Skip to main content
Back to AI Academy
FREENo

Demand Forecasting in Practice

A ~35-minute course for CSCOs, S&OP heads, demand planners — landscape, accuracy and the MAPE trap, model choice, demand sensing, promo uplifts, NPI, S&OP integration, 2-quarter rollout.

8

Chapters

~35 min

Duration

Intermediate

Level

No

Certification

Who this is for

For CSCOs, S&OP heads, demand planners, operations directors past basic AI literacy.

How this course works

  • 8 audio-narrated slide chapters · ~35 min of focused content
  • Trust trip-wires on every play — what not to cross
  • Free verifiable certificate on completion

What you'll walk out with

Specific outcomes from this course — no fluff.

  • The honest landscape — M5 Forecasting Competition reality, where AI beats classical and where it doesn't
  • Accuracy and the MAPE trap — why FVA (Forecast Value Add) is the metric that matters
  • Model choice — when classical wins, when ML wins, when LLM-augmented wins
  • Demand sensing in practice — Coca-Cola / Unilever / Carrefour cases, 20-40% promo-week lift
  • Promotional + event uplifts — clean promo calendar + LightGBM uplift model + cannibalisation
  • New product introduction — P&G analog-product approach, ML does the math, judgement picks the analogs
  • S&OP integration — narrative + accountability, FVA analysis, the quote-or-cut rule against LLM hallucination
  • A 2-quarter rollout with 4 trust trip-wires — no unexplainable model, no unrecorded override, no unsourced AI narrative, no set-and-forget

Course content

8 chapters · ~35 min

01

The forecasting landscape

Gartner SCT25 2025: 78% have AI-in-forecasting initiative, 41% measurable lift. McKinsey: 15–30% MAPE reduction (top quartile), 30–50% top decile. M5 LightGBM domination, classical competitive on aggregates. SAP IBP/Oracle/Blue Yonder/o9/Kinaxis. Tetlock: calibration > confidence.

02

Accuracy, bias, and the MAPE trap

APICS 2024-25: median bias +5 to +12% (over-forecasting, structural). MAPE asymmetry — biased toward under-forecasting, catastrophic on intermittent demand. M5 used RMSSE. Use sMAPE/MASE/RMSSE by context. APICS: 1pp MAPE = $0.5–1.5M per $100M.

03

Choosing the right model

The 3-variable decision tree: history × signal × granularity. LightGBM for long-history granular, ETS/ARIMA/Theta for aggregated, Croston (or Croston-LightGBM hybrid, Boylan 2024) for intermittent, analog-lookup for short history. Ensemble for robustness. ML weak past 6 months.

04

Demand sensing

Blue Yonder + o9. POS, weather, ad spend, social signals. 5–15% MAPE reduction on 4-week horizon. Confidence-weighted overrides + cap on magnitude. Signal hygiene (POS latency, mapping) + FVA shadow-mode before production. Noise-chasing failure mode.

05

Promotional & event uplifts

Coca-Cola / Unilever / Carrefour 2024-25 cases: 20–40% accuracy improvement on promotional weeks. Pattern: clean promo calendar + LightGBM uplift model + explicit cannibalisation modelling. Calendar drift (planned ≠ executed) is the failure mode.

06

New product introduction

P&G NPI case via McKinsey 2024. Analog-product lookup + first-4-week early signal. Category expert picks analogs (judgement layer), ML does the math. Lock forecast for first 4 weeks, then re-tune. 3 buckets: analog-rich, analog-thin, genuinely novel.

07

Connecting forecast to S&OP

SCC 2025: only 38% connect ML forecast to executive S&OP decisions. The gap is narrative + accountability. LokadTech 2025 LLM hallucination case (phantom promotion). FVA analysis (Gilliland/SAS) measures where in the chain accuracy is added or lost. Quote-or-cut for S&OP commentary.

08

Making it stick: your demand-forecasting roadmap

2 use cases · 2 quarters · 1 FVA discipline. Q1 = baseline ML on top 100 SKUs + FVA. Q2 = demand sensing OR promo uplift. 4 trust trip-wires: no unexplainable model, no unrecorded override, no AI narrative without source check (LokadTech rule), no "set and forget". Interactive roadmap builder.

Want this delivered inside your organisation?

The course is the starting point. The same content powers a 4-week pilot, an org-wide rollout, or a continuous build engagement — set up on your data, with your team, by Gennoor Tech.