COMPARE METHODS
Monte Carlo vs sensitivity: a modeled frequency or a tested response
Sensitivity answers what happens at selected input values. Monte Carlo estimates the frequency of outcomes under stated input distributions. Both reuse the financial model; only the second introduces a probability model for the uncertain inputs.
| Question | Sensitivity | Monte Carlo |
|---|---|---|
| What happens at price 9 and 11? | Directly evaluate those prices | The endpoints may never be sampled exactly |
| How often is EBITDA below 180? | A grid alone cannot answer this | Count qualifying valid outcomes under the supplied distributions |
| What must I specify? | Input values or ranges | Input distributions, settings and interpretation threshold |
| What is not established? | Likelihood of the tested inputs | Real-world likelihood without defensible probability assumptions |
On this page
Use the same economics to isolate the difference
In the synthetic model, volume is 100, variable cost is 5 EUR per unit and fixed cost is 300 EUR. The price is the only changing assumption. EBITDA = 100 × Price − 800. The native financial calculation is identical for both methods.
Sensitivity evaluates the selected endpoints. Monte Carlo draws prices uniformly between those endpoints and reports a frequency. The following rows answer different questions; the 200 EUR response span cannot be converted into an event probability.
| Method | Inputs tested | Result and meaning |
|---|---|---|
| Sensitivity | Price 9 EUR and 11 EUR | 100.00 EUR and 300.00 EUR EBITDA; selected endpoint responses |
| Monte Carlo | Uniform 9–11 EUR; 5,000 runs; seed 42 | 1983/5000 valid outcomes below 180 EUR EBITDA (39.66%) |
Separate an analytical probability from a sampled estimate
Under the explicitly supplied uniform assumption, EBITDA < 180 is equivalent to Price < 9.8. The qualifying price interval has width 0.8 inside a total width of 2, giving 0.8 / 2 = 40%.
The actual seeded simulation yields 39.66%. Sampling variation explains why it differs from 40%. Neither figure establishes that the business has that measured real-world risk: change the assumed price distribution and the modeled frequency changes.
Where Tornado and driver rankings fit
A Tornado compares response spans across chosen input ranges. Its longest bar identifies the largest modeled range effect under those ranges, not the most probable event. An unrealistically wide range can dominate that comparison.
Stochastic driver rankings use statistical association within the run. Association is not causal attribution. The underlying financial relationships and input assumptions still require review. For a model with interactions, examining deterministic responses can help explain a distribution rather than compete with it.
Choose a first method and know what remains unresolved
Start with sensitivity when the assumptions are not ready for defensible probability distributions or you need a transparent boundary test. Choose simulation when the decision requires a conditional threshold frequency and you can justify the distributions. An uncertain model does not become well specified merely because a distribution can be selected.
BalanceCheat currently samples configured inputs independently while retaining modeled links. It does not fit a correlation matrix. Invalid outputs are excluded from statistics and must be investigated; cancellation produces an incomplete run.
Continue with the relevant workflow
Use the native Monte Carlo workflow to reproduce the frequency. Read Sensitivity for deterministic controls and Tornado for range comparisons. For coherent management cases, use sensitivity vs scenario analysis.
The Monte Carlo concept guide explains the method; the Analyze overview provides the wider method selection.
Synthetic examples. Model-based results were reconciled with the current BalanceCheat calculations; assumptions, units and boundaries are specified in each example.