Cloud-Modus: Backtests & Sweeps laufen nur lokal Single Run und Mega-Sweep brauchen lokale Parquet-Daten + Megasweep-State (11 GB, nicht in Supabase). Das Sweep-Archiv ansehen → Sweep-Explorer in der Karte.
Ergebnis einordnen? Jeder Sweep wird an der 7-stufigen Bewertungs-Leiter gemessen — PBO < 0.2 (ehrliche Selektion) und ein sauberes Monte-Carlo-Risikoprofil entscheiden, ob ein Kandidat live darf. Fachbegriffe sind hier überall hoverbar.
Evidenz-Leiter →

Mega-Sweep Configuration

Automated 3-phase combinatorial optimiser. Tests all valid condition combinations, optimises parameters, and validates winners with walk-forward analysis.

3-Phase Process:
Phase 1 — Coarse Sweep: All condition combos x intervals (default params). Find which structures work.
Phase 2 — Fine Sweep: Top-N winners with parameter grid search. Find optimal params per structure. With Multi-Window Training, each combo runs on every training window — only configs profitable in ≥ min wins advance.
Phase 3 — Robustness: Top-N tested across a fixed multi-regime suite (full years 2021–2025 + Y2026 YTD + bull / bear / crash / range buckets, ~11 windows). Training windows are excluded from Phase 3 OOS. Ranking: all-profitable first, then worst-case score, then average.

Scoring: return_pct x MAR x pf_bonus (hard gates on trades/return/PF). Annualized returns shown alongside raw returns.
Lade…