Why Oncology Supply Chains Require Specialised Demand Forecasting

Demand is clinically structured but operationally uncertain

Oncology medicine demand is not simply a smoother version of general pharmaceutical demand. It can be influenced by treatment pathways, patient cohorts, line of therapy, dosing schedules, body weight or surface area, diagnostic decisions, formulary changes, capacity at treatment centres, clinical-trial activity and product substitution. Some signals are visible in advance; others change quickly at an individual site.

This combination creates a forecasting challenge. A national trend may suggest stable demand while a particular hospital experiences a short-term change in referrals or protocol. A low-volume medicine may be clinically critical even when ordinary statistical measures treat it as noise. A high-value cold-chain product may carry more financial and wastage risk than its unit volume suggests. The forecast must therefore support several decisions, not only the expected quantity.

Separate demand, allocation and availability

A useful planning model distinguishes what customers are likely to require, what stock the organisation intends to position and what authorised, released inventory can actually be supplied. Those are different measures. Demand may exist without an approved source. Inventory may be physically present but quarantined. An order may be submitted but outside the customer's contract or credit scope.

When these concepts are collapsed into one number, users lose trust. The system should label forecast demand, confirmed orders, allocated stock, released availability, safety stock and expected inbound supply separately. Each value needs a timestamp, unit of measure and source. Exceptions should be visible: delayed inbound stock, short expiry, temperature events, quality holds or unconfirmed supplier quantities.

• Forecast demand by product, site and relevant time horizon.
• Keep clinical demand signals separate from commercial orders.
• Expose release status and batch constraints before availability.
• Model expiry and cold-chain risk alongside quantity.
• Record human overrides and the reason for each decision.

Data quality determines model quality

Forecasting projects often begin with a choice of algorithm. In regulated supply, the more important first question is whether the inputs are defined, lawful, timely and representative. Order history can reflect past shortages rather than true demand. A zero may mean no need, no stock, a substitution or an unrecorded manual order. Product identifiers may change across supplier, customer and warehouse systems. Returns and cancellations can distort apparent consumption.

The data layer should reconcile product master records, units, pack sizes, customer locations, order states, batches, lead times and source-system timestamps. It should identify missing or unusual data before training or scoring. Where external information is considered, the organisation must document its permitted use, update frequency and relationship to the decision. Patient-identifiable data should not be pulled into an inventory model merely because it appears predictive.

Choose metrics that reflect the operating decision

A model can have an attractive average error and still perform poorly where the business most needs it. Oncology portfolios may contain intermittent or low-volume demand, so aggregate accuracy can be dominated by common items. Evaluation should examine product and customer segments, service-critical lines, bias, missed demand, excess stock, expiry exposure and performance over time.

The cost of an under-forecast and an over-forecast may differ materially. For a clinically important medicine, a missed requirement can create urgent sourcing and continuity pressure. For an expensive short-dated product, excess stock may create avoidable wastage. The planning objective should express those trade-offs rather than assume every unit error has equal consequence.

Human oversight is part of the model

Forecasts should inform accountable pharmaceutical and supply-chain decisions. They should not silently trigger purchasing or customer commitments without controls. A planner needs to understand the main input signals, uncertainty range, recent model performance and known exceptions. If the recommendation is overridden, the system should capture who changed it, why and what happened next.

This record improves both governance and learning. Repeated overrides may reveal a missing data source, a new treatment pattern or a process that sits outside the platform. A sudden model change may indicate data drift rather than a real market shift. Review thresholds can direct attention to high-risk decisions while leaving routine lines efficient.

A responsible maturity path

A credible forecasting roadmap begins with descriptive visibility: clean order, inventory, batch, expiry and lead-time data. The next stage may introduce statistical baselines and scenario planning, evaluated alongside current human processes. More advanced machine-learning methods become useful only when they deliver consistent improvement for defined decisions and can be monitored in operation.

Deployment should include model versioning, access control, test evidence, performance monitoring, data-drift checks, fallback procedures and business continuity. Users should see when data is stale or an integration is unavailable. A forecast should never be displayed as current stock, a guaranteed delivery date or a clinical recommendation.

NovaPharm's business plan identifies AI-assisted demand forecasting as a future capability. The right ambition is not to advertise an algorithm before it exists. It is to create the data, process and quality foundation from which a useful model can be developed. In oncology supply, responsible forecasting is less about technical spectacle and more about disciplined uncertainty management.

Early pilots should be deliberately narrow. A small product group with clear history and engaged operational users can expose data and workflow problems before the system influences wider purchasing. The pilot should compare recommendations with a documented baseline, examine failure cases and confirm that users understand uncertainty. Only a stable, monitored result should progress to another category or customer group. That maturity path is slower than a demonstration dashboard and considerably more useful to the people accountable for medicine continuity. It also creates auditable evidence for deciding whether further investment is justified. Governance should define who can approve expansion, pause the model or return to the manual fallback when performance moves outside agreed limits.