Sensitivity analysis: identifying the assumption that changes the decision.
Evidence status
What was reproduced, and what it does not prove.
The public JSON and the Python run verify the single parameters, the order of impacts and the switching rules of the four published scenarios. They verify neither the statistical intervals, nor the elasticities, nor the probability of the scenarios.
- Technical specification of the teaching sensitivity analysis, specification in French (opens in a new tab)
v1, 31 August 2026 · CC BY 4.0 · specification in French · Method document written by marketing-mix.net. The figures are illustrative: they are not market measurements.
Key distinctions
Three conditions before calculating.
Defensible ranges
Use estimation intervals, contracts, history or operational bounds. An identical arbitrary variation gives more importance to parameters according to their unit.
Linked parameters
Price, volume and margin; reach and frequency; budget and cost of purchase can move together. Varying them separately produces impossible scenarios.
Tipping criterion
Observe the change of decision, not only the variation of the result. A recommendation can lose 30% of gain while remaining clearly preferable.
Method
Moving from local sensitivity to robustness
- 01Name the central decision
Publish base scenario, result, constraints and alternative; without a stable reference, the variants become a collection of forecasts.
- 02Class the unknowns
Separate estimated parameters, controlled decisions, constraints and external events; their treatment and their range are not identical.
- 03Test one by one and then jointly
The tornado locates the sensitive parameters; the linked scenarios then check the plausible combinations.
- 04Compute the value of information
Compare the cost of a test with the avoidable loss if the central assumption is false.
The tornado of contribution gain
Each bar shows the deviation from the central gain of €122,000. Elasticity and margin matter more than the model's secondary parameters.
The pilot becomes preferable to a direct roll-out.
A positive recommendation but dependent on elasticity
The Noroa reallocation produces +€122,000 at the centre. The plausible variation of media elasticity moves the result from −€92,000 to +€71,000, that of margin from −€64,000 to +€58,000.
Under prudent elasticity alone, the gain remains +€30,000. Combining prudent elasticity and low margin, the result becomes −€34,000. A six-week test costs €18,000 and can halve the uncertainty on elasticity.
Do not deploy the target directly. Fund the pilot: its cost is lower than the potential loss of the adverse joint scenario.
Acceptance conditions
What must be true to act.
- 01
The ranges reflect real uncertainty or operational bounds.
- 02
The tornado is completed by plausible joint scenarios.
- 03
The threshold of decision change is explicit.
- 04
The value of additional information is compared with the cost of the test.
Limits
What this analysis does not prove.
- A local analysis does not always capture non-linearities or strong interactions.
- The scenarios do not cover unimagined events.
- Ranges that are too narrow turn sensitivity into confirmation of the central scenario.
References
Works cited.
- Saltelli, A. et al. (2008), Global Sensitivity Analysis.
- Keeney, R. L. & Raiffa, H. (1976), Decisions with Multiple Objectives.
- Ben-Tal, A., El Ghaoui, L. & Nemirovski, A. (2009), Robust Optimization.