Skip to content

Latest commit

 

History

History
82 lines (65 loc) · 4.66 KB

File metadata and controls

82 lines (65 loc) · 4.66 KB

Scoring & Player Valuation Upgrade

Requirement: kickers and defenses are still overvalued. This doc specifies the fix and the research behind it. Implementation is phase v2.2. Companion: VALUE_ENGINE.md.

Research grounding (why K/DEF are overvalued)

External consensus and our own league context agree:

  • K and D/ST have the highest week-to-week and year-to-year variance and the lowest predictability of any position. The top-10 at both positions turns over almost completely year to year. (Fantasy Points, FantasyPros VORP)
  • Because they're freely streamable off waivers, their true value over replacement collapses — the replacement-level K/DEF is nearly as good in expectation as the "elite" one. Accepted practice: draft them last, stream all season.
  • Projection models overestimate busts and underestimate booms, and TD/turnover-dependent scoring (exactly what K/DEF are) is the most variable. A point projection treats a volatile D/ST like a stable WR — that's the bug. (ParlaySavant, FF Analytics)

Diagnosis: v1 ranks K/DEF on projected point total with a per-position replacement baseline that's too generous. It rewards a high mean without penalizing (a) low predictability and (b) the trivially-available replacement. So elite-projected K/DEF float up the board.

The fix — three coordinated changes

1. Predictability-discounted value

Each player gets a predictability score ρ ∈ [0,1] from the Projector: how reproducible is this projection? Derived from historical year-over-year rank stability for the position + this player's own variance + share of points from high-variance sources (TDs, turnovers, return/defensive TDs). Value is discounted toward replacement by low predictability:

adj_value = replacement + (raw_VORP) * f(ρ)        # f(ρ) ≈ ρ^k, k tuned by backtest

K and D/ST have structurally low ρ → their VORP is heavily compressed toward 0. Stable RB/WR/QB keep most of their VORP. This is position-agnostic — it just happens to hit K/DEF hardest because the data says it should.

2. Replacement level from real waiver availability

Set K/DEF replacement at streamer level, not "12th-best." Because every team streams, the realistically-available K/DEF each week is close to the positional median, not the worst starter. Concretely: replacement baseline = the expected points of a matchup-optimized streamer (top-of-waiver each week), which sits near the position's upper-middle — crushing the gap between "elite" and "replacement."

3. League-specific signal kept (but bounded)

Our league has distance-based kicker scoring (50+ = 5, 60+ = 6) and a yardage-allowed D/ST component (uncommon) — so some differentiation is real: kickers with big legs / on high-scoring offenses, and genuinely stout low-yardage defenses do carry more value. The model keeps a KickerProjector/DefenseProjector that captures this, but the predictability discount + streamer replacement bound how high it can lift them. Net effect: the best K/DEF are a modest edge, not a mid-round pick.

Position treatment summary

Pos Replacement basis Predictability ρ Net v2 effect
QB league slot demand (superflex ⇒ ~24th) high unchanged (correctly premium)
RB/WR slot demand incl. FLEX/OP med–high unchanged
TE slot demand medium unchanged
K weekly streamer (≈ upper-median) low strongly compressed
D/ST weekly streamer (≈ upper-median), bounded by yardage signal low strongly compressed

Validation

  • Backtest 2021–2025 (shared harness with DRAFT_LOGIC.md): does v2 value, used to draft, beat v1 value on season points-for, and stop spending early picks on K/DEF?
  • Calibration check: do realized boom/bust frequencies match the emitted distributions (reliability diagram)? Especially for K/DEF.
  • Sanity: no K or D/ST should rank above a startable offensive player in superflex draft value except in extreme, well-justified cases.

Acceptance (v2.2)

  • K/DEF draft value compressed to "last few rounds" range across 2021–2025 backtests.
  • Predictability score surfaced in the UI (the player-card "predictability meter") so the why is legible.
  • No regression to QB/RB/WR/TE valuations vs. v1 on backtest.