## Summary

The **9+3 forecast** (9 months of actuals plus 3 months forecast) was far the most accurate. It landed just $3K over on revenue (0.02%) with a MAPE of only 0.63%, versus the earlier versions which ranged from 1.4% to 3.2% variance.

**The bias pattern:** All versions ran consistently high on revenue. Budget overshot by 3.2%, driven by July missing by $132K (15%). The 3+9 and 6+6 versions improved as more actuals rolled in, but still forecast above actual in most months.

**Operating expenses** were tighter across all versions—the Budget itself landed nearly flat, and 6+6 and 9+3 each came within 0.08% of actual. The real misses were on revenue, not cost control.

Full results with tables and month-by-month detail are in `output/accuracy-note.md`.