Sensitivity Analysis: Stress-Test a Startup Plan to Failure

Sensitivity analysis measures how far a plan’s outcome moves when one assumption moves; a stress test asks how far the assumptions can move before the plan breaks. Both answers depend on the ranges tested. Set each range from evidence, define the break as the point where the next round fails, and measure the distance to it, alone and in combination.

A tornado chart that moves every input by ±10% answers neither question, because it ranks line items by their size. A test that stops at zero cash looks for the break in the wrong place. Banking regulators rewrote their stress-testing rules after 2007 and now start from the other end: the Basel Committee asks banks to find “what scenarios could challenge the viability of the bank” (Basel Committee, 2009). What follows applies that method to a startup plan, with a worked example.

Four tests, four different questions

Test	Question it answers	How it is run	What it misses
One-at-a-time sensitivity analysis (the tornado chart)	Which assumption moves the outcome most?	Move each input to a low and a high value with the others at plan, then rank the swings	Combinations, and any input whose range was set too narrow or too wide
Global sensitivity analysis	Which assumptions drive the uncertainty in the outcome?	Vary every input at once over its range, thousands of times, and split the outcome’s variance by input	Correlations between inputs, unless they are modelled; it needs a range for every input
Stress test	Does the plan survive a severe but plausible shock?	Apply a chosen shock or scenario and recompute	Every shock nobody chose to run
Reverse stress test	What would make the plan fail, and how likely is that?	Define failure first, then search for the inputs that produce it	Any failure outside its definition, so the definition must be written down first

Source: Saltelli et al. (2019); Basel Committee on Banking Supervision (2009, 2018); FCA Handbook, SYSC 20.2.

Sensitivity analysis with ±10% ranges ranks line items, not risks

The Basel Committee’s description of the basic test is exact. Sensitivity tests “generally shock individual parameters or inputs without relating those shocks to an underlying event or real-world outcomes”, and its own example moves equity prices by fixed steps “(such as by 10%, 20%, 30%)” (Basel Committee, 2009). Andrea Saltelli and seven co-authors reviewed sensitivity analysis as it is practised. They describe the display as “a tornado plot (a bar chart which shows the effect on the output of varying each assumption by a fixed amount in either direction)” (Saltelli et al., 2019). Each bar’s length depends on two numbers: how strongly the outcome responds to the input, and how far the input was moved. A fixed percentage makes the second number arbitrary.

A hypothetical SaaS company has $3.0M of ARR and $14.0M in the bank after its Series A. Its plan books $125k of new ARR a month and loses 12.5% of revenue a year to churn. It earns a 75% gross margin, collects in 45 days and spends $560k a month on operating expenses, rising 1.5% a month with hiring. The churn figure is the median in Lighter Capital’s 2025 benchmarks, drawn from the connected accounts of 155 private B2B SaaS companies in the lender’s portfolio (Lighter Capital, 2025). Over 24 months the plan reaches $4.94M of ARR, 28.4% a year against Lighter’s 28.29% median, and ends with $3.81M of cash (our arithmetic).

The same plan, two tornado charts: uniform ranges put churn last, evidence puts it first

Assumption	Plan	Swing in month-24 cash at ±10% (rank)	Evidence-based range, 25th to 75th percentile	Swing (rank)	Share of variance
Operating expenses vs plan	$560k a month, +1.5% a month	$3.21M (1)	98% to 102% of plan	$0.64M (2)	27%
Gross margin	75%	$1.22M (2)	73% to 77%	$0.32M (4)	6%
New ARR booked	$125k a month	$0.37M (3)	20% below to 5% above plan	$0.46M (3)	14%
Annual gross revenue churn	12.5%	$0.12M (4=)	5.48% to 24.53%	$0.94M (1)	54%
Days sales outstanding	45 days	$0.12M (4=)	40 to 60 days	$0.27M (5)	4%

Source: hypothetical company. Churn median and quartiles from Lighter Capital (2025), annual revenue churn, 155 companies, 2020 to 2024 pooled; the other ranges are hypothetical. Share of variance: first-order shares of month-24 cash from 100,000 Monte Carlo draws, estimated by binning. All figures our arithmetic.

Under ±10%, operating expenses top the chart because they are the largest number in the model, $16.0M over two years, and gross margin comes second for the same reason. Churn ties for last. Evidence reverses the order. Lighter’s interquartile range for revenue churn runs from 5.48% to 24.53%, 7.6 times as wide as the ±10% band around 12.5% (our arithmetic). A company with little churn history of its own could sit anywhere in its peers’ middle half. Operating expenses get a fifth of the ±10% band: the company decides them every month, so their uncertainty is execution error rather than the size of the budget.

The bookings range is skewed downward on purpose. Benchmarkit’s 2026 survey of B2B SaaS companies found that planned growth had exceeded the previous year’s actual growth “For 4 years in a row”, with a 26% median plan for 2026 against 20% achieved in 2025 (Benchmarkit, 2026). A board shown the first chart spends its meeting on expense discipline and gross margin, two lines that the second chart ranks second and fourth, while the line that can sink the plan gets a $0.12M bar.

But banks that chose their own stress tests chose mild ones

Honest ranges fix the ranking. They do not fix severity. In its 2009 review of bank stress testing, the Basel Committee found that scenarios “tended to reflect mild shocks, assume shorter durations and underestimate the correlations” between risks. Before the crisis, it added, “severe” scenarios typically produced losses of “no more than a quarter’s worth of earnings (and typically much less)”. Risk managers who proposed harsher tests met resistance: “Scenarios that were considered extreme or innovative were often regarded as implausible by the board and senior management” (Basel Committee, 2009).

Northern Rock made the point in Britain. Sir Derek Wanless told the House of Commons Treasury Committee that the bank’s “stress tests at the time were sufficient”, and that it had been working through 20 scenarios signed off by its board. “Fifteen of those scenarios involved liquidity risk, including two where securitisation became a particular problem. What did not happen was that we stress-tested the scenario of what has actually happened”. Adam Applegarth, the chief executive, described the extra tests the regulator had requested as mainly about credit, such as a 40% fall in house prices. They were not focused on the event “deemed implausible, which was the rapid and long-lasting closure of global markets”. The committee concluded that the board’s stress testing of its business model “proved to have been thoroughly inadequate” (Treasury Committee, 2008).

Twenty scenarios, chosen by the people being tested, missed the one that happened. A startup is not a bank, but a founder’s downside case, written by the founder for a board meeting, has the same kind of author.

Therefore regulators now run the test backwards, from failure

The fix was to stop letting the tested choose. Basel’s Principle 9 asks every bank to determine “what scenarios could challenge the viability of the bank (reverse stress tests)”. Such tests “start from a known stress test outcome (such as breaching regulatory capital ratios, illiquidity or insolvency) and then asking what events could lead to such an outcome”. The committee noted that before the crisis most senior managers saw the exercise as “of little value” because the event seemed remote. It listed the areas that gain most, including “new products and new markets which have not experienced severe strains” (Basel Committee, 2009). The phrase was written for a bank’s business lines, and it fits a startup more closely still, since the whole company is a new product in a market that has not yet tested it.

Britain wrote the method into its rulebook. From 14 December 2010, firms in scope had to “carry out stress tests and scenario analyses that test its business plan to failure” (FCA Handbook, SYSC 20.2). That chapter lapsed at the end of 2021. The PRA still requires reverse stress tests of banks (PRA, SS31/15), and the FCA expects them of investment firms with complex businesses (FCA Handbook, MIFIDPRU 7 Annex 1). The part a founder should copy is the definition of failure: “the point at which the firm’s existing shareholders are unwilling to provide new capital”. The guidance adds that “Such a point may be reached well before the firm’s financial resources are exhausted.”

For a venture-backed company, shareholders refusing new capital means the next round falling through, so the break has two tests. In the worked example the plan breaks if month-24 cash falls below $2.0M, about four months of the burn at that point. It also breaks if ARR growth falls below 15.25% a year, the 25th percentile in Lighter’s data, which takes month-24 ARR below $3.98M (our arithmetic). Both thresholds are assumptions a board should argue over; the method only needs them written down.

How far each assumption can move on its own before the plan breaks

Assumption	Plan	Breaks the plan at	Test that fails first	Distance to break, in standard deviations
Annual gross revenue churn	12.5%	25.2%	Growth	0.71
New ARR booked	$125k a month	$79.7k a month (36% below plan)	Growth	1.22
Operating expenses vs plan	100%	111.3%	Cash	3.80
Days sales outstanding	45 days	176 days	Cash	5.91
Gross margin	75%	52.7%	Cash	7.51
Closest combination: churn 22.2% with bookings 9% below plan			Growth	0.62

Source: the hypothetical company in the tornado table above. Each input’s standard deviation is the adverse half of its interquartile range divided by 0.674, the ratio for a normal distribution. All figures our arithmetic.

The last column measures the distance from an assumption’s planned value to the value that breaks the plan, in that assumption’s own standard deviations.

distance to break = (break value − plan value) ÷ standard deviation of the assumption

A tornado bar says how much the outcome moves over a range; the distance to break says how much room is left.

Two results stand out. The plan breaks on growth before it breaks on cash. With churn at 25.2%, the company still holds $3.18M at month 24, $1.18M above its floor, but its growth has fallen below what investors will fund. And the assumption closest to the break is churn, the one the ±10% chart placed last. Its break lies just beyond the peer 75th percentile of 24.53%, so a company that knows no more about its churn than its peers’ spread has about one chance in four of breaking the plan on churn alone.

But the most likely failure is a mix of moderate misses

A table of single-input break values is still a one-at-a-time sensitivity analysis, and Saltelli’s group is blunt about those. In ten dimensions, moving factors one at a time “leaves over 99.75% of the input space unexplored”, with the result that “uncertainty is systematically under-estimated and sensitivity is wrongly estimated”. In their review of 280 highly cited papers, 34% of sensitivity analyses moved one factor at a time and 41% used global methods; the rest were unclear or had none. The authors judge the misuse “at least as serious as the misuse of the p-test in statistical modelling” (Saltelli et al., 2019).

Structural engineering has an exact answer to the question a reverse stress test asks. The first-order reliability method is also known as the Hasofer-Lind reliability index, after a 1974 paper by Abraham Hasofer and Niels Lind (Hasofer and Lind, 1974; First-order reliability method, Wikipedia). It rescales each uncertain input into standard deviations, and the combinations that cause failure form a boundary. The most likely failure is the design point: “the failure point with the highest probability density, which often falls at the closest point to the origin of the limit state function” (Line sampling, Wikipedia).

The mapping to a plan holds term for term: the origin is the plan, the boundary is the break, and the closest point is the reverse stress test’s most likely failure. The method’s own assumptions mark its limits here: startup ranges are rarely normal, and in a downturn churn and bookings move together.

In the example, the closest failure point is churn at 22.2% with bookings 9% below plan, both inside their interquartile ranges and 0.62 standard deviations from the plan, against 0.71 for churn alone. A Monte Carlo run of 100,000 versions of the plan, with every input drawn independently from a distribution fitted to its range, breaks the plan in 31.8% of draws. Of those failures, 99.3% miss the growth test, and in 13.6% no single input is beyond its own break value. Read line by line, the break-value table above shows one exposed assumption; read as a whole, the plan has worse than one chance in four of failing.

Evidence changes the distance. Suppose two years of the company’s own data confirm the 12.5% median and put churn’s interquartile range at 10% to 16%. Churn would then sit 2.44 standard deviations from the break, and bookings would become the nearest single assumption, at 1.22. The share of failing draws would fall from 31.8% to 14.5%. The plan has not changed; what the company knows about it has.

Is a one-at-a-time test enough for a near-linear startup model?

The best case for the simple chart comes from its critics. Saltelli’s group concedes that “If all models were linear, an OAT or derivative based approach would be adequate”. Basel credits sensitivity tests with “a fast initial assessment of portfolio sensitivity to a given risk factor” (Saltelli et al., 2019; Basel Committee, 2009). A startup plan is mostly sums and products. In the worked example the first-order variance shares account for the whole variance within estimation error, so interactions are small. The evidence-based tornado also ranks the five assumptions in exactly the order of their variance shares.

Saltelli’s concession holds for the ranking. Once the ranges come from evidence, a one-at-a-time sensitivity analysis is a fair first screen, and in our judgement most companies would gain more from fixing their ranges than from computing variance indices. It does not hold for the distance to the break, because a break is a threshold and thresholds are crossed by combinations, as the gap between 0.62 and 0.71 standard deviations shows. The Monte Carlo also errs on the optimistic side, since it draws the inputs independently. Shocks that move several inputs together call for coherent scenarios, which Basel’s 2018 principles require to be internally consistent and backed by a narrative (Basel Committee, 2018). That is the work of scenario analysis rather than sensitivity analysis.

Write the break down before running the test

Before anyone opens the model

Each range needs a written source: the company’s own forecast errors where it has two years of history, a peer spread where it does not, and a downward skew on bookings. A board shown ±10% on every line is looking at line items ranked by size. The break comes next, defined before any test runs as a cash floor and the growth rate the next round needs, and that written definition belongs in the board pack. The 3-statement model template and scenario planner below is a starting point; put each range and its source on the input sheet.

When the results reach the directors

Each assumption’s distance to the break belongs in the pack in standard deviations, followed by a search for the closest combination; a spreadsheet Monte Carlo add-in or a short script will do, and the three assumptions nearest the break deserve a line each. The pack should then name the combination of misses that breaks the plan first, and what the company will do, and by when, if that combination starts to appear. That response needs a trigger agreed in advance, for example a trailing churn rate that, if crossed by month 9, moves the company to a lower-burn plan. Set it early: the UK rule asks for measures that allow for “the time that the firm would have to react to these events and implement those measures” (FCA Handbook, SYSC 20.2.5G).

A board pack should show the distance to failure

A sensitivity analysis earns its place in a board pack when it states three things: where each range came from, what counts as failure and how far the plan is from it. The ±10% tornado chart states none of them. Banks were made to learn this after 2007, when the scenarios they had chosen proved too gentle and the one they had rejected arrived. A startup can borrow the method without a regulator: ranges from evidence, failure defined as the next round falling through, and the distance measured alone and in combination. In the worked example the nearest failure sat 0.62 standard deviations from the plan, and the plan broke in 31.8% of 100,000 draws, a share that drawing the inputs independently probably understates.

Sources

• Basel Committee on Banking Supervision (2009). Principles for sound stress testing practices and supervision. Bank for International Settlements.
• Basel Committee on Banking Supervision (2018). Stress testing principles. Bank for International Settlements.
• Benchmarkit (2026). 2026 B2B SaaS and AI-Native Performance Benchmarks.
• Financial Conduct Authority. FCA Handbook, SYSC 20.2, Reverse stress testing requirements (in force 14 December 2010 to 31 December 2021), and MIFIDPRU 7 Annex 1.
• Hasofer, A. M. and Lind, N. C. (1974). Exact and invariant second-moment code format. Journal of the Engineering Mechanics Division 100(1), 111–121.
• House of Commons Treasury Committee (2008). The run on the Rock, Fifth Report of Session 2007–08, HC 56-I.
• Lighter Capital (2025). 2025 B2B SaaS Startup Benchmarks.
• Prudential Regulation Authority (2025). Supervisory Statement SS31/15, The Internal Capital Adequacy Assessment Process (ICAAP) and the Supervisory Review and Evaluation Process (SREP), chapter 4.
• Saltelli, A., Aleksankina, K., Becker, W., Fennell, P., Ferretti, F., Holst, N., Li, S. and Wu, Q. (2019). Why so many published sensitivity analyses are false: a systematic review of sensitivity analysis practices. Environmental Modelling & Software 114, 29–39.
• Wikipedia. First-order reliability method; Line sampling.

The company in the tornado and break-value tables is hypothetical; only its churn median, churn range and growth hurdle come from published data, Lighter Capital’s. Every model output given without a cited source, from the tornado swings and break values to the closest combination and the Monte Carlo of 100,000 draws (independent split-normal inputs fitted to each range), was produced by CX Cash from those inputs.