The forecast is a set of decisions, not a number
A sponsor asking for a demand forecast on a vaccine still in clinical development is rarely asking what sales will be in year five. They are asking whether to fund a late-stage trial, how much fill-finish capacity to reserve, what price to defend, or whether a partner's offer is fair. Each of those decisions has a threshold, and the forecast is only useful if it tells the sponsor which side of the threshold they are likely on and what would move them across it.
That changes how the work is built. A single point forecast invites a single argument about whether it is too high. A forecast built as a chain of explicit assumptions, each with a low and a high case and a named source, invites the argument that matters: which assumption is carrying the decision, and how much do we know about it.
Throughout this piece we use one running example: a hypothetical vaccine against a mosquito-borne virus that causes seasonal illness in some regions and is a risk for travelers. The example is illustrative. No figures here describe a real product or a real disease forecast.
Start with who could be vaccinated
The first input is the eligible population, and it comes in two forms that behave very differently.
Age-based populations are defined by birth year: an infant cohort, adolescents at a given age, adults over a given age. They are easy to count from census estimates, they replenish every year, and they are reachable through existing visit schedules such as well-child checks or annual adult visits. When a vaccine fits into an age-based slot, the population is large and the counting is not where the risk lives.
Risk-based populations are defined by a condition, an exposure or an occupation: people with a particular chronic disease, laboratory workers who handle the virus, military personnel deploying to endemic regions, travelers to affected areas, residents of counties with local transmission. These are harder to count, and the count depends on the data source. Claims data finds diagnosed conditions but misses the uninsured. Travel statistics count trips, not travelers, and not the subset who visit a clinic before leaving.
For the mosquito-borne example, the population list might read: travelers to regions with active transmission, residents of areas with recurring local outbreaks, laboratory staff, deploying service members, and possibly older adults if severe disease concentrates in that group. Each segment gets its own row in the model, its own count, its own source and its own uncertainty. Merging them early is the most common way a forecast loses the ability to explain itself later.
Epidemiology inputs, and what they are actually for
Analysts new to vaccine forecasting often spend the first month on disease incidence. Incidence matters, but mostly indirectly. Disease burden drives whether a recommendation is made, how broad it is, and what price a payer will accept. It rarely drives demand directly, because people get vaccinated against a disease they are unlikely to encounter if their clinician recommends it and their plan covers it.
The epidemiology inputs worth getting right are these: incidence by age and geography, the share of infections that become severe, hospitalization and death among severe cases, the seasonality of transmission, and how much of the burden falls on the groups a recommendation would target. For many infectious diseases, surveillance counts only a fraction of infections, so the burden estimate will itself be a range. That range feeds the recommendation scenarios, not the uptake curve.
There is one place epidemiology does drive demand directly: outbreaks. A visible local outbreak produces a short spike in demand among people who would otherwise never have sought the vaccine. We treat that separately, later in the build.
Recommendation scenarios drive everything
In the United States, the type of national recommendation a vaccine receives shapes coverage, clinician behavior, reminder systems and standing orders more than any other single factor. Before modeling uptake, define the recommendation scenarios explicitly. Four are usually enough.
Routine, age-based. Everyone in an age group is recommended to receive the vaccine. This is the strongest case. Electronic health record prompts, quality measures and standing orders tend to follow, and the vaccine becomes part of the visit rather than a conversation.
Risk-based. The recommendation applies to people with specific conditions, exposures or occupations. Demand now depends on someone identifying the person as eligible, which means the clinician has to know the indication and the record has to carry it. Uptake in risk-based groups is almost always lower than in routine groups, and slower to build.
Shared clinical decision-making. The vaccine may be given based on a discussion between the clinician and the patient. There is no default. In our experience this category produces the widest range of outcomes, because uptake depends on whether clinicians bring the topic up at all, and many do not without a prompt.
No national recommendation. The vaccine is licensed and available, perhaps recommended for travel by travel medicine guidance, but not part of routine practice. Demand is mostly private, often cash pay, and concentrated in travel clinics and pharmacies.
Coverage rules for commercial plans, public programs for children and Medicare have each been linked to national recommendations in different ways, and those rules have been in motion. Check the current state of each rule at the time the forecast is built, write down the version you used, and treat a change in the linkage as a scenario rather than a footnote.
Uptake assumptions come from analogs, chosen with care
A vaccine that does not exist has no uptake history, so the forecast borrows one. The choice of analog is the single assumption most likely to be challenged by a reviewer, so it should be made deliberately and documented.
A good analog matches on the dimensions that drive uptake: the same recommendation type, a similar target population, a similar dose schedule, a similar setting of care and a similar payer channel. A routine adolescent vaccine is a poor analog for a risk-based adult vaccine even if the disease looks similar. A travel vaccine given in a single dose at a pharmacy is a poor analog for a two-dose series that requires a return visit.
Pick two or three analogs rather than one, and use them to set the low and high cases rather than averaging them. Record the peak uptake each reached, the time it took to get there and the conditions at the time: whether coverage was in place at launch, whether supply was constrained, whether the recommendation changed partway through. Multi-dose schedules need their own adjustment, because completion of the series is always lower than initiation, and the drop between doses is a separate assumption.
Price scenarios and who pays
Price enters a vaccine forecast twice. It sets revenue directly, and it affects whether a recommendation is made at all, because recommending bodies and payers look at cost-effectiveness. A high price can narrow a routine recommendation to a risk-based one, which can reduce volume by more than the price gained. That interaction belongs in the model explicitly, as linked scenarios rather than independent dials.
Then split volume by payer channel, because each one has its own price, its own purchasing behavior and its own timing.
- Public purchase for children. Vaccines for eligible children are bought through federal contracts at negotiated prices and distributed through state immunization programs. Volume is large and the price is lower than the private sector price.
- Commercial insurance. Clinics and pharmacies buy the vaccine and are reimbursed by the plan. Uptake here depends on whether providers are willing to stock it and whether reimbursement covers their cost.
- Medicare. Some vaccines are covered under the medical benefit and others under the drug benefit, which changes whether the dose is given in an office or at a pharmacy.
- Government and institutional buyers. The military, public health departments and research institutions buy directly, often in a few large orders a year.
- Cash pay. Travel vaccines without coverage are often paid out of pocket. Price sensitivity is real, and the channel is small but steady.
- International. Procurement agencies and national programs abroad buy on tender, usually at tiered prices. Treat this as its own model if it matters to the decision.
For the mosquito-borne example, a large share of early demand might sit in cash pay and travel clinics, with a much larger but uncertain public and commercial channel that only opens if a risk-based recommendation for residents of affected areas follows.
Outbreaks make demand lumpy
For vaccines against diseases with outbreak dynamics, the baseline forecast is only half the picture. An outbreak in a populated area can produce months of demand in a few weeks, followed by a collapse when transmission ends. A government may place a large one-time order to build a reserve, then nothing for years.
Do not average outbreak demand into the baseline. A smoothed number describes a year that never happens, and it misleads capacity planning in both directions. Model outbreaks as an overlay: an annual probability of a significant outbreak, a plausible size range for the demand spike, and a constraint on how much of that spike could actually be supplied given manufacturing lead times. That last constraint is often the binding one, and it is the one sponsors most want to see.
The scenario drivers in one table
Every driver should appear in the model with a low case, a high case and the evidence behind each. This table is the backbone of the forecast and the page a reviewer will read first.
| Driver | Low case | High case | Where the evidence comes from |
|---|---|---|---|
| Recommendation | Shared clinical decision-making or travel-only | Routine for an age group plus risk-based groups | Burden and cost-effectiveness evidence, precedent for similar vaccines |
| Eligible population | Narrow risk definition, counted from claims | Broad definition including residents of affected areas | Census estimates, claims data, travel statistics, occupational counts |
| Peak uptake | Lowest credible analog | Highest credible analog with full coverage at launch | Analog launches matched on recommendation type and setting |
| Time to peak | Slow provider stocking, coverage delayed | Coverage in place at launch, pharmacy access from day one | Analog timing, payer and pharmacy rules |
| Price and channel mix | Price pressure narrows the recommendation | Price supports a broad recommendation | Analog prices by channel, cost-effectiveness thresholds |
| Outbreaks | No significant outbreak in the horizon | One large outbreak plus a reserve purchase | Historical outbreak frequency, supply constraints |
Drivers That Usually Move the Answer Most
Relative influence on the forecast, in our judgment. Not measured data. Outbreaks can dominate for a vaccine whose demand is mostly reactive.
The build, in the order it should be done
The sequence matters because each step constrains the next. Teams that start with the uptake curve end up reverse engineering populations to fit it.
Forecast Build, Illustrative Eight Weeks
The first step is the one most often skipped and the one that makes the rest useful. A forecast built without named thresholds tends to come back as a central estimate with a band around it, which leaves the sponsor to work out what it means for their decision on their own.
Present ranges against the thresholds that matter
Sponsors do not need the forecast to be right. They need to know whether the decision is safe across the plausible range, and if it is not, which assumption would have to be true for it to fail.
Three presentation tools do most of the work. A scenario table with three to five named, internally consistent scenarios, each with peak annual doses, time to peak and revenue by channel. A tornado chart showing how far the key output moves when each driver swings from its low to its high case, holding the rest at base. And a threshold view: for each decision, the share of scenarios that clear the bar, and the single assumption that flips the answer.
For the mosquito-borne example, the threshold view might show that a commercial launch clears the sponsor's return target in every scenario that includes a risk-based recommendation for residents of affected areas, and in none that are travel-only. That one sentence is worth more to the sponsor than the full revenue curve, because it tells them where to spend the next dollar of evidence: on the burden data that would support the broader recommendation.
Keep scenarios internally consistent. A high-price case paired with a routine recommendation and peak analog uptake is a scenario nobody believes, and including it stretches the range in a way that makes the whole forecast look less serious.
Before you trust the forecast
Whether you built it or bought it, run the forecast through these checks before it goes in front of a decision-maker.
- Every eligible segment has its own row, source and range
- Recommendation scenarios are written out in words, not implied by an uptake percentage
- Each analog is named, with the reason it was chosen and the conditions at its launch
- Series completion is modeled separately from first-dose uptake
- Price and recommendation are linked, not independent
- Outbreak demand sits in an overlay, not averaged into the baseline
- The coverage rules used are dated, with the version cited
- The write-up names the assumption that flips each decision
A forecast that passes these checks can be defended line by line. One that does not will be defended by its author's confidence, which lasts until the first reviewer asks where the peak uptake came from.
Where vaccine forecasts usually go wrong
The failures repeat. The eligible population is counted once from a convenient source and never challenged. A routine-recommendation analog is applied to a risk-based indication because it was the analog the team had on hand. Uptake is modeled as a percentage of the whole population rather than of the people who reach a provider who stocks the vaccine. The outbreak year from three years ago is quietly treated as normal. The forecast reaches the board as a single line, and the range lives in an appendix nobody opens.
Each of these is fixed by the structure above: separate segments, explicit scenarios, deliberate analogs, an overlay for outbreaks, and a presentation built around the decision rather than the curve.
Frequently asked questions
Far enough to support the decision in front of you, with ranges that widen as the horizon extends. A forecast made before late-stage trial results should be honest that the recommendation, price and label are all unknown, and should show how demand changes across each. The value is in identifying which uncertainties matter, not in narrowing the year-five number.
Usually the type of recommendation the vaccine receives, because it shapes coverage, clinician behavior and whether the vaccine becomes a default part of the visit. The eligible population definition and the choice of analog for peak uptake come next. Disease burden matters mainly through its effect on the recommendation.
Match on what drives uptake: recommendation type, target population, dose schedule, setting of care and payer channel. Choose two or three analogs to set low and high cases, document why each was chosen, and record the conditions at its launch, such as whether coverage was in place and whether supply was constrained.
No. Model it as a separate overlay with an annual probability, a plausible size range and a supply constraint. Averaging outbreak demand into the baseline produces a smooth number that describes a year that never happens and misleads capacity planning in both directions.