| Vose Software

Industry: Healthcare and Epidemiology
Product: ModelRisk
Application: Medical Research Portfolio — Stochastic NPV Across 7 Themes


A $42M Research Office Asking Whether the Portfolio Pays for Itself

An academic medical centre allocates a $42M annual indirect-cost budget across seven research themes. The dean's deterministic spreadsheet — expected publications × expected R01 hit rate × expected translation rate × expected license value — reports a comfortably positive five-year NPV and concludes the portfolio is value-accretive. The Monte Carlo simulation on the same inputs tells a sharper story: the mean portfolio NPV is barely above break-even at roughly +$2M, the median is about −$112M, and the probability of a positive NPV is only ~30%. The thin positive mean is carried entirely by a fat right tail — the rare years in which one or two themes land major license events. The deterministic answer is right on average and wrong on every other useful question: what's the probability the office breaks even? What's the downside the trustees should be prepared for? Which themes are doing real lifting versus which are being carried by one improbable jackpot?

Research-portfolio NPV distribution — academic medical centre, 7 themes

The shape is the whole point. The mean sits above the 70th percentile — the tell-tale signature of a fat-right-tail process where the average is not the typical outcome. The P10 is about −$175M (a five-year stretch where licensing largely fails to materialize within the discount horizon — the full indirect spend with nothing to show for it), the P50 is about −$112M (the typical portfolio underwater in PV terms), and the P90 is about +$289M (one or two themes produce major license events). The probability of a positive NPV is ~30% — the portfolio is roughly a two-in-three bet to lose money in PV terms over five years, redeemed only by the size of the upside when it lands.

The research-office leadership team rebuilt the portfolio-NPV view in ModelRisk. The deterministic dashboard still runs — it is the basis for the annual report to the trustees — but every quarterly portfolio-review cycle now starts with the simulated NPV distribution.

Four uncertainties stacked

Research-portfolio cash flow is a four-layer probabilistic stack:

Publication. Annual probability of producing at least one publishable result per theme — between 0.22 (Rare-disease genomics, slow but high-leverage) and 0.40 (Infectious disease Dx, fast cycle time). Modelled as Bernoulli per year, conditioned on grant continuity.

Grant renewal. Probability that a published theme attracts follow-on NIH R01 funding within the horizon — Beta(20, 8) with mean ~71%, calibrated against the centre's 8-year hit-rate. Beta is the right family here: bounded on [0, 1], event-count parameterized.

Translation. Probability that an R01-funded theme produces a licensable IP asset — between 8% (mental-health digital, low IP density) and 22% (cardiac devices, established translation pathway). Beta posteriors per theme.

License value. Conditional on translation, expected license value is LogNormal. The right family is critical: license outcomes are bounded below at zero, right-skewed, and dominated by rare large wins. For rare-disease genomics the simulation anchors each licensable asset on a LogNormal with σ_log = 1.20 and scales it to the size of a sustained, multi-year licensing programme — so a theme that translates lands a heavy-tailed stream rather than a single small payment, and the rare large wins dominate the portfolio.

A deterministic spreadsheet that multiplies these means produces an answer that is mathematically wrong in two ways: it ignores the cliff at zero (translation = 0 → license value = 0, regardless of the conditional mean), and it ignores the LogNormal tail that dominates the portfolio.

Per-theme contribution

Per-theme expected license value vs annual indirect spend

The chart reveals the structural problem. Rare-disease genomics consumes 18% of indirect spend ($7.6M/year) and contributes the highest expected license value (~$103M) with a P90 of well over $150M — but its expected contribution is dominated by the tail. Mental-health digital consumes 10% of spend ($4.2M/year) and produces an expected license value of only ~$9M, which does not cover its own indirect costs across the horizon at the modelled rates. Oncology imaging (~$80M) and Cardiac devices (~$67M) are the steady mid-tier producers.

What actually drives the portfolio NPV

Tornado — drivers of portfolio mean NPV

The single biggest mover is grant-renewal probability — raising the renewal Beta from 0.5 to 0.85 moves portfolio mean NPV by ±$22M. Rare-disease license tail ranks second: tightening the LogNormal σ_log from 1.0 to 1.4 changes mean NPV by ±$18M, almost entirely on the upside. Time-to-license (5 vs 9 years from grant to license) is third — the discount-rate compounding bites disproportionately at the long-cycle themes. Oncology translation rate ranks fourth. That ranking is what told the leadership team that the highest-leverage operational lever was the grant office (improving renewal rates), not portfolio rebalancing.

Three rebalance scenarios

Portfolio rebalance scenarios — NPV CDF

Moving 4 percentage points of indirect spend from Mental-health-digital to Rare-disease-genomics raises mean NPV by roughly $10M — but almost entirely in the right tail, leaving the probability of breaking even essentially unchanged at ~30%. Moving 4 percentage points from Cardiac-devices to Oncology-imaging lifts mean NPV by only ~$2M and likewise barely moves P(NPV > 0) — the steadier but smaller rebalance. The simulation is what allowed the trustees to see those two options as different in shape, not just different in mean: the same headline metric can hide whether a reallocation buys tail upside or central-tendency improvement.

What the model changed

  • Grant office headcount expanded (one additional pre-award officer at $145,000 loaded) as the highest-leverage NPV lever — the simulated +0.05 renewal probability improvement is worth ~$3M in mean NPV terms.
  • Mental-health-digital allocation cut from 10% to 6% of indirect spend; the freed capacity reallocated to oncology imaging and rare-disease genomics in a 1:1 ratio, lifting mean NPV by roughly $10M while the breakeven probability holds near ~30%.
  • Five-year translation milestones added for each theme; themes failing to hit milestones on schedule are reviewed for closure rather than maintained at flat funding.
  • Trustee reporting restructured to include the simulated P10/P50/P90 alongside the deterministic mean, and the P(NPV > 0) metric is now the headline number on the quarterly report.
  • License-revenue-sharing model with PIs restructured to reflect the LogNormal nature of the upside — flat percentages were replaced with tiered shares that pay more at the tail, aligning incentives with the actual distribution.

ModelRisk Functionality Used

  • Bernoulli-per-year publication generation conditioned on grant continuity, replacing the deterministic "expected publications" point estimate with a real binomial count.
  • Beta posteriors on grant renewal (Beta(20, 8)) and translation probabilities per theme — the only defensible family for bounded rate parameters, and one that produces a tight P10 the deterministic model cannot show.
  • LogNormal license values per theme with theme-specific σ_log between 0.80 (cardiac devices) and 1.20 (rare-disease genomics), reflecting the empirical fact that successful translations cluster around modest values with a long tail of major wins.
  • Per-theme cash-flow aggregation to portfolio NPV with a 10% discount rate over the 5-year horizon and a 7-year license-to-PV adjustment, producing the roughly +$2M mean / −$112M median split that drove the trustee restructure.
  • Tornado on portfolio NPV identifying the grant office as the highest-leverage lever, ahead of any portfolio rebalance — a counter-intuitive finding the deterministic model could not have surfaced.
  • Scenario CDF comparison of three rebalance options under common random numbers, isolating reallocation effect from sampling noise.

A research-portfolio NPV that reports its mean and stops is a portfolio reporting its tail dressed up as its centre. Monte Carlo simulation in ModelRisk is what lets the research office answer the question its trustees actually ask — not "what is the expected return?" but "what is the probability we are underwater, and what is the cheapest lever to move that number?"