Demand planning for pharmaceutical and biotechnology organizations, built around the reconciliation between commercial, finance and supply rather than around a single number that none of them believe.
Commercial forecasts the opportunity. Finance forecasts the commitment it will defend. Supply forecasts units at a SKU and a location. The three numbers differ, and they should, because they answer different questions.
Most demand planning improvement programmes set out to produce a single number the organization agrees on. It is the wrong objective, and pursuing it produces a figure that is politically acceptable and operationally useless.
The three forecasts exist because three functions carry different risk. Collapsing them removes the information that each function needed.
Three functions hold three numbers and no reconciliation
So the meeting argues about the number instead of about the assumption.
Brand demand is converted to SKU demand by nobody in particular
Which is where a large share of the actual error enters.
Accuracy is reported, bias is not
So a forecast that is consistently wrong in one direction looks merely imprecise.
Patent expiry and biosimilar entry sit inside a trend model
When they are discrete events with uncertain timing, not a slope.
The consensus process has never been tested against a baseline
So nobody knows whether it adds accuracy or only effort.
Shipments are being read as demand
When a distributor building inventory looks exactly like a market that grew.
A revenue forecast is compared against a unit forecast
And the gross to net bridge that explains the gap has no owner.
Inventory position is absent from the demand conversation
So a supply plan is built on demand that is already partly on a shelf.
Compute your bias by product for the last several cycles. Not the error, the signed bias. Take the forecast and the actual for each product for as many past cycles as you hold, keep the sign, and average it. Error will already be reported somewhere. Bias almost certainly is not. If a product shows persistent directional bias across many cycles, that is not a modelling finding and no model change will address it. It is telling you how the forecast is being used, and that is a conversation with people rather than with data.
Statistical method, model selection, uncertainty quantification and validation are covered in depth on our Forecasting and Statistical Modeling page. This page is about the planning process those models feed, which is where most of the recoverable error actually sits.
Step 1
Assess the planning process and its decision points
Step 2
Name the unit and the horizon for every forecast in the process
Step 3
Integrate commercial, prescription, supply and inventory data
Step 4
Standardize product and market hierarchies across functions
Step 5
Harmonize to a common grain, and keep the conversion explicit
Step 6
Establish a demand baseline from clean history
Step 7
Segment products and markets by forecastability
Step 8
Develop statistical forecasts at the grain each decision needs
Step 9
Apply machine learning where the data volume supports it
Step 10
Simulate the events that are not trends
Step 11
Reconcile the three forecasts, and attribute every delta
Step 12
Validate assumptions against an independent view
Step 13
Visualize the plan with its uncertainty and its deltas visible
Step 14
Monitor accuracy, bias, stability and forecast value added
Step 15
Optimize the process, not only the models
Test whether your consensus process beats a naive baseline. Take a naive statistical forecast of each product, the simplest defensible one, and compare its accuracy against your final consensus forecast over the same past periods. This is forecast value added, and it is the only honest test of whether the planning process earns its cost. A material share of organizations find that the consensus process adds effort and subtracts accuracy on at least some product segments. Knowing which segments those are lets you stop reviewing them, which is the rare improvement that saves money and improves the number at the same time.
Statistical and machine learning method detail sits on our Forecasting and Statistical Modeling page. What follows is the planning capability built on top of it.
Commercial demand
Pharmaceutical demand and commercial sales forecasting at brand, market and channel level
Prescription demand modelling with reporting lag and revision behaviour handled explicitly
Patient-based and epidemiology-based forecasting, with the population basis documented
Supply and operational demand
SKU-level demand planning, with the brand to SKU conversion as an owned and measured artifact
Inventory demand analytics including channel and pipeline position
Supply and demand gap analysis at the grain the supply decision is made
Demand sensing from leading indicators, where the indicator has been shown to lead
Scenarios and planning process
New product launch forecasting with falsification triggers attached to each assumption
Scenario planning for entry, patent expiry, access change and promotional change
Forecast reconciliation across commercial, finance and supply with delta attribution
Integrated business planning analytics, including forecast value added measurement
The third column is the point of the table. Knowing which forecast a request belongs to settles most of the argument about what the right answer is.
| Use case | What it answers | Primarily serves |
|---|---|---|
| Launch demand forecasting | What to manufacture and commit to with no history | All three, which is why it is hardest |
| Brand sales projections | What the brand will deliver in value | Commercial and finance |
| Patient-based forecasting | What the population implies for demand | Commercial |
| Prescription volume forecasting | What prescribing behaviour implies, net of reporting lag | Commercial |
| Specialty therapy planning | Demand for low-volume, high-value, lumpy products | Supply, with commercial input |
| Generic or biosimilar entry | Erosion shape and timing as a scenario | All three, as a scenario set |
| Patent expiry forecasting | The cliff, its timing distribution and the floor | Finance and supply |
| Payer access change | Demand effect of a coverage or formulary movement | Commercial and finance |
| Seasonal demand planning | Within-year pattern at SKU and location | Supply |
| Inventory requirement forecasting | What to hold, where, net of channel position | Supply |
| Commercial budget planning | The committed number and what it rests on | Finance |
The technical difficulty here is rarely the forecasting engine. It is that commercial, finance and supply each hold a different product hierarchy, a different market definition and a different calendar.
What each layer contributes
Ingestion from enterprise planning, supply chain, CRM, prescription, licensed commercial, access and inventory sources, each on its own cycle
A landing layer retaining each source as delivered, including the revision history, because a restated history silently changes every baseline built on it
A hierarchy layer holding the three hierarchies and the mappings between them as versioned artifacts, rather than as a translation done inside reports
A baseline layer with known distortions marked: stocking, shortage, prior-period revision and incomplete recent periods
A forecasting layer running statistical and machine learning methods at the grain each decision needs, with method and parameters versioned
A reconciliation layer holding all three forecasts, the deltas between them, and the assumption attributed to each delta
Platforms and systems
Enterprise planning and supply chain: SAP and other enterprise planning platforms, inventory and distribution systems
Commercial: CRM platforms, pharmaceutical sales and prescription datasets
Licensed commercial data sources including IQVIA and equivalents
Patient population, epidemiology and market access datasets
Platforms: Microsoft Fabric, Azure Data Lake, Databricks, Snowflake, AWS
Modelling: Python-based statistical forecasting, AI and machine learning models
Delivery: Power BI and enterprise planning integration
Ask which product hierarchy the last planning meeting used. Put the commercial product hierarchy, the finance hierarchy and the supply hierarchy side by side for one brand and count the places they disagree: how products roll up, what counts as a market, where presentations sit, and when a period starts and ends. Then ask which of the three the last reconciliation was performed in. In most organizations the answer is that the translation happened in a spreadsheet owned by one person. That is both the integration problem and the explanation for most of the arguing.
Governance here is mostly about reproducibility and about the licensed and patient data the forecast consumes. Which privacy frameworks apply, and what each dataset permits, is determined by your legal, privacy and compliance functions.
Reproducibility
- Forecast version control covering numbers, assumptions, method, parameters, data vintage and approval
- Standardized product and market hierarchies with versioning and change control across all three functions
- Historical data integrity, with restatements held as versions rather than as corrections
- Assumption documentation as a required artifact of every published forecast, not a supporting note
Model transparency
- Method and parameter choices recorded with the forecast, so a model change is distinguishable from a market change
- Model validation performed to the standard your analytics function sets, with evidence retained
- Responsible AI governance covering training data eligibility, drift monitoring and human approval of the published plan
Data rights and privacy
- Licensed commercial and patient data restrictions encoded per source: audience, grain, combination, derivation and retention
- GDPR and other applicable privacy regulations, and HIPAA where protected health information is handled, as your privacy function determines
- Patient-derived inputs used at cohort level only, with no individual-level data entering the commercial planning estate
- Model training eligibility recorded per dataset at onboarding rather than decided at project time
Process control
- Scenario approval workflow, so a scenario cannot become the plan without a named approver
- Scenarios structurally distinguished from the approved plan, because an upside or downside case circulating in a deck will be quoted as the plan by somebody within a fortnight
- Role-based access over forecasts, scenarios and the commitment number, which is frequently price-sensitive internally
- Override logging, because an unlogged manual adjustment is the most common untraceable source of bias
Accuracy is the one everybody reports. On its own it tells you least, because it does not tell you the direction, the consistency or whether the process is adding anything.
| Failure | How it presents |
|---|---|
| Persistent directional bias | Accuracy looks acceptable while every period misses the same way |
| An unlogged override pattern | The statistical forecast is sound and the published plan is not |
| Instability between cycles | The same future period moves repeatedly, and supply stops trusting it |
| Negative forecast value added | The consensus process is less accurate than a naive baseline |
| A hierarchy changed without versioning | Forecast and actual are compared at different grains |
| History was restated | The baseline moved, and the forecast change has no other explanation |
| Conversion error misattributed | Brand forecast was close, SKU plan missed, and the brand gets blamed |
| Shipments were read as demand | Channel stock build looks like growth, and the correction looks like decline |
| Revenue missed on mix, not volume | The demand forecast gets revised to fix a problem it did not cause |
| Aggregate accuracy hid SKU error | The brand number looks fine while the supply plan is unusable |
| Model drift | The method stopped fitting and the output is still being published |
| A scenario quietly became the plan | A downside case entered the commitment with no approval recorded |
The outcomes below follow from one change: the three forecasts kept, reconciled, and every delta attributed to a named assumption.
Outcome measures
Forecast reliability improved, and improvable, because a miss can be attributed to a specific assumption rather than to the market
Bias reduced, which is usually a larger and cheaper gain than accuracy improvement and is rarely attempted
Supply and demand coordination improved, with the brand to SKU conversion owned and measured separately
Market scenarios evaluated as scenarios, so an entry or access event is a switch rather than a rebuild
Manual forecasting effort reduced, including review steps that measurably subtract accuracy
We will compute bias and forecast value added on your own history, by product segment, and tell you where the planning process is earning its cost and where it is not. If it is working, we will say so.
CaliberFocus builds demand planning for pharmaceutical and biotechnology organizations around the reconciliation between commercial, finance and supply. We keep all three forecasts, attribute every delta to a named assumption, own and measure the brand to SKU conversion, model entry and access events as scenarios, and measure bias and forecast value added so the planning process can prove what it adds.
- Commercial, finance and supply forecasts reconciled with every delta attributed
- Bias and forecast value added measured, not only accuracy
- Brand to SKU conversion owned and measured separately
- Versioned forecasts with assumptions, method and data vintage retained
Security & Compliance
