Forecasting Software Fits the Curve. The Error Is in the Assumptions.

On 2 February 2024 AbbVie reported that US net revenue for Humira fell 34.7 per cent in 2023, to $12.16 billion, in the drug's first year facing biosimilars. Yet in November 2023, IQVIA data prepared for the Biosimilars Council showed biosimilars holding just 1 per cent of adalimumab volume. Other recent biosimilars had averaged 22 per cent eleven months after loss of exclusivity. The brand kept most of its patients and lost its price. IQVIA's later summary is that Humira “maintained brand share post-LOE but lost 60% of net sales” through discounting.

That gap is the real subject of any conversation about pharma forecasting software. A share-erosion curve borrowed from earlier biosimilars would have got the volume badly wrong and still missed the revenue. What moved adalimumab was a run of decisions: rebate contracts, a pharmacy benefit manager's formulary, and a biosimilar sold under that manager's own label. No curve-fitting tool predicts those. The useful questions are what each kind of forecast needs, what the vendors really sell, and where AI earns its place.

What does a pharma forecast actually have to model?

Most commercial forecasting in pharma is one of three jobs, which share software and little else.

A patient-based forecast is a chain of multiplications, so steps that each look reasonable can compound into a large miss. For pipeline assets the chain starts with a probability of success, and Phase 2 success rates for AI-discovered molecules look like everyone else's.

Rx forecasting is the opposite case. It runs on prescription audits and claims data, so there is plenty of history, and the question is how far the recent past predicts the next quarter. On the supply side the same idea is called demand sensing, and its limits show up in what DSCSA data can and cannot predict.

Pre-LOE forecasting sits in between. The brand has history, but the inputs that decide the erosion are other organisations' future choices.

How wrong are pharma forecasts, and why?

The most cited evidence is old but has not been replaced. Three McKinsey authors, writing in Nature Reviews Drug Discovery in October 2013, compared sell-side consensus forecasts with actual sales. Their methods note explains that the consensus was the average of up to six analysts, taken from EvaluatePharma. Of 260 drugs with a consensus forecast in the year before launch, 54 were forecast more than 160 per cent above their actual peak sales. IQVIA, summarising the same study in 2020, says actual peak sales differed from predictions by 71 per cent a year before launch, and forecasts were still 45 per cent off six years after.

These are analyst forecasts from 2002 to 2011, not company ones. Treat them as a measure of how hard the problem is, not of your own team.

A rare-oncology launch case in our library shows the problem from the other side. PharmaForceIQ reports that “Year one revenue was triple launch forecast” and presents this as a campaign result. Read as forecasting evidence, the launch forecast was out by a factor of three. With no control group, nobody can separate the campaign's effect from a forecast that was simply too low.

The forecasting research explains why tools have not closed the gap. The M competitions have compared forecasting methods since 1982. In the first three, simple statistical methods beat complex ones. In M4 in 2018, only two sophisticated methods beat the simple ones. In M5, run in 2020 on Walmart sales, all 50 top methods were machine learning. The organisers' conclusions stress what was different about the data: many related series, plus prices, promotions and calendar events. A launch or a loss of exclusivity offers neither.

Where AI and machine learning genuinely improve a forecast

In M5 the winning team, an equal-weighted combination of LightGBM models, beat the best statistical benchmark by 22.4 per cent across 42,840 series over 28 days. But only 415 of 5,507 teams, 7.5 per cent, beat that exponential smoothing benchmark at all. For the top 50 teams, the average gain was about 40 per cent at the top of the hierarchy and about 3 per cent for single products in single stores.

Retail is not pharma, but the shape carries over. Weekly prescriptions by brand, region and channel, with known formulary and promotion calendars, look like M5. Prescriber-by-pharmacy-by-day does not.

The pharma case studies fit that pattern, with the usual caveats about vendor claims:

• Amazon Pharmacy and AWS. AWS reports a daily mean absolute percentage error of 5 per cent against an industry target of 10 per cent or less. AWS is writing about a sister company, gives no before figure and does not say at what level the error is measured. The Amazon Pharmacy case does show that short-horizon operational demand suits machine learning.
• Beghou for an oncology brand. The model separates refill cadence and insurer delays from real trend in daily sales. No accuracy figure is published. The use is sensible: it stops a commercial team chasing noise.
• Axtria for a women's health portfolio. Sub-national demand forecasts with demographic and event variables. The page claims “less than average” prediction error with no number and no baseline.

Large language models help differently. They read rather than predict. In the Novartis case with Dataiku, time to insight from hundreds of 20-page interviews with healthcare professionals fell from 21 days to two. In a separate spreadsheet workflow, data ingestion fell from a week to six hours. Both are time measures, and the Dataiku page gives no accuracy figure. Faster reading means fresher share assumptions, which is worth having. But a faster input is not a more accurate one. IQVIA's own forecasting page sells its machine learning on speed too: “minutes, not hours”.

Where machine learning cannot help is the adalimumab problem. A model trained on recent biosimilar launches would have expected around 22 per cent share at month eleven. The actual figure was 1 per cent. Then CVS Caremark, whose parent had set up a biosimilar subsidiary called Cordavis in August 2023, removed Humira from its major national commercial formularies from 1 April 2024. Biosimilars reached 18 per cent of adalimumab volume in April 2024, and the Cordavis-branded Hyrimoz accounted for over 70 per cent of biosimilar prescriptions written in April and May. No training set contained that decision.

A forecast model in use also drifts. Commercial forecasts are not GxP systems, but the monitoring plan for a deployed model transfers well. So does freezing and versioning the model, so that variance comes from the market, not a silent method change.

What pharma forecasting companies and software actually sell

The market sorts into four groups. This is not a ranking, and the examples are only those with a public page we could check.

The acquisition note says Norstella plans to use generative AI to synthesise insights across J+D's models, a plan rather than a measured result. The J+D oncology case in our library describes a biotech moving from "finance-led" models to a patient-flow method before a Phase III launch. That is the right change for a pre-launch oncology asset, but the public page gives no accuracy figures. It shows a method adopted, not a forecast improved.

Ask every group the same question: what was the accuracy on periods the model did not see, against a simple benchmark? Without that, the claim is about speed or convenience, which may still be worth paying for.

What the published erosion evidence says

For small molecules the evidence is good and reasonably consistent.

Biologics are different. AbbVie's full-year 2024 results show US Humira down a further 41.3 per cent, to $7.14 billion. IQVIA's tracking attributes part of the fall to AbbVie moving patients to its newer drugs. The number of former Humira patients switching to Skyrizi or Rinvoq grew 120 per cent from 2022 to 2023.

In Europe the first question is whether anyone is coming. IQVIA's January 2026 report for the European market counts about 100 biologics losing European exclusivity by 2032. Of these, 79 per cent have no biosimilar in development, and IQVIA expects only about 10 per cent to face biosimilar competition. The same series observed in 2015 and 2016 that the link between biosimilar share and price reduction is weak. Its price figures are list prices, before rebates.

Small-molecule erosion follows a documented curve once you know the date and the number of entrants. Biologic erosion depends on who enters, which payers move and what the brand owner does. Those are inputs to a forecast, not outputs of a model.

What a defensible pre-LOE forecasting framework looks like

A defensible framework is mostly a discipline about evidence, in eight parts.

1. Date the event from primary records. For small molecules that means the Orange Book, FDA's Paragraph IV certifications list, the 30-month stay that applies if the brand sues, and any settlement entry dates disclosed in filings. For biologics it means first licensure: 42 U.S.C. 262(k)(7) bars a biosimilar approval taking effect until 12 years after it.
2. Count the entrants by date. The number of Paragraph IV filers and tentative approvals drives the price curve. For biologics, check the pipeline before assuming any competition.
3. Choose the archetype before the analogues. IQVIA's questions are the right ones: small molecule or biologic, how many competitors and when, and whether the drug goes through retail, specialty pharmacy or a clinic. A hypertension pill can be substituted automatically within weeks. A clinic-administered oncology infusion can keep share longer. Match analogues to the archetype, not the therapy area.
4. Forecast volume and net price separately. Humira is the warning. A share curve alone hides a defended-volume, collapsed-price outcome.
5. Treat payers and channels as scenarios. A formulary decision is a step change, not a trend. Give each major payer a stated assumption and a range.
6. Treat your own company's response as an input. An authorised generic, brand-for-generic contracting and moving patients to a successor product are choices your company makes. Each changes the curve, and each needs lead time.
7. Put the Medicare price date on the same timeline. A negotiated price can bite before the patent does.
8. Attach a trigger to every assumption. A trigger is an observable event that moves the forecast to another scenario: a Paragraph IV filing, a tentative approval, a settlement, an at-risk launch, a formulary announcement. Watching for these is reading work. Language models help there, provided a person confirms each change before the forecast moves.

IQVIA's LOE planning advice is that contracting, promotion and sales strategies need to start at least four to five years before expiry, and that many brands start only two years out.

What binds and what does not

Position current at 12 October 2026. A commercial forecast is not a GxP record, so computerised-system rules do not apply to the model itself. What binds is the law that sets the dates and prices the model has to use.

Two cautions. First, the FDA drafts point towards cheaper biosimilar development, which could mean more entrants for later biologic LOEs. IQVIA expects such streamlining to help at the margin without closing the gap. Model it as a scenario, not a base case, until the guidances are final. Second, the EU reform is still completing adoption. The Council's note of 21 September 2026 puts its first-reading position forward, with Parliament expected to approve it at second reading. Until the texts are published and in application, current EU protection rules govern.

What this means in practice

In lead-time order, starting with the cheapest step:

Measure your own error first. Take your last launch forecast and your last LOE forecast and score them against actuals. Break the miss down by assumption: entry date, entrant count, share, net price, payer. This takes days, and it tells you whether to spend on software, data or judgement.

Five or more years before expiry, build the event timeline. Pull patent, exclusivity and Paragraph IV records from FDA sources, settlement dates from filings, and biologic licensure dates. Add Medicare negotiation eligibility on the same chart.

Four to five years out, fix the archetype and the company's own choices. Choose analogues by archetype. Decide on authorised generics, contracting posture and any successor product, because each changes the curve and none can be started late.

Two to three years out, write payer and entrant scenarios with triggers. Give each scenario an observable trigger and a named owner, and set up the watching.

In market, use machine learning where the data is dense. Weekly Rx forecasting by brand, region and channel is a fair test. Hold out recent periods and compare against a simple benchmark. If the model cannot beat exponential smoothing on your own data, do not buy it.

When buying, ask for held-out accuracy and full assumption visibility. Treat any accuracy claim without a baseline as a claim about speed, then decide whether speed is what you lack.