×

OverADP — How to Use the War Room

OverADP is an AI-powered fantasy football draft assistant. It uses machine learning trained on historical NFL data plus current roster overlays, teammate target competition, draft-value rookie signals, prior-year contract signals, and conformalized quantile regression for calibrated confidence intervals. Target Intel uses opportunity-cost draft math that adapts to roster settings, live picks, VBD, ADP, flex paths, and player availability before your next turn. Updated June 2026 with current Sleeper roster/depth data and a clean 825-player draft board.

Understanding the UI

Player Pool (Left Panel)

Each player row has two action buttons:

• DRAFT — Click to add this player to YOUR team.
• TAKEN — Click to mark this player as drafted by an OPPONENT (removes them from the board).

Column	Meaning
VBD	Value Based Drafting — projected points above the replacement level at each position. Higher = more valuable relative to a waiver-wire fill-in.
PROJ	Projected fantasy points for the full season (half-PPR by default; scoring format is configurable).
CI	Conformalized 80% Confidence Interval [low–high]. Validated on held-out seasons: when we say the 80% CI is 110–180 pts, the truth actually lands in that range ~80% of the time. Rookies have wider CIs (1.8× widening); second-year players 1.3×.
UNC	Uncertainty: low / medium / high — computed as a position-relative quartile on relative CI width (range width ÷ projection). Low = a tight projection range. High = wide CI or near-replacement-level depth.
DRAFT RISK	low / medium / high — an explainable draft-value score using CI uncertainty, recent injury history, rookie/second-year unknowns, ADP overpay, and low VBD at cost. It is a decision aid, not a medical projection.
BYE	Bye week for this player's team. Watch for conflicts with your other starters.
ADP	Average Draft Position from Fantasy Football Calculator — where the market values this player. Sleepers = model rank better than ADP; busts = worse.

Draft Board Sort

Offensive players (QB/RB/WR/TE) are sorted by an opportunity-cost draft score. VBD and projections measure value, ADP measures price, and the board estimates whether a player is likely to make it back to your next pick. Kickers (K) and Defenses (DEF) are sorted by projected points since they have no meaningful VBD signal.

My Team (Center Panel)

Roster slots are driven by your league configuration (Settings). Players are auto-assigned in priority order:

• QB → QB slot(s), then Superflex (if configured)
• RB → RB1, RB2 (or however many RB slots), then Flex, then Bench
• WR → WR slots, then Flex, then Bench
• TE → TE slot(s), then Flex, then Bench
• K / DEF → their dedicated slots (configurable count)

Bye weeks are shown next to each player. Red = bye conflict with another starter at the same position.

Right Sidebar Panels

Panel	What It Shows
Target Intel	Top 5 recommendations from opportunity-cost draft math, blended with VBD, live roster holes, flex needs, positional scarcity, ADP edge, sleeper/bust flags, draft risk, and late-round K/DEF timing.
Positional Scarcity	Drop-off from elite to replacement level per position. Higher dropoff = position is scarce = draft early.
Bye Conflicts	Warns when multiple starters share a bye week. Critical = no bench coverage. Warning = bench available.
Sleepers & Busts	Compares model rank vs ADP. Sleepers = model ranks higher than market. Busts = market overvalues.
Handcuff Alerts	Backup RBs on the same team as your starting RBs. If your starter goes down, the handcuff takes over.
Opponent Drafts	Tracks what positions opponents are taking — watch for position runs.

Key Concepts

VBD (Value Based Drafting)

VBD = Projected Points − Replacement Level Points. The replacement level is the projected points of the last player typically drafted at each position (including bench). This means even bench-worthy players get positive VBD, helping you compare value across all rounds — not just the early ones.

Target Intel

Target Intel is separate from the point-projection model. It asks a draft-specific question: given your roster, league settings, current pick, next pick, ADP price, and the players still available, who is worth taking now?

The recommendation layer is ADP-guarded: ADP remains the price anchor, but the app boosts players with enough VBD, roster need, and scarcity to justify taking them now, especially when the next-turn estimate says they probably will not make it back.

Roster settings matter. No-flex, multi-flex, 2QB, unusual starter counts, and very shallow or deep benches all change the opportunity cost and next-pick math directly.

Positional Scarcity

Measures the drop-off from the elite tier to replacement level at each position. If QB has a small dropoff but TE has a huge one, you should prioritize TE — because the quality falls off a cliff after the top TEs are gone.

How it's calculated: For each position, OverADP finds the projected points of the #1 player (elite) and the projected points of the last starter that would be rostered across all teams (replacement). The difference is the scarcity score.

Scarcity Level	What It Means	Draft Strategy
High Scarcity	Huge gap between elite and replacement	Draft this position early — waiting means getting a much worse player
Medium Scarcity	Moderate dropoff	Balanced — draft when value aligns with your pick
Low Scarcity	Shallow dropoff — many comparable options	Safe to wait — replacement-level players are almost as good as "starters"

Example: In a 12-team league starting 1 TE, the #1 TE (Kittle, 136 pts) and the #12 TE (~97 pts) differ by 39 pts. But the #1 QB (Lamar, 245 pts) and #12 QB (~170 pts) differ by 75 pts. However, QB scarcity is often lower because there are only 12-14 starting QBs rostered, while TE has a steeper talent cliff at the top. The scarcity panel shows you which positions have the steepest cliffs so you can draft accordingly.

Confidence Intervals (CQR)

The 80% CI is built via Conformalized Quantile Regression — we train separate quantile models at the 10th and 90th percentiles, then apply a conformal adjustment learned on a held-out calibration season to guarantee marginal coverage.

Empirical coverage targets 80% marginal coverage. The latest calibration lands around 81-82% by position after the conformal adjustment; raw quantile intervals were meaningfully undercovered before that correction.

Rookies and 2nd-year players get a 1.8× / 1.3× width multiplier on top of the calibrated interval to reflect their scarcer historical data.

Model vs ADP

Our model explains 63% of variance in fantasy points on held-out seasons. ADP alone (a learned curve from draft position to expected points) explains only 6% — roughly 10× worse. Walk-forward validation: train 2021 to N-1, test season N, averaged across 2022-2025.

Position	Model MAE	ADP MAE	MAE Edge	Model R²	ADP R²
QB	65.0	107.8	-39.7%	0.57	0.00
RB	38.0	65.0	-41.5%	0.66	0.12
WR	30.6	51.2	-40.3%	0.66	0.10
TE	22.3	39.6	-43.8%	0.64	0.03

When the model and ADP disagree strongly, it flags sleepers (model rank better than ADP) and busts (ADP rank better than model).

Handcuff Alerts

A handcuff is a backup RB on the same NFL team as your starting RB. If your starter gets injured, their backup typically takes over the starting role and inherits most of the carries — making them extremely valuable as insurance.

Why it matters: If you drafted Saquon Barkley in the 1st round and he goes down in Week 3, his backup (who you can draft in the last round for almost nothing) suddenly becomes a top-15 RB. Without the handcuff, you're scrambling on waivers.

Handcuff Type	Description	Priority
Must Handcuff	Your RB1's backup — protect your 1st/2nd round investment	High — draft in late rounds
Consider	Your RB2/RB3's backup — nice insurance but less critical	Medium — if bench space allows
Available	Other backup RBs still on the board	Low — only if you have empty bench slots

Strategy: Draft your stud RB's handcuff with one of your last picks. It's like buying insurance — you hope you never need it, but if you do, it saves your season. The Handcuff Alerts panel shows you which backups are still available for RBs already on your roster.

Model Architecture

OverADP currently runs tuned CatBoost models per position, selected after walk-forward experiments against simpler baselines. The production models use per-position feature sets, temporal sample weighting (more recent seasons matter more), and leakage-controlled lag features so current-season outcomes never feed their own predictions.

Calibrated confidence intervals come from a parallel Conformalized Quantile Regression pipeline (see Confidence Intervals section above).

Model	Role
CatBoost (production)	Ordered boosting for the QB/RB/WR/TE projection models.
Target Intel	Opportunity-cost next-pick scoring for live draft recommendations.
ADP Baseline	Held out as the market comparison, not used as a target leak.
Conformal Quantiles	Parallel interval pipeline for the 80% confidence ranges.
K/DEF	Consensus/regression-to-mean board support, not overfit ML projections.

Key Features (June 2026)

• Current roster overlay — Sleeper roster/depth data keeps teams, bye weeks, and post-draft movement aligned with the board.
• Target / carry competition — Sum of teammate prior-year production on current roster. Captures FA moves and scheme shifts directly.
• College & draft pipeline — Draft capital, draft-pick value, combine athleticism, college per-game production, Power 5 flags, and rookie interaction features.
• Injury-adjusted regression — Separates injury-driven declines from talent-driven declines; flags bounce-back candidates.
• Contract signals — Prior-year contract context helps capture role stability, especially in WR/TE experiments.
• Conformal 80% CIs — Latest calibration lands around the 80% target across positions.
• ADP matching — Robust via football_name (handles apostrophes, suffixes, nicknames).

Data Sources

• nflverse (via nfl_data_py) — Seasonal stats, rosters, team stats, OL metrics, snap counts, injuries, depth charts
• Fantasy Football Calculator — ADP (half-PPR) and bye weeks
• FantasyPros — K and DEF projections (ML is not reliable for these; we use consensus instead)
• Pro-Football-Reference — Combine data, draft capital, college production stats

Accuracy (Walk-Forward Validated)

Trained on 2021 through N-1, tested on season N, averaged across 2022-2025. Features are strictly lagged to prevent leakage. Calibration set for CQR is the most recent season in the training window. Numbers refreshed June 2026 after the roster/depth, draft-value, and contract-signal updates.

Position	Model MAE	ADP MAE	MAE Edge	Model R²	ADP R²
QB	65.0	107.8	-39.7%	0.57	0.00
RB	38.0	65.0	-41.5%	0.66	0.12
WR	30.6	51.2	-40.3%	0.66	0.10
TE	22.3	39.6	-43.8%	0.64	0.03
Overall	39.0	65.9	-41% avg	0.63	0.06

The model explains roughly 10× more variance than ADP alone. The strongest MAE edge is TE (-43.8%), with RB (-41.5%), WR (-40.3%), and QB (-39.7%) all clearing a roughly 40% error reduction versus ADP.