On - Forecasting Discussion (Giannis/Federico) - 2026-01-30¶
Metadata¶
- Date: 2026-01-30
- Company: On
- Attendees: Federico Morando, Art Koci, Simon Jonsson, Giannis
- Type: Customer Call
- Source: Notion Customer Success Meetings
Summary¶
Forecasting Model Architecture Overview¶
- Current V2: global setting -> entities -> accounts -> categories with transactions
- Forecasting at transaction-level categories, then ensembling for better category predictions
- Bottom-up reconciliation using neural, statistical, ML and deep learning models
- Model selection based on 13-month historical performance evaluation
- V3 planned: multi-level scenario planning, human-interactive assumptions, more sophisticated modeling
Federico's Presentation Goals & Timeline¶
- Final month at On Running — building case for Palm's continued partnership
- Presentation to senior director (Christoph) end of January
- Content: technical deep-dive, 8-week Palm vs Kyriba accuracy comparison, accuracy across forecast horizons
- Check-in scheduled February 11th
Model Performance & Improvement Process¶
- 181% improvement in VMAPE across global level
- Clustering approach: categories grouped by characteristics (volatility, density, sparsity)
- Different treatments per cluster type
- Upcoming dashboard features: VMAPE, weighted MAPE, scaled MAPE, signal-to-noise ratio, mean absolute error, exportable data
Integration Priorities¶
- BigQuery integration critical for AP/AR data to improve model accuracy
- Cash pool forecasting identified as major gap — needs better categorization, country-specific considerations
- APAC and Brazil particularly challenging
- IC transaction forecasting in development pipeline
Synced from Notion: https://www.notion.so/2f8c47335ffb81238669c93e48ad7f3b