IBP: Demand Forecasting
ML model accuracy 88→95%, planning cycle 12→4 days, ₽263–550M/year impact
What doesn't work
Freight transportation planning was manual: commercial directors manually determined volumes by clients and routes. Production plan accuracy by clients — 88%, by routes — 59%. Planning cycle — 12 days. Labor costs — ~180 person-days per year.
Architectural approach
Automated demand forecasting across the network (unconstrained demand) with ML models. Integration into S&OP process: scenario planning, freight base optimization. 3 MVP phases: forecast for PV → expansion to KR → all RPS regions. Corporate architect on the project team.
What made it hard
Commercial directors had planned manually for years and didn't trust ML forecasts — deployment required running both approaches in parallel for 6 months until ML proved superiority. Route accuracy was 59% — the model performed worse than humans on this dimension until seasonal features were added. Board justification: the accuracy-to-revenue correlation (0.798) was the only argument that convinced.
My role & contribution
CTO / Technical Director
Corporate architect on the project. Designed ML forecast integration with Optimizer and Navigator. Defined the 3-phase MVP strategy. Justified the business case to the Board (₽263-550M/year impact, accuracy-to-revenue correlation = 0.798).
How it looks
System architecture
How it works
ML models for client and route forecasting. Automated route volume calculation replacing manual process. Integration with Optimizer and Navigator. Daily freight forecast for scheduling. Prototype → MVP 1.0 (PV) → MVP 2.0 (PV+KR) → MVP 3.0 (all RPS).
Why this way
ML forecast + S&OP instead of pure BI automation
Automate existing manual process without ML (dashboards + alerts)
BI automation: speeds up the same manual process but doesn't improve accuracy. ML model: new forecast quality that humans can't achieve manually. Accuracy-to-revenue correlation = 0.798.
Accuracy +7 pp, cycle 3× faster, ₽263–550M/year impact. Foundation for company's S&OP process
Results
- 01
- Client accuracy: 88% → 95%
- 02
- Route accuracy: 59% → 75%
- 03
- Planning cycle: 12 days → 4 days
- 04
- Freed 779 person-days/year (~3 FTE)
- 05
- ₽263–550M/year total impact
Impact on business
Three impact areas: S&OP scenarios (₽150–300M/year from correct obligation determination), marginal revenue increase (accuracy-to-revenue correlation = 0.798), 5–10% reduction in over-provisioning (₽113–250M/year). Total: ₽263–550M/year. Freed 779 person-days for commercial directors.
Algorithms & patterns
Technologies
- Python
- ML Pipeline
- Anaplan
- SAP BW
- S&OP Framework