Solution / Demand forecasting
Plan for demand. Account for uncertainty.
Demand forecasting predicts future demand at a grain and horizon that match a real planning decision, such as replenishment or production scheduling. Quorlytic builds an honest baseline, backtests against historical outcomes with stockouts and promotions accounted for, and exposes forecast uncertainty rather than a single confident number.
By QuorlyticUpdated
Connect forecasting to replenishment, inventory or production planning. Build a repeatable evaluation before committing a business process to a new model.
Problem, approach and outcome.
| Problem | Specify product/location grain, planning horizon, lead time and review cadence. A daily SKU forecast and a monthly category forecast support different decisions and should not be judged by the same aggregate score. |
|---|---|
| Approach | Observed sales may be constrained by inventory. Include stock availability, returns, promotions and assortment changes where reliable. Record missing or unavailable inputs rather than filling the history with invented certainty. |
| Typical outcome | Planning objective and data suitability review |
The approach
- 01
Match the forecast to the decision
Specify product/location grain, planning horizon, lead time and review cadence. A daily SKU forecast and a monthly category forecast support different decisions and should not be judged by the same aggregate score.
- 02
Separate sales from demand
Observed sales may be constrained by inventory. Include stock availability, returns, promotions and assortment changes where reliable. Record missing or unavailable inputs rather than filling the history with invented certainty.
- 03
Backtest against a baseline
Use rolling time-based evaluation and compare with simple seasonal or existing planning rules. Measure bias and error by relevant segments. Watch for leakage from information that would not have been available when the forecast was made.
- 04
Give planners a controlled override
Expose forecast intervals and known limitations. Record the reason for an override and measure its effect over time. Monitor both statistical behavior and the inventory or service decision the forecast supports.
- 05
Planning analytics beyond the demand number
Forecasting is one input to a planning decision that also involves lead time, service level, capacity, working capital and shelf life. Making those constraints explicit shows which forecast improvements actually change a decision, and which only change a chart. Inventory, capacity and service-level trade-offs are usually where the value sits.
- 06
Suppliers, network and capacity belong in the picture
In logistics and manufacturing the constraint is often not demand itself but the network that has to absorb it: lane capacity, supplier lead time, production changeover and storage limits. Planning analytics becomes useful when the same model carries demand, capacity and cost together, and when exceptions point to a lane, plant or supplier rather than a percentage.
Reference implementation
Retail demand forecasting.
An illustrative architecture for connecting demand signals to a planner decision.
Business problem
Demand planning is fragmented across sales, inventory, promotions and calendar signals.
Data sources
- POS / Orders
- Inventory
- Promotions
- Product catalogue
- Calendar
Platform
- Ingestion and validation
- Modelled datasets
- Feature generation
Analytics & ML
- Baseline
- Forecast and confidence interval
- Driver analysis
Human decision
Planner review with a controlled override path.
Operational action
Replenishment, allocation and inventory planning.
Technology examples
Snowflake · Databricks · dbt · Python · Power BI — illustrative, not mandatory.
Illustrative reference implementation — not a client case study. Architecture varies by environment, data quality and operating requirements.
What the work produces
- Planning objective and data suitability review
- Baseline and rolling backtest
- Forecast output with uncertainty and limits
- Planner review and monitoring workflow
Delivery stages
- 01Sales & availability
- 02Baseline & features
- 03Backtest & segment
- 04Plan & learn
Rebuild what the planner could know at the forecast origin.
Time-ordered training rows can still leak future information. A promotion calendar revised after a planning meeting, or weather observed after the target day, can make a historical forecast look more capable than the decision-time system. This illustrative replenishment fixture freezes the information available at each forecast origin before comparing an approach with the planner's existing baseline.
Scope and assumptions: The fixture concerns historical evaluation at an agreed product/location grain and replenishment horizon. It does not forecast a real product or promise an error reduction. Calendar plans, observed outcomes and later corrections need distinct versions; their availability depends on the source and the actual planning workflow.
| Dimension | Input or evaluation boundary | Evidence available at the historical decision |
|---|---|---|
| Forecast origin | Record the meeting cutoff, target horizon, product/location grain and data-arrival deadline. A sale occurring before the cutoff but arriving afterwards may be unavailable to that forecast even though its business date is earlier. | Keep event time, availability time and source revision where possible. Reconstruct the accepted input snapshot instead of filtering only on the transaction date in a current warehouse table. |
| Known future plans | A promotion approved before the origin can be a legitimate future input. Use the plan version that was visible then, rather than the final calendar that later absorbed cancellations or changed discounts. | Retain the plan revision and approval/availability timestamp. Report how the forecast handles a planned promotion that subsequently changes; do not substitute the observed promotion silently during evaluation. |
| Unknown future conditions | Realised weather, final stock availability and actual supplier delays occur after the decision. If an operational forecast would use an earlier weather prediction or lead-time estimate, the backtest needs that earlier estimate too. | Separate scenario assumptions from available predictors. When an old estimate was not retained, identify the affected evaluation as a limited reconstruction rather than presenting hindsight inputs as a deployable forecast. |
| Rolling comparison | Advance the origin and recreate each input snapshot. Fit transformations on training data only and respect the time needed for labels to mature. Compare candidate and baseline on the same origins and eligible items. | scikit-learn's TimeSeriesSplit provides ordered folds for equally spaced samples; it does not reconstruct feature availability. Keep fold boundaries, any gap and the horizon policy with the as-of ledger. |
| Missing or censored history | A stockout can hide demand; a new item can lack a seasonal history. Keep those cases visible, define which baseline is possible and avoid removing difficult items only from the candidate's evaluation. | Report coverage, bias and error by relevant segment and origin, with the exclusions shared by both approaches. State whether the evaluation target is observed sales or a separately justified estimate of unconstrained demand. |
At narrow widths, scroll the worksheet to read every column. With the table focused, use the arrow keys to scroll.
Worked decision: a candidate's apparent advantage relies on the final promotion calendar, while the baseline uses the plan known at the meeting. Rebuild both with the earlier plan version before comparing them. If that historical version was not retained, keep the uncertainty explicit and begin retaining it for future origins. This is an evaluation design decision, not evidence that one forecasting method has won.
- Treat historical data corrections as versions. Later corrected sales may be appropriate outcome truth while still being unavailable as an earlier model input; these two roles should not be conflated.
- Keep planner overrides with their decision-time reason and information set. Evaluate the statistical forecast and the final overridden plan separately so a good or bad intervention is not automatically attributed to the model.
When the planning requirement is broader than one forecasting workflow, explore advanced analytics services with the decision-time evidence already defined.
Official references for this example:
Questions
Questions worth asking.
What accuracy improvement should we expect?
No percentage is promised before evaluating your history and current baseline. The first deliverable establishes what can be measured fairly and whether a more complex model offers enough improvement to justify operating it.
Is demand forecasting the same as supply chain analytics?
No. Demand forecasting estimates what will be wanted; supply chain analytics covers how the network responds — inventory, capacity, lead time, service level and cost. Forecasting is one input, and it only earns its place when it changes a planning decision someone owns.
Your next data decision
Start with a clear problem.
Build what comes next.
Bring the business question.
We'll help shape the engineering brief.