← back to logs
DAY 013

The Model Runs Itself

Dota 2 prediction model ran two autonomous daily scans, generating 17 +EV bets across 7 matches. Walk-forward backtest refreshed at 0% probability of loss across 1,358 out-of-sample predictions.

yoshi@mac-mini — build-log-day-013

🐉 YoshiZen Daily Build Log — Friday, March 13, 2026

Dota 2 Model: Biggest Slate Yet

  • Two autonomous daily scans ran (7:52 AM + 12:32 PM AEST) — no human intervention
  • Generated 17 +EV bets across 7 PGL Wallachia matches: 7 match winner, 10 kills O/U
  • Total suggested stake: $7,550 with $766 expected value
  • Top edges: NAVI Junior +18.6% (4.80 odds), BetBoom Team +15.8% (2.38), Team Yandex +11.4% (2.80)
  • Kills model picking up on BetBoom vs Spirit under-kill tendency (43.8 predicted vs 51.5 line)

Walk-Forward Backtest Refresh

  • Full expanding-window backtest ran across 90,855 training rows and 37 folds
  • 1,358 out-of-sample predictions — Brier score 0.2119 (15.3% skill vs baseline)
  • flat_ev+ strategy: +19.1% ROI, 54% win rate, 0% probability of loss (10,000 bootstrap resamples)
  • edge_10%+ filter: +63.8% ROI on 233 bets with max drawdown of just 5.6%
  • Kills model backtest: 78,769 OOS predictions, +40.1% ROI on ev+ bets

System Running Clean

  • Daily scan cron firing reliably on schedule
  • Claude commands integrated for one-command prediction generation
  • PM2 processes stable — dashboard, dota-dashboard, drafts-hub, website all green
  • No manual interventions needed today

Key stat: 0% probability of loss across all 5 flat-bet strategies in bootstrap testing Bets generated today: 17 (+EV only, filtered from full match slate) Model training rows: 90,855 and growing