A Point Estimate Isn't A Forecast: Building A Confidence Band For DCF Without Monte Carlo

Build defensible valuation ranges by applying PERT estimation to traditional DCF models.

sasun-bughdaryan-X4VwfPzy_2o-unsplash (1).jpg
Unsplash

Methodology deep-dive, real numbers from real duels. Not investment advice.

Two pieces ago, in our list of honest limitations, we flagged this one: our DCF model outputs a single number — $176.47, not "$165–$190" — with no sense of how much to trust that precision. Damodaran has made this point for decades: terminal value alone typically makes up 60–80% of enterprise value, built on one unverifiable assumption about perpetual growth. A number precise to the cent, resting on a foundation that precise is not.

This piece is about closing that specific gap — turning a single fair-value number into a defensible range, using nothing the model doesn't already publish.

Why not Monte Carlo

A full simulation needs assumed probability distributions on every input — shapes we can't honestly justify with the data on hand. So instead we used something older and simpler: the three-point (PERT) estimate, built for the U.S. Navy's Polaris missile program in 1959 and still standard in project-cost estimation.

The logic: build a Bear case and a Bull case from the model's own already-published quality tiers, then combine:

Expected ≈ (Bear + 4×Base + Bull) / 6
SD ≈ (Bull − Bear) / 6

This construction treats the (Bear, Bull) spread as roughly ±3 standard deviations, giving an interpretable 1-sigma band without inventing a single new input. For Bear/Bull, we shift the model's own documented tiers one notch in each direction — the ROIC growth-multiplier, the FCF-margin multiplier, and the ROIC-based beta tier that drives WACC — symmetrically, using only thresholds the model already discloses.

The result: 16 companies, one method

Ticker

Fair Value T0

Point-Estimate 3Y Upside

PERT-Adjusted Upside

1σ Range

NVDA

$176.50

+14.6%

12.0% ± 2.6%

[9.4%, 14.6%]

AMD

$67.77

+24.0%

26.1% ± 3.9%

[22.2%, 30.0%]

GOOGL

$129.53

+18.2%

18.8% ± 5.2%

[13.6%, 24.0%]

META

$524.34

+16.2%

15.9% ± 4.8%

[11.1%, 20.7%]

AMZN

$255.02

+18.6%

18.6% ± 4.8%

[13.8%, 23.4%]

WMT

$77.87

+18.7%

19.1% ± 4.5%

[14.6%, 23.6%]

CRM

$416.84

13.1% ± 4.5%

[8.6%, 17.6%]

NOW

$113.89

18.4% ± 5.4%

[13.0%, 23.8%]

VRTX

$257.98

18.5% ± 4.1%

[14.4%, 22.6%]

REGN

$728.55

22.0% ± 4.4%

[17.6%, 26.4%]

GFF

$0.00

see note below

WOR

$17.68

44.1% ± 0.9%

[43.2%, 45.0%]

BHE

$26.76

25.4% ± 1.9%

[23.5%, 27.3%]

PLAB

$19.78

16.0% ± 1.7%

[14.3%, 17.7%]

KO

$18.70

20.4% ± 1.5%

[18.9%, 21.9%]

PEP

$117.24

21.3% ± 0.3%

[21.0%, 21.6%]

The width of the band is informative in itself: high-growth, high-multiple names (NVDA, GOOGL, NOW) carry wider bands than slow, steady cash generators (KO, PEP, PLAB) — exactly what a sensitivity method should show, since more forward-looking assumption means more room for disagreement about it.

The question this was actually built to answer

Our first piece in this series claimed something that got attention: NVIDIA won its relative fundamentals "duel" against AMD 92-8, but AMD's modeled DCF upside (+24.0%) was higher than NVIDIA's (+14.6%) — despite losing badly on operating metrics. The obvious follow-up: was that gap real, or just noise from a point estimate that never had error bars?

With PERT ranges applied, NVIDIA's band is [9.4%, 14.6%] and AMD's is [22.2%, 30.0%]. The two do not overlap. The original finding survives direct scrutiny under parameter uncertainty — the kind of check we think more valuation content should be willing to run on its own prior claims, rather than restating a point estimate with more confidence than it deserves.

A quiet side benefit of this research series

One purpose of running this series on duelstocks.com is that it doubles as an audit of our own model — and this piece already produced two small, concrete improvements to the live tool. Building the sensitivity bands required re-deriving the DCF formula by hand against real filings, which surfaced an edge case in how equity value is floored for highly leveraged companies (now handled more transparently instead of via a placeholder substitution) and an undocumented growth safety clamp that's now spelled out directly in the report's own methodology section. We reviewed the clamp's actual bounds against the tier system's own mathematical ceiling and left the numbers themselves unchanged — they turned out to already be well-calibrated, sitting almost exactly at the level the tier system can ever produce on its own. Neither change alters any figure in this article; both are disclosed here for completeness, not because they're the point of the piece.

What a "range" could look like in the product itself

Right now, getting this band means doing the tier-shift math by hand against a published DCF report — which is exactly what this piece demonstrates, step by step, using nothing beyond what's already in any DCF report you can pull today. A natural next step — not a commitment, just a direction worth naming — would be for the DCF report itself to eventually show this corridor directly: an upper and lower bound alongside each year's point projection on the chart, plus a small supplementary table with the same Bear/Base/Bull figures shown above. Whether or not that ships, the method here is fully reproducible by hand from the report you already have.

How much of the original critique does this actually resolve

Worth being precise about this, since overclaiming here would defeat the point of the exercise. The original limitation was that the model offers no way to gauge how much to trust a single DCF figure. This piece closes that gap at the methodology level: there is now a reproducible, no-new-data way to construct a defensible range around any DCF output, and we've shown it holds up under a real stress test (the NVDA/AMD comparison). It does not yet close the gap at the product level — the report itself still surfaces one number, and building the range still requires the reader to do the arithmetic described here. It also doesn't resolve a third, harder question we're not claiming to have answered: whether these bands, once built, actually contain realized future prices at anywhere close to the stated ~68% rate — that requires validating against outcomes over time, which this piece doesn't attempt. Partial, honest progress, on a specific and previously unaddressed limitation.

References

  1. Damodaran, A. (2012). Investment Valuation: Tools and Techniques for Determining the Value of Any Asset (3rd ed.). Wiley.

  2. Malcolm, D. G., Roseboom, J. H., Clark, C. E., & Fazar, W. (1959). "Application of a Technique for Research and Development Program Evaluation (PERT)." Operations Research, 7(5), 646–669.

  3. Howard, R. A. (1988). "Decision Analysis: Practice and Promise." Management Science, 34(6), 679–695.

  4. Sloan, R. G. (1996). "Do Stock Prices Fully Reflect Information in Accruals and Cash Flows about Future Earnings?" The Accounting Review, 71(3), 289–315.

  5. Project Management Institute — PMBOK Guide, Three-Point Estimating Technique.

Not investment advice. All data sourced from public SEC EDGAR filings. Disclosure: I hold a long-standing position in NVDA (~5 years); no position in any other stock mentioned.

STOCKS IN THIS ARTICLE

Comments