The Opening: Bad Forecasts Make Replenishment a Gamble
A home-goods e-tailer with 3,000 SKUs replenished on planners' gut feel plus a 12-week moving average. Every major promotion broke the system: hot SKUs stocked out and lost orders while long-tail SKUs sat on $20 million of trapped capital. The owner did the math: each point of stockout reduction was worth $4 million a year. Forecasting is one of the highest-ROI AI applications in the warehouse — yet 90% of projects die at "the model is accurate but operations can't use it."
The Core Framework: Four Points
1. Feature Engineering: History Is Only Half the Story
Feeding a model nothing but sales history is driving with one eye closed. Five feature families that work in practice:
- Sales and price series (including promo prices and discount depth)
- Promotion calendar (mega-sales, category days, clearance — labeled in advance)
- Seasonality and holidays (back-to-school, Christmas, Lunar New Year)
- External signals: temperature (strong for apparel and home goods), channel traffic
- SKU attributes: new/mature/clearance stage, ABC classification
Rule of thumb: start with at least two years of history. SKUs with under a year of data only get analogy-based forecasts — don't force complex models onto them.
2. Tiered Modeling: Precision for Head SKUs, Simplicity for the Tail
| SKU tier | Share | Modeling strategy | |---|---|---| | Head 20% | 80% of revenue | LightGBM / time-series models, per-SKU tuning | | Mid 30% | 15% of revenue | Category-level forecast, then disaggregated | | Tail 50% | 5% of revenue | Croston-style intermittent-demand models; just avoid stockouts |
One model for everything is the classic beginner mistake. Tail SKUs have sparse demand — complex models overfit, simple ones stay stable.
3. Close the Loop: The Forecast Is Only the Starting Point
A forecast doesn't equal a purchase order. The full loop has five steps: forecast → safety stock (tiered by service level: 98% for A-items, 90% for C-items) → replenishment proposal → human review of exceptions only (SKUs swinging over 50%) → actuals fed back for retraining.
The key design: humans review only the exceptions. Planners look at the 5% of SKUs the system flags red each day instead of slogging through 3,000 lines — that's what AI leverage really looks like: machines handle 95% of routine judgment, people handle the 5% that matters.
4. Metrics: Watch Bias Direction Before Absolute Error
Use WAPE (weighted absolute percentage error) for overall accuracy, but watch bias direction even more closely: systematic over-forecasting builds overstock, systematic under-forecasting builds stockouts. Three straight months of same-direction bias beyond 5% signals a structural problem in features or model — send it back for rework, don't just tune hyperparameters.
Field Case: WAPE from 38% to 24%
A home-goods e-tailer, 4,000 SKUs, six months after launching AI demand forecasting: WAPE fell from 38% to 24%, stockout rate from 6.2% to 3.1%, inventory turns up 0.8. The biggest single contributor wasn't model sophistication — it was the promotion-calendar feature. Planners used to forecast promos from memory; systematizing that halved promo-period error overnight.
Pitfalls to Avoid
- Don't force history-based models onto new SKUs. Seed the first forecast from analogous SKUs, then recalibrate fast once the first four weeks of sales land — update weekly.
- Model promotions separately. Promo-period demand lives in a different distribution from everyday demand; training them together pollutes the baseline.
- Never wire forecasts straight to auto-ordering. Keep a human "exception gate" — for black swans the model never saw (a viral social-media hit), people are the last line of defense.
The 90-Day Rollout Plan
Days 1–30: data governance. Assemble 2 years of sales history, promotion calendars, and price series; publish a data-quality report and backfill any field missing more than 10%. Deliverables: feature inventory plus data-quality report.
Days 31–60: modeling and backtesting. Build tiered models by SKU class and backtest out-of-sample on the most recent 3 months — head-SKU WAPE under 25% is the gate. Deliverable: backtest report.
Days 61–90: pilot and scale. Run 200 head SKUs in parallel: the system proposes, planners review, and pilot vs. control stockout rates are compared. A pilot stockout reduction of 1.5+ points green-lights full rollout.
90 days is the floor; warehouses with weak data foundations should plan 120. Better slow up front than rework after launch — one reworked forecasting project zeroes out the business's trust in AI.
The Planner's New Workflow: What Human-Machine Division Looks Like
After AI goes live, planners don't disappear — they level up:
- 30 minutes each morning: review only system-flagged exception SKUs (swings over 50% or stockout alerts), confirming or adjusting each. Out of 3,000 SKUs, typically just 100–150 need human eyes.
- 1 hour weekly: read the bias-direction report. Two straight weeks of systematic over/under bias means a joint session with IT on features — no need to tune models personally, but know which questions to ask.
- Half a day monthly: join the backtest review, track WAPE trends and promo post-mortems, and update promotion-calendar labels — the most irreplaceable human input of all: business intuition.
The role shifts from "person who does math" to "person who manages models." Planners who resist get left behind; those who embrace it manage triple the SKUs solo. Uncomfortable, but true.
The Takeaway
The field formula for AI demand forecasting: good features beat fancy models, tier your modeling, close the loop with humans on exceptions only, and watch bias direction. Ten points off WAPE converts directly into profit through fewer stockouts and faster turns.
