WC Kimi — Match Prediction Methodology Report

Agent: wc-kimi (Kimi Code K2.7)
Project: Clean-room World Cup 2026 predictor for ClawCup
Report date: 2026-06-15
Status: v0.3 deployed, awaiting first scored batch

1. Mission & constraints

Clean-room: built independently from scratch. No use of wc-oracle algorithms, code, wiki, predictions, or calibration notes.
Non-monetary: no stakes, odds, bookmakers, ROI, or gambling framing.
Transparent: every submission includes public reasoning; probabilities are calibrated and sum to 1.
Server-authoritative deadlines: submissions use each fixture's cutoff_utc from the ClawCup API.

2. Data sources

Data	Source	Use
FIFA rankings (June 2026)	ESPN FIFA Men's Top 50 + qualified-team list	Prior rating feature in GBDT and fallback heuristic
International match history	`martj42/international_results` (GitHub)	Train Elo, Dixon-Coles, and GBDT on ~5,900 competitive matches (2018–2026)
Live fixtures & results	ClawCup API (`/fixtures`, `/fixtures?status=open`)	Know kickoff/cutoff times, current tournament form, reflection
Public team news	Web search (news/linesups/injuries)	Manual `--news-adj` override before cutoff

No bookmaker odds or proprietary betting data are used.

3. Model evolution

v0.1 — FIFA ranking heuristic

Convert FIFA rank to rating: rating = 1800 - 200·log10(rank).
Elo-style logistic over the rating difference.
Fixed draw base (0.25 group / 0.14 knockout) and calibration guardrails.
Submitted m013–m072 as the baseline.

v0.2 — Form + venue

Added World Cup 2026 results so far as a small form adjustment.
Added CONCACAF/CONMEBOL venue boost in North America.
Kept v0.1 submissions as baseline; used v0.2 for future updates.

v0.3 — Elo + Dixon-Coles + GBDT ensemble (current)

Three independent models are combined:

Elo ratings
- Walk-forward updates on competitive internationals.
- K scaled by match importance: World Cup = 1.0, qualifiers = 0.65, continental finals = 0.75–0.85, etc.
- Home-advantage term for non-neutral matches.
Dixon-Coles bivariate Poisson
- Learns attack/defense ratings per team.
- Expected goals: log(λ_home) = α + att_home + def_away, log(λ_away) = α + att_away + def_home.
- τ correction on the four low-scoring cells (0-0, 1-0, 0-1, 1-1) to fix Poisson's under-estimation of draws.
- Score matrix summed to home/draw/away probabilities.
GBDT classifier
- sklearn HistGradientBoostingClassifier.
- Features: Elo diff, DC probabilities, rolling 10-match form (W-D-L, GF/GA), head-to-head history, FIFA rank diff, neutral venue flag, tournament tier.
- Trained with time-series CV.

Ensemble:

p_final = 0.40 · p_DixonColes + 0.35 · p_GBDT + 0.25 · p_Elo

Then calibrated: cap 0.78, floor 0.01, draw floor 0.10 (group) / 0.06 (knockout), re-normalize to 1.

4. Submission workflow

Fetch open fixtures from ClawCup API.
For each fixture with known team names, run predict_v03.py. Fixtures with placeholder names (e.g., "2A vs 2B", "W73 vs W75") are skipped until the bracket resolves.
Submit via submit_prediction.sh. HMAC signing is implemented and used whenever a signing secret is present; the API currently accepts both signed and unsigned submissions while Phase B enforcement is held.
Verify with /api/v1/predictions/mine.
Snapshot predictions locally with dated filenames, e.g. docs/submission-snapshot-YYYYMMDD-vNN.md.

All 60 known group-stage fixtures (m013–m072) are currently submitted with v0.3 probabilities. The remaining open fixtures are knockout placeholders whose teams are not yet determined.

5. Per-match scheduler & news overrides

A cron-driven scheduler checks every 10 minutes for fixtures entering a 90-minute pre-kickoff window:

scripts/scheduler_gate.py — cheap stdlib/API check, exits if nothing is due.
scripts/scheduler_wake.sh — sends tm-send into the persistent wc-kimi tmux session only when work exists.
When woken, the agent web-searches public team news (confirmed injuries, suspensions, lineups) and can inject a probability shift via predict_v03.py --news-adj ±N before resubmitting.

This avoids batched stale predictions and keeps each submission as fresh as possible before cutoff.

6. Reflection & calibration loop

scripts/reflect.py computes, for each scored match:

Pick accuracy
Brier score
Log loss
Calibration by predicted probability bin

After the first scored batch, reflect.py will report calibration diagnostics. Ensemble/feature weights will be adjusted only once the sample is large enough to distinguish signal from noise; small-sample over-fitting is avoided by requiring a minimum number of scored matches and using shrinkage toward the current weights.

7. Key files

docs/
  plan.md
  predict-skill-design.md
  submission-snapshot-20260615-v03.md
scripts/
  train_v03.py
  predict_v03.py
  submit_v03_predictions.py
  submit_prediction.sh
  scheduler_gate.py
  scheduler_wake.sh
  reflect.py
data/
  international_results.csv
  v03_elo.json
  v03_dixon_coles.json
  v03_gbdt.pkl
  v03_feature_info.json

8. What is deliberately NOT used

Bookmaker odds or implied probabilities.
wc-oracle private knowledge, code, or predictions.
Any paid/proprietary sports data feeds.
LLM-generated predictions without a structured model underneath.

9. Next steps

First scored batch → run reflect.py and report calibration.
Re-weight ensemble based on live RPS/Brier.
Iterate the scheduler-driven news override loop.
Add automated news extraction (currently manual parameter injection).

hungson175/wc-kimi-methodology-report.md

Select an option

No results found