Forecast module
Versioned, explainable, self-tuning demand forecasting at SKU × customer × month grain.
Driver decomposition tells you why the number is what it is. What-If lets you move a promo and see the new curve in seconds. Accuracy backtests run continuously, so a drifting model gets flagged before it costs you a quarter — not after.
Sound familiar?
We built this module by sitting with demand planners and listening to what they complain about on Monday morning. Each pain below maps to a specific page in the dashboard — no buzzwords, no roadmap promises.
"The forecast lives in one analyst's spreadsheet and breaks the day they're on vacation."
A versioned, auto-rerun pipeline owns the forecast end-to-end. Every model and every input is reproducible — no single point of failure, no copy-paste between tabs.
→ Pipeline
"When the number's off, nobody can tell us why."
Driver decomposition breaks each forecast into its components — seasonality, trend, promotional lift, holiday effects, ad-hoc events. You see, at SKU × customer grain, exactly which signal moved.
→ Intelligence
"Every promo change kicks off a full day of refitting."
What-If lets you tweak assumptions — shift a promo window, change pricing, add a new SKU — and see the forecast move in seconds. Compare scenarios side-by-side before you commit to AOP.
→ What-If
"We won't know if the forecast was right until the month closes."
Accuracy backtests run continuously by brand, segment, and period. MAPE / RMSE / bias trends update with every retrain so you can flag a drifting model before it costs you a quarter.
→ Accuracy
"Every quarter we re-pick the model and retune the seasonality blend."
A self-tuning loop monitors every series. Stable series stay on their current model; drifting series get relearned automatically. You see the retrain queue, not a black box.
→ Learning
"Promoting a new model to production is a leap of faith."
Model Performance vets a candidate model against the incumbent on MAPE / RMSE / bias before promotion. Side-by-side, version-by-version — no production surprises.
→ Model Performance
What's inside
Every page reads from the same underlying engine, so the number you see on the Overview is the same one driving What-If, the same one being backtested on Accuracy, and the same one whose drivers are decomposed on Intelligence.
Headline KPIs across the whole module — actual vs forecast vs AOP, confidence distribution, monthly detail. The single screen the planning team opens every Monday.
Stage-by-stage breakdown of the generation pipeline — read raw sources, blend in causal drivers, validate against guardrails, publish to downstream consumers. Click any stage to see what ran, when, and what it produced.
Backtest accuracy by brand × segment × period. MAPE / RMSE / bias trends with month-over-month deltas. Diagnose where the model is reliable and exactly where it drifts.
Driver decomposition — seasonality, trend, promotional lift, holiday effects, ad-hoc events. Each forecast point comes with its 'why' attached, at SKU × customer × month grain.
Scenario planning. Tweak a promo, shift a launch date, model a new pricing tier — see the forecast move in seconds. Pin scenarios side-by-side before you commit to the AOP.
Self-tuning state — which series are stable, which are being relearned, and the most recent retrain status. The system surfaces its own work; you don't have to babysit it.
MAPE / RMSE / bias trends across model versions. Used to vet a candidate model against the incumbent before promotion — every release goes through this gate.
What you'll surface
Each view is opinionated about the next question. You don't go looking for the answer — the page brings it forward.
How the pipeline runs
01
Pull from the canonical gold layer — sales, customer master, product master, holidays, promos. No copy-pasting between files; the same data the rest of the dashboard sees.
02
Hierarchical middle-out: brand × segment × SKU forecasts reconciled top-down and bottom-up. Causal drivers (promo, holiday, price) attached as features.
03
Guardrails check the candidate forecast against bounds, the prior period, and the AOP. A drift flag pauses the publish step until a planner reviews it.
04
Versioned outputs hit the dashboard and any downstream consumer. Every prior version is queryable — no overwrites, no surprises.
What we need from you
Required sources give you the volume curve. Holiday and promo calendars turn on causal driver decomposition; the AOP enables plan-vs-forecast variance overlays.
| Source | Description | Required |
|---|---|---|
| sales | Sales transactions | Required |
| product_master | Product master | Required |
| customer_master | Customer master | Required |
| holidays | Holiday + event calendar | Optional |
| promo_calendar | Promotional calendar | Optional |
| aop | Annual Operating Plan | Optional |
Full schema details on the Data Schema page once you sign in.
Request a demo and we'll walk you through the modules with sample data shaped like yours.
Request a demo