways: round-1 WRs both hit more and bust less than round-1 RBs; the "RB dead zone" (rounds 5-9) is visible (RB top-12 rate 5-14%); and the dominant risk at the top of the draft is injury, not talent.

4.2 Breakout indicators — Evidence: strong

• Year in league / age / draft capital (Fantasy Footballers, 2000-21, 117 WR breakouts = first top-24 PPR PPG season, source): 77.9% of WR breakouts occur by year 3; 91.5% by age 26; 85.5% were drafted in NFL rounds 1-3; hit rates by NFL round: R1 56.1%, R2 25.3%, R4-7 4.8%; year-2/3 breakouts averaged 9.6 PPR PPG the season before breaking out.
• Draft capital hit rates (FF Trading Room, WRs 2013-21, source): "great or better" careers R1 37%, R2 22.5%, R3 8%, R4 3%; the R2->R3 drop is bigger than R1->R2. Fantasy Footballers draft-capital study (2000-18, first three seasons, source): RB1 rate R1 25.5% / R2 14.0% / R3 6.0% / R4 1.4%; RB2 rate R1 55.3% / R2 28%; WR1 rate R1 5.3% / R2 2.4%; WR2 rate R1 18.4% / R2 10.7% / R3 3.7%; TE top-12 rate R1 26.1%. First-round RBs average 17.1 opportunities/game vs 10.9 for day-2 RBs.
• Age curves for the up-slope (ESPN peak/decline study): WR PPR PPG +43% from year 1 to year 2 (since 2000) and another +24.5% at year 5 vs year 4; TE +98.5% from year 1 to year 2 and then flat through year 7.
• Sticky opportunity metrics (Fantasy Classroom, 2012-22, source): for starting WRs, year-over-year r >= 0.70 for target share, air-yards share, and WOPR (1.5 x target share + 0.7 x air-yards share); receiving TDs r ~ 0.40-0.54 (the flukiest stat); separation/cushion metrics have "extraordinarily poor" correlation with fantasy points.
• Vacated targets: widely used, but the only evaluative statement I found (Fantasy Footballers) is that vacated targets are "descriptive... not in the least bit predictive"; no hit-rate study exists. Use prior-year target share plus draft capital instead of vacated-target arithmetic.
• New play-caller: 17 offenses have some play-calling discontinuity in 2026 (CBS, 4for4 tendencies), but I found no published effect-size study; treat as a widening of the distribution (more variance), not a shift in the mean, unless the coordinator's neutral pass rate / pace history says otherwise.

4.3 Bust / downturn indicators — Evidence: strong for age and TDs, mixed for workload

• RB age. ESPN: 25.2% PPG drop from age 28 to 29 (since 2000); among 11 RBs since 2010 reaching 28-29: -16.8% (26->27), -8.4% (27->28), -15.3% (28->29); RBs 28+ get only 13.8% of league touches now (vs 16.4% in 2012). PFF (Narrative Street, 2010-16): for 150+ carry RBs, PPG peaks at 25, holds through 28, drops at 29; for all RBs the wall is age 31, not 30. Fantasy Footballers lifecycle: 8.7% of peak seasons at 28, 5.2% at 29, 3.8% at 30. Rule: the RB cliff is the age-29 season for workhorses; age 28 is a yellow flag, not a red one.
• WR age. Fantasy Life (2011-23, source): prime = 24-27 (100% of career-best 3-yr stretch); 28-29 comparable to years 21-23; 31+ is 30-40% below prime. Elite WRs (3+ WR1 seasons): at 30, 16.4 PPG and 0% cliff rate; ages 31-32 ~50% cliff risk; age 33, 0% WR1 finishes. "Good" WRs (3+ top-24, <3 WR1): coin-flip cliff risk at every age 30+. Rule: elite WRs are fine at 30, coin-flips at 31-32; non-elite WRs are coin-flips at 30.
• TE age. Peak 27.45 (Apex); elite TEs' average "age-out" season (last with 50+ catches) is 33.55 (Footballguys); FantasyPros/ESPN: gradual decline after 30, cliff after 32. Rule: TE 31+ is a mild flag, 33+ is a real one.
• QB age. Rushing fantasy production falls 25.7% at age 27 vs 26, plateaus 27-29, and is down ~2/3 by 30 and ~80% by 33 (ESPN, Dynasty Edge); pocket passers age fine.
• Workload (the "Curse of 370"). Fantasy Life, 30 RBs with 350+ touches 2015-24 (source): next season 16.6 PPR PPG (avg RB14.7, median RB12), only 4 of 30 (13%) improved their PPG, YPC 4.7 -> 4.1, games played avg 13.3 / median 16, 12 of the last 13 played 13+. PFF, 21 RBs with 400+ touches 2006-20 (source): 5 of 19 (26%) failed to play 14 games. Harstad (Footballguys) shows the 370 threshold was cherry-picked and that, 2009-13, 300+ carry top-12 RBs repeated as top-12 more often (53%) than <300 carry top-12 RBs (39%). Rule: big workload = expect regression to ~RB12-15 and a ~25% chance of a 3+ game absence, but that is mostly ordinary regression from a career-best year, not a workload curse.
• TD regression. Expected TDs (xTD) predict next-year TDs better than actual TDs (r-squared 0.283 vs 0.275) and are far stickier (0.382 vs 0.276) (reported in the FanDuel/Fantasy Life TD-regression explainers, FanDuel Research; the Fantasy Points xTD article is paywalled). Of players with 10+ TDs in a season over the past decade, only 11% scored more the next year, losing 5.2 TDs on average; of the top-25 positive TD-differential seasons, one player scored more the next year (avg -52%), and 18 of the bottom-25 improved (avg +63%) (same explainers; attribution to a specific original dataset is unverified). Rule: replace last year's TDs with xTD in every projection input.
• New team. PFF free-agent study since 2010 (source): share that beat ADP after switching teams: WR 35% (but only 31% improved PPG), RB 6 of 35 (17%), TE 0 of 15, QB 3 of 11. Rule: discount TEs and RBs on new teams the most; WRs least.
• QB change. No effect-size study found; the mechanism (target share is quarterback-relationship dependent) is well documented (PFF catchable target rate). Treat as added variance.
• Injury history. Draft Sharks sells an ML injury predictor (about); no public validation. Sports-medicine ML in soccer reaches AUPRC ~0.83 with daily load data (PMC) that you do not have. The only strong fantasy-relevant base rate is Fantasy Squawk's: the injury, not the talent, is what busts top picks. Use position x age hazard rates rather than player-specific injury narratives, except for chronic soft-tissue or in-season status.

5. 2026-specific data

5.1 Vegas win totals and implied scoring (all 32 teams) -> data/vegas_2026.csv

Win totals are DraftKings season O/U as published by FOX Sports on 2026-09-09; May openers from Yahoo (2026-05-16); PFF model wins from PFF (2026-07-06); implied season PPG from the Sharp Football implied-team-totals tool as displayed 2026-09-10 (their "Pts/Game" column; I could not confirm whether it is a pure projection or blends Week 1 lines, so treat it as a ranking more than a level).

Notable September moves vs May: CIN +1, NE +1, GB -1, SF -1, LAC -1, JAX -1, CLE -1, ARI -1, MIA -1. Implied-PPG ranking (Sharp) top 8: LAR, BUF, DET, CIN, DAL, BAL, SF, GB; bottom 5: CLE, NYJ, ARI, MIA, LV.

5.2 Weeks 15-17 schedule and published strength grids

Opponents are in data/schedule_w15_17_2026.csv (matches ETR's and FTA's tables). Published ratings:

• FTA schedule-adjusted playoff SOS (higher = easier), all 32 (FTA): JAX +7.9, MIN +6.9, LAR +5.5, TEN +4.9, BAL +4.6, CLE +4.0, NO +3.4, NYG +2.4, PIT +2.0, KC +1.6, BUF +1.3, CAR +1.1, LAC +0.8, DAL +0.7, ATL +0.6, CHI +0.6, DEN +0.5, ARI +0.4, CIN +0.2, DET -0.5, LV -0.7, IND -0.7, HOU -1.0, GB -1.3, TB -1.7, SEA -2.6, PHI -2.9, MIA -3.5, NYJ -4.4, NE -7.3, WAS -9.3, SF -9.3. (Per-opponent values in the same table: e.g., DAL's defense +22.5, MIN -20.5, HOU -14.1, PHI -13.8, WAS +14.2, CIN +13.0.)
• ESPN (Clay, roster-based): easiest weeks 15-17: WAS, NO, ARI; hardest: PHI, SEA, SF. Note the direct contradiction with FTA on WAS (Section 3.3).
• Posted game totals for weeks 15-17 quoted by Fantasy Life (source): DAL @LAR 52.5, DAL-JAX 51.5, DAL-NYG 49.5 (all domes); DET @MIN 46.5, DET-NYG 48.5, DET @CHI 49.5; LAR-DAL 52.5, LAR @SEA 47.5, LAR @TB 48.5; CIN-CAR 47.5, CIN @IND 52.5, CIN-BAL 51.5; CHI 51.5/47.5/49.5 (all outdoor cold); LV three domes but totals <= 42.5. ETR's full grid of implied team totals by week is paywalled (ETR).
• Venue count, weeks 15-17 (my tally from the schedule; dome/roof teams = ARI, ATL, DAL, DET, HOU, IND, LV, LAR, LAC, MIN, NO): 3 dome games: DAL, LV, ARI. 2 dome games: ATL, DET, JAX, NYG, LAC, NO, MIN (DET @MIN, vs NYG home, then @CHI outdoors). 3 cold-outdoor games: CHI (@BUF, GB, DET), GB (MIA, @CHI, HOU), NE (@KC, @NYJ, DEN), PIT, CLE, BUF (CHI, @DEN, @MIA - the MIA game is warm), NYJ (2 cold + ARI dome).

6. Recommended for a one-night implementation (ranked by expected impact per hour)

1. Two-objective player value (1 hr, highest impact). value = w1 x proj(weeks 1-12) + w2 x proj(weeks 15-17) x env_mult, where env_mult comes from the posted week 15-17 game totals (or, until posted, Sharp implied PPG and dome flags: +7-10% for dome games, -5 to -8% for likely sub-25 F games, kickers -1 FG-equivalent in cold). Set w2 so that weeks 15-17 carry roughly 40-50% of total value for the starters you expect to have in that lineup (justification: Section 0, the championship pool is 55% first place and 95% 3-week variance).
2. Dynamic VORP with survival probabilities (1.5-2 hrs). From your FFPC ADP mean/SD, compute P(player available at my next pick); pick to maximize value_now minus E[best value at position at next pick]. This replaces static VORP and captures most of what MCTS/beam search would add.
3. Projection blend + ADP shrinkage + xTD (1 hr). Equal-weight your projection sources, shrink 35-40% toward ADP-implied points, and replace TD components with xTD-based expectations. Store between-source SD per player.
4. Age / regression / new-team flags as distribution adjustments (45 min). RB age-29+ workhorse: -10 to -15% mean and higher injury hazard; 350+ touches: cap mean at ~RB12-15 equivalent (only 13% improve); WR 31-32 non-elite: 50% cliff mixture; TE 33+: mixture; new-team TE/RB: -10% mean; year-2/3 WRs with R1-R2 draft capital and 9-10+ PPG last year: fatten the right tail (77.9% of breakouts occur by year 3, R1 hit rate 56%).
5. Stack bonus, championship-weighted (30 min). Add value for QB+WR / QB+TE same-team pairs proportional to their week 15-17 game totals (r ~ 0.3), and zero for QB+RB. Prefer to hold the stack QB as QB2 if it costs mean in weeks 1-12.
6. Opponent model with need and run terms (1 hr, medium). ADP-softmax x roster-need x recent-run bump, parameters fit on data/boards/. Exploit the JDM result: do not chase QB/K/DST runs.
7. Lightweight two-stage Monte Carlo (2-3 hrs, high impact but only if time remains). Section 2.4 design; output P(qualify), P(top-200), P(1st), and marginal EV per candidate. Even a crude version will reorder the top-60 meaningfully versus mean projections.
8. Skip tonight: full MCTS/beam search (marginal over item 2), minimax regret (wrong objective for this contest), ownership/leverage modeling (ownership is capped at 1/12 per player; the structural uniqueness is automatic), preseason SOS beyond the week 15-17 totals (year-over-year r = 0.16-0.26).

Sources

Game theory / optimization

• Keuter, MCTS for fantasy drafts: https://medium.com/data-science/using-monte-carlo-tree-search-for-your-fantasy-football-draft-6509b78a1c20 ; performance: https://medium.com/@ykeuter/performance-of-monte-carlo-tree-search-for-fantasy-football-drafts-269075d08088
• Jensen, Simulating the Snake: https://bcjense6.medium.com/simulating-the-snake-an-ai-assisted-fantasy-football-draft-strategy-4064c98940f7
• Stathole Sports Monte Carlo draft strategies: https://statholesports.substack.com/p/were-off-to-the-monte-carlo-to-find
• JDM 2022 drafting behavior study: https://www.cambridge.org/core/journals/judgment-and-decision-making/article/drafting-strategies-in-fantasy-football-a-study-of-competitive-sequential-human-decision-making/2AB841B3F446833348D784C0FC54DAD2 (open: https://www.sas.upenn.edu/~baron/journal/22/220318/jdm220318.html)
• FPL data-driven framework (arXiv): https://arxiv.org/pdf/2505.02170
• PSU thesis, optimization approach to drafts: https://honors.libraries.psu.edu/catalog/6788dos5531
• Harstad, Bayes and Bob: https://www.footballguys.com/article/HarstadFiT3
• Braun, Bayesian fantasy football: https://nathanbraun.com/bayesian-fantasy-football/
• FFA 12-season projection accuracy: https://fantasyfootballanalytics.net/2026/08/we-analyzed-12-seasons-of-fantasy-football-projections-heres-what-we-found.html ; uncertainty: https://fantasyfootballanalytics.net/2024/09/fantasy-football-projections-and-uncertainty.html ; optimization sim: https://fantasyfootballanalytics.net/2013/08/optimization-simulation.html
• Minimax regret MDPs: https://arxiv.org/pdf/2012.04626 ; DR regret: https://arxiv.org/pdf/2412.15406
• 4for4 TE premium value: https://www.4for4.com/2024/preseason/understanding-tight-end-value-te-premium-fantasy-leagues-ffpc
• Draft Sharks FFPC first 6 rounds: https://www.draftsharks.com/article/ffpc-main-event-strategy-guide--how-to-attack-the-first-6-rounds-2022

Tournament theory

• FFPC official rules: https://myffpc.com/cms/public/play/main-event-official-rules/ ; rules explained: https://myffpc.com/cms/public/play/main-event-rules-explained
• fantasymojo FFPC Data Warehouse: 2025 retro https://www.fantasymojo.com/index.php/2025/05/29/2025-main-event-retro/ ; 2025 championship deep dive https://www.fantasymojo.com/index.php/2025/11/25/2025-main-event-and-fantasypros-championship-round-teams-deep-dive/ ; Big Gorilla 2026 https://www.fantasymojo.com/index.php/2026/04/12/2026-ffpc-big-gorilla-preview/
• ETR Best Ball Mania Manifesto: https://establishtherun.com/best-ball-mania-manifesto-a-guide-to-winning-big-on-underdog-fantasy/
• Underdog Network, Uniqueness: https://underdognetwork.com/football/best-ball-research/uniqueness-in-best-ball-tournaments
• RotoWire stacking data 2022-25: https://www.rotowire.com/football/article/does-stacking-work-in-fantasy-football-what-four-years-of-data-say-about-drafting-correlated-players-2026-131409
• 4for4 BBM VI QBs: https://www.4for4.com/2026/preseason/how-winners-draft-quarterbacks-underdog-best-ball-mania ; RBs: https://www.4for4.com/2026/preseason/how-winners-draft-running-backs-underdog-best-ball-mania
• Fantasy Footballers Underdog construction: https://www.thefantasyfootballers.com/articles/optimal-roster-construction-on-underdog-fantasy-fantasy-football/
• Fantasy Points Underdog guide (stack advance rates): https://www.fantasypoints.com/nfl/articles/2024/underdog-fantasy-best-ball-draft-guide
• RotoViz tools / FFPC explorers: https://www.rotoviz.com/tools-2/ ; https://www.rotoviz.com/2020/08/extreme-zero-rb-tactics-in-the-ffpc-main-event/
• Stokastic leverage: https://www.stokastic.com/articles/nfl-dfs/nfl-dfs-leverage-game-theory ; 4for4 ownership: https://www.4for4.com/2018/preseason/leveraging-ownership-nfl-dfs
• Playoff-odds Monte Carlo examples: https://ffwrapped.com/fantasy-football-playoff-odds-calculator ; https://spreadsheetsolving.com/fantasy_football_monte_carlo/

Environment

• Fantasy Life wind: https://www.fantasylife.com/articles/best-ball/does-wind-matter-in-fantasy-football
• PFF wind/completion: https://www.pff.com/news/fantasy-football-the-factors-week-14-2017 ; PFF dome QBs: https://www.pff.com/news/fantasy-football-the-factors-week-2-2017
• Sharp Football weather: https://www.sharpfootballanalysis.com/sportsbook/weather-impact-on-nfl-betting/
• The Spax weather: https://www.thespax.com/nfl/analyzing-the-effect-of-weather-in-the-nfl/
• Fantasy Footballers weather mythbusters: https://www.thefantasyfootballers.com/articles/the-fantasy-football-mythbusters-whether-weather-really-matters/ ; dome advantage: https://www.thefantasyfootballers.com/articles/the-dome-field-advantage-in-fantasy-football/
• AFA temperature and FGs: http://www.advancedfootballanalytics.com/2012/01/temperature-and-field-goals.html
• DraftKings Network dome vs outdoor 2020-24: https://dknetwork.draftkings.com/2025/09/02/how-does-playing-in-a-dome-vs-outdoor-stadium-impact-fantasy-football-production/
• Fantasy Index kickers in domes: https://fantasyindex.com/2015/07/07/factoid/kickers-in-domes
• 4for4 defenses repeat: https://www.4for4.com/2025/preseason/do-defenses-repeat-fantasy-football-performances
• Fantasy Football Blueprint SOS: https://www.fantasyfootballblueprint.com/2026/08/07/strength-of-schedule/
• ESPN 2026 schedule strength: https://www.espn.com/fantasy/football/story/_/id/48755471/fantasy-football-2026-nfl-schedule-strength
• FTA playoff SOS 2026: https://fantasyteamadvice.com/nfl/strength-of-schedule/playoff
• ETR 2026 playoff schedules: https://establishtherun.com/fantasy-football-playoff-schedules-3/
• Fantasy Life playoff schedules to target: https://www.fantasylife.com/articles/fantasy/4-fantasy-football-playoff-schedules-to-target
• PFF Vegas win totals and game script: https://www.pff.com/news/fantasy-football-metrics-that-matter-vegas-win-totals ; Fantasy Points game script: https://www.fantasypoints.com/nfl/articles/season/2021/vegas-and-2021-gamescript-dependency
• Fantasy Footballers implied totals: https://www.thefantasyfootballers.com/articles/over-undervalued-fantasy-football-players-based-on-vegas-implied-totals/
• Winks pace: https://www.nbcsports.com/fantasy/football/news/article-going-deep-projecting-plays-game-and-offensive-pace ; PFF neutral pass rate 2026: https://www.pff.com/news/projecting-2026-fantasy-volume-through-neutral-situation-pass-rates
• FiveThirtyEight home field: https://fivethirtyeight.com/features/fans-are-back-at-nfl-games-but-home-field-advantage-isnt-yet

Player signals

• Fantasy Squawk ADP vs reality 2015-25: https://fantasysquawk.com/blog/adp-vs-reality-2015-2025
• Fantasy Footballers WR breakouts: https://www.thefantasyfootballers.com/analysis/anticipating-the-breakout-wide-receiver-trends-breakout-candidates-fantasy-football/ ; draft capital: https://www.thefantasyfootballers.com/articles/draft-capital-its-correlation-to-early-career-fantasy-production/
• FF Trading Room WR draft-capital hit rates: https://www.fftradingroom.com/362/Hit-Rates-for-Dynasty-Wide-Receivers-Based-on-NFL-Draft-Capital
• Fantasy Classroom WR sticky stats: https://fantasyclassroom.org/Blogs/sticky-stats/wr-sticky-season-totals
• ESPN peak/decline ages: https://www.espn.com/fantasy/football/story/_/id/37933720/2023-fantasy-football-players-peak-decline-quarterback-running-back-wide-receiver
• PFF aging RBs: https://www.pff.com/news/fantasy-football-narrative-street-when-do-aging-running-backs-hit-the-fantasy-wall
• Fantasy Life WR age: https://www.fantasylife.com/articles/redraft/how-does-age-impact-wr-performance-in-fantasy-football
• Footballguys elite TE decline: https://www.footballguys.com/article/2025-when-to-expect-an-elite-tight-end-to-decline-fantasy-football ; elite WR: https://www.footballguys.com/article/2025-when-to-expect-an-elite-wide-receiver-to-decline-fantasy-football
• Dynasty Edge age curves (EPA): https://thedynastyedge.com/2025/06/21/nfl-age-curve-study-epa-trends-by-position-rb-wr-te-qb-fantasy-football-analysis-2014-2024/
• Fantasy Life 350+ touch RBs: https://www.fantasylife.com/articles/fantasy/the-history-of-fantasy-football-rbs-after-a-massive-touch-season
• PFF 400+ touch RBs: https://www.pff.com/news/fantasy-football-what-happens-the-season-after-a-running-back-has-400-touches-hartitz-2022
• Harstad RB workloads: https://www.footballguys.com/article/HarstadRBWorkloads
• FanDuel Research TD regression: https://www.fanduel.com/research/touchdown-regression-what-it-is-and-how-to-use-it-for-player-prop-bets-fantasy-football ; Fantasy Points xTD (paywalled): https://www.fantasypoints.com/nfl/articles/2023/xtd-touchdown-regression-candidates
• PFF free agents switching teams: https://www.pff.com/news/fantasy-football-metrics-that-matter-free-agents-new-teams
• PFF catchable target rate / QB effect: https://www.pff.com/news/fantasy-football-expected-fantasy-points-wide-receivers-catchable-target-rate-2026
• CBS new play-callers 2026: https://www.cbssports.com/fantasy/football/news/new-play-caller-nfl-fantasy-football-rankings/ ; 4for4 play-calling tendencies: https://www.4for4.com/2026/preseason/how-play-calling-tendencies-affect-fantasy-football
• Draft Sharks injury predictor: https://www.draftsharks.com/injury-predictor/about

2026 data

• FOX Sports DK win totals (2026-09-09): https://www.foxsports.com/stories/nfl/2026-nfl-win-totals-over-unders-all-32-squads
• Yahoo win totals at schedule release (2026-05-16): https://sports.yahoo.com/nfl/betting/article/2026-nfl-schedule-release-odds-overunders-for-all-32-teams-ravens-rams-have-nfls-highest-season-win-total-013716986.html
• PFF projected win totals (2026-07-06): https://www.pff.com/news/bet-nfl-betting-2026-pff-projected-win-totals-for-all-32-teams
• Sharp Football implied team totals tool: https://www.sharpfootballanalysis.com/fantasy/nfl-implied-team-totals-tool/