Scenario Modeling in Pharma: Translating Uncertainty into Decisions
Pharma companies are already good at building forecast models. Scenario analysis, sensitivity testing, and Monte Carlo simulation are standard tools.
The harder problem often comes before the model is built — and after it leaves the spreadsheet.
Before the model, different stakeholders may be asking different questions: How much demand could the market generate? How much can we manufacture? What should we put in the budget? What number can leadership communicate with confidence?
After the model is built, those different questions are often compressed into one number. A board member asks, “So what are we actually putting in the plan?”
The range disappears — and with it, the information leadership needed to decide well.
Eli Lilly's Zepbound launch offers a useful illustration of why that matters.
The Same Drug, Two Opposite Misses
Zepbound was approved by the FDA in November 2023 for chronic weight management and reached U.S. pharmacies weeks later. Early analyst expectations put 2024 U.S. sales at around $2 billion. Lilly reported $4.9 billion — more than double what was expected. (1, 2)
Supply, not demand, was the constraint leadership was already flagging. Lilly told investors on their Q4 2023 earnings call that demand would continue to outstrip tirzepatide supply throughout the year. By Q1 2024, they reiterated the same supply issues, even as they still managed to beat expectations and raised guidance. (3, 4) The uncertainty wasn't whether the market wanted the drug; it was how much of that demand the company could actually convert into shipped product.
Then the story changed.
In the fourth quarter of 2024, Zepbound generated approximately $1.9 billion in revenue, below analyst expectations of roughly $2 billion or slightly more, depending on the consensus date. Lilly subsequently described supply and channel dynamics as contributors to quarter-to-quarter variability in tirzepatide revenue. (5, 6)
In fact, this pattern is not unique to Zepbound. A review of 130 major drug launches found that fewer than 24% were within 30% of their pre-launch analyst revenue forecast by Year 5; 27% exceeded forecast by at least 2x, while 24% delivered less than half of expectations. (7) The findings are striking not because forecasts are missing, but because they miss in both directions and often by wide margins.
The takeaway isn't that Lilly's forecasters got it wrong. It's that forecast error stays stubbornly common even with sophisticated tools, because the miss can come from different places — different drivers each time.
A Single Number Hides Both the Drivers and the Decisions Riding on Them
In the above Zepbound example, the forecast comparisons are not identical: one is a full year outcome versus early launch expectation, while the other is a quarterly result versus consensus. But together they illustrate an important point. A forecast can miss because underlying demand was underestimated or it can also miss because supply, inventory, pricing, access or timing changed the amount of demand that ultimately became reported revenue.
Revenue is the outcome at the end of a chain:
Revenue
These drivers don't necessarily move together. Strong demand doesn't automatically translate into higher reported revenue if supply is constrained. Revenue can temporarily rise because inventory builds in the channel. Access changes can increase the number of patients treated, while pricing or channel mix determines how much revenue is ultimately realized from those patients.
Lilly's own disclosures show this clearly — 2025 U.S. Zepbound revenue grew 175%, driven by higher demand but partly offset by lower realized prices. (8) Revenue tells us what happened; the underlying drivers tell us why.
But even once revenue is broken into its drivers, a second question remains: who decided what those drivers should be in the first place — and for what decision?
A single forecast is often asked to answer several different decisions at once, and each one pulls the number in a different direction:
Capacity planning
A capacity-planning decision needs a number high enough to justify building manufacturing capacity ahead of demand.
Budget
A budget decision needs a more conservative number that won't leave finance exposed if growth is slower than expected.
Communications
A communications decision needs a number that supports a credible growth story to the board or to investors.
These aren't three biased versions of one true forecast. There are three different decisions, each of which genuinely requires a different assumption set — and in some cases, a different forecast entirely. Treating them as one number that should satisfy all three is where the trouble usually starts, well before anyone opens a spreadsheet.
How to Build a Scenario Model That Changes the Decision, Not Just the Slide
A scenario model earns its place only if it informs what decision is being made — which team gets funded, how much capacity gets built, what the board is told to expect. Here's how to build one that does that instead of dissolving into a single number on the way to the meeting.
Separate the forecast from the target.
A forecast says what's likely, based on evidence. A target says what the organization wants to achieve. A budget says what it's prepared to spend. These are different things and should be allowed to disagree. When leadership needs a single planning number, that's fine — but it shouldn't be relabeled as “the forecast.”
Make every assumption traceable to the decision it feeds.
Every major input should have a clear source — observed data, external evidence, internal data, expert judgment, or a management decision — and a clear link to which decision it's meant to inform. In practice: “initiation rate: internal prescriber survey, n=200” versus “initiation rate: matched to a comparable GLP-1 launch” versus “initiation rate: management judgment, no external data” carry very different confidence levels, and a reviewer should be able to see which one a number rests on. “That's roughly what analysts expect” is not a source — it's a consensus, and it should be labeled as one, not treated as evidence.
Let the decision decide how many scenarios you need — not a template.
Don't default to three or five scenarios out of habit. Start by identifying which uncertainties could actually change the decision on the table. If one uncertainty dominates — as supply capacity did for Zepbound in 2024 — build the model around stress-testing that one thing deeply, instead of spreading attention evenly across five inputs that don't matter equally to the decision at hand.
Test which drivers matter before running a full simulation.
Run simple sensitivity checks first — move one variable at a time and see what happens. Only once you know which two or three drivers actually move the decision should you build a full Monte Carlo simulation around them. Running thousands of iterations on a poorly understood set of assumptions just makes weak guesses look statistically impressive.
Model how assumptions move together, not just their individual ranges.
If higher demand is likely to strain supply, or broader access is likely to lower realized price, say so explicitly in the model. Treating linked variables as independent understates real-world constraints on the decision — this is one of the most common weaknesses in published forecasting reviews.
Bring the range into the room, not just the number.
This is where most of the value is lost. When the model reaches a leadership meeting, show the distribution of outcomes, the central estimate, and — most importantly — what would have to be true for results to land above or below that range. Lilly's own actions show what that looks like in practice: having flagged supply as the dominant constraint on its Q4 2023 call, the company announced an additional $5.3 billion to its Lebanon, Indiana site in May 2024 — more than doubling its investment there to $9 billion — specifically to expand tirzepatide API capacity. (9, 10) That wasn't a forecasting exercise; it was a capacity decision made because the range of possible demand outcomes had already made clear which constraint would decide the outcome. Don't let the conversation collapse into “just give us one number” before it reaches that kind of decision. A single number thrown into a plan quietly becomes a target, and every future result gets measured against it, even though it was never meant to be a promise.
The goal is not to eliminate uncertainty or simply make the range wider or narrower. It is to make the range meaningful — so management can understand what could happen, what would drive it, and make a better decision when the future is uncertain.
Prepared by PharmSight Research & Analytics for informational purposes only. Numbered citations link to the underlying sources.