Skip to content

DRAFT_LOGIC.md

Latest commit

 

History

History
72 lines (57 loc) · 3.87 KB

DRAFT_LOGIC.md

File metadata and controls

72 lines (57 loc) · 3.87 KB
Raw

Draft Logic Redo + Backtesting (2021–2025)

Requirement: redo the draft logic; the autodraft hoards K/DEF and starves offense (QB/WR/RB/TE), and the bench is the weakest part. Phase v2.4. Depends on the v2.2 value upgrade.

What's wrong in v1

The autodraft (draftAI.ts) picks "best available by need via VORP." Two failures:

  1. K/DEF hoarding: with v1's over-generous K/DEF value (fixed in v2.2) and a naive need model, bots take kickers/defenses too early and too often, leaving startable offense on the board.
  2. Bench is an afterthought: late picks default to whatever has the highest raw value — often a second/third K/DEF or a low-upside flier — instead of upside offensive depth, bye/injury cover, and handcuffs that actually win leagues.

The redo — a need-and-upside draft policy

Best-pick-by-value, replaced with an objective that scores each candidate on:

pick_score = marginal_starter_value          # how much it raises my projected starting lineup
           + bench_value(candidate, roster)    # upside/cover value if it's a bench pick
           − positional_overfill_penalty       # diminishing returns past slot+reasonable depth
           − scarcity_adjusted_opportunity_cost # what I'd lose by not taking the run-position now

Key pieces:

  • Starting-lineup marginal value: value a pick by how much it improves the optimal starting lineup (incl. FLEX/OP superflex), not raw VORP. A 3rd RB adds little to a full lineup; a needed QB in superflex adds a lot.
  • Bench model (the focus): bench slots are valued explicitly by (a) upside (ceiling / boom, not mean), (b) handcuff/cover correlation to your starters (RB handcuffs, backup QB in superflex), and (c) bye-week and injury redundancy. A bench pick's worth = expected starts × value-when-started × availability — which makes a high-ceiling WR4 beat a 2nd kicker every time.
  • K/DEF cap: at most one K and one D/ST until the final rounds; their compressed v2.2 value plus an explicit late-round gate stops early hoarding.
  • Positional caps & scarcity runs: respect league max per position; model superflex QB runs (opponents are QB-hungry) so the bot doesn't get left behind on QBs.
  • Opponent model: bots draft need-aware with superflex-correct QB demand (feeds Monte Carlo too).

Backtesting harness (2021–2025)

A reproducible offline harness — the core deliverable of v2.4, shared with SCORING.md.

  1. Data: nflverse historical seasons 2021, 2022, 2023, 2024, 2025 (actual fantasy points under our league rules — superflex, half-PPR, distance-K, yardage-D/ST).
  2. Setup: reconstruct each season's draft pool + realistic ADP; simulate a 12-team superflex snake draft with our policy vs. baselines.
  3. Baselines to beat: v1 autodraft policy; raw-VORP; ADP-follow; "K/DEF last" heuristic.
  4. Metric: season points-for of the optimally-set lineup each week (start-sit done greedily from actuals), plus head-to-head record vs. the field, averaged over seeds.
  5. Ablations: turn off the predictability discount / bench-upside / K-DEF cap individually to confirm each helps.
  6. Output: a tuned parameter set (f(ρ) exponent, bench weights, scarcity curve) and a report (docs/modeling/backtest-report.md, written when v2.4 runs).

Note: backtests are executed in v2.4, not in this documentation phase. This doc specifies the harness, metrics, and acceptance so the work is turnkey.

Acceptance (v2.4)

  • Across 2021–2025, the v2 policy beats v1 and raw-VORP on mean season points-for and H2H record (report with seeds + CIs).
  • Autodraft no longer takes K/DEF before the last rounds; benches are dominated by upside offense + cover, not duplicate K/DEF.
  • Live + offline draft boards consume the same policy (thin consumers — D7), so manual and synced drafts get identical recommendations.