Help - Sensitivity Analysis (sensitivity)

Click here to open the calculator: Sensitivity Analysis

Sensitivity Analysis

Purpose

sensitivity() estimates which uncertain inputs drive your output most.

Monte Carlo alone answers "how wide is the output range?".
Sensitivity analysis answers "which assumptions are causing that range?"

In short:

  • Monte Carlo = output uncertainty distribution
  • Sensitivity = contribution/importance of each input to that uncertainty

When to use

Use this calculator when:

  • you have multiple uncertain assumptions,
  • you need driver ranking for decisions,
  • you need direction (+/-) and strength of each driver.

Typical use cases:

  • pricing and demand planning,
  • project economics (CAPEX/OPEX uncertainty),
  • operational risk analysis.

Background knowledge (quick)

What this calculator estimates

This calculator builds many sampled scenarios from your input distributions, evaluates your model, and then estimates input importance using:

  • SRC (Standardized Regression Coefficient): linear effect size after standardizing variables.
  • PRCC (Partial Rank Correlation Coefficient): monotonic relationship strength, controlling for other variables, based on ranked data.

Why two methods?

  • SRC is easy to interpret and strong when relationships are roughly linear.
  • PRCC is often more robust when relationships are monotonic but not strictly linear.
  • If both agree on top drivers and sign, confidence is higher.

What sensitivity scores are not

  • They are not percentages of total variance.
  • They do not prove causality beyond your model structure.
  • They depend on your chosen input distributions/ranges.

If you widen or narrow a variable’s uncertainty range, its sensitivity score can change substantially.


Inputs

Expression (xpr)

Your business/model equation. Default model:

(price - unit_cost) * demand - fixed_cost

Vary By (variation_target)

  • Variables (v): vary expression variables directly.
  • Parameters (p): vary named parameters inside function/calculator calls.

Inputs table (inputs)

Use one row per uncertain variable with these columns:

Variable | Distribution | Param 1 | Param 2 | Param 3

Supported distributions:

  • normal: Param 1 = mean, Param 2 = stdev
  • uniform: Param 1 = low, Param 2 = high
  • triangular: Param 1 = low, Param 2 = high, Param 3 = mode
  • lognormal: Param 1 = mu, Param 2 = sigma

Default example input:

price      | normal     | 120   | 8    |
unit_cost  | normal     | 70    | 6    |
demand     | triangular | 700   | 1300 | 1000
fixed_cost | normal     | 25000 | 3000 |

Analysis options

  • trials: number of samples (start with 500+, often 2000+ for stable ranking)
  • method:
  • src (Standardized Regression Coefficients)
  • prcc (Partial Rank Correlation Coefficients)
  • both
  • target_column: required when multiple numeric outputs are returned
  • correlated_variables + correlation: optional correlated pair
  • top_n: keep top-ranked drivers only
  • random_seed: reproducible runs
  • result_cells: pick numeric cells if expression returns tables

Outputs

Summary

Shows:

  • method used,
  • target metric analyzed,
  • trials requested / used / failed,
  • SRC R² (when SRC is computed).

About SRC R²:

  • closer to 1.0 => linear approximation explains most variation in target
  • low value => target behavior may be nonlinear/interaction-heavy; rely more on PRCC and model-specific checks

Table (ranked)

Main output columns include:

  • Variable
  • SRC, SRC Abs, SRC Rank (if SRC used)
  • PRCC, PRCC Abs, PRCC Rank (if PRCC used)
  • Primary Score, Primary Abs, Direction, Rank

How ranking is chosen:

  • method='src': ranking is based on SRC absolute value
  • method='prcc': ranking is based on PRCC absolute value
  • method='both': primary ranking favors PRCC-style monotonic signal while still showing SRC for cross-checking

Chart

A tornado-style directional impact chart for quick driver comparison.


Interpretation guide

Core interpretation

  • Larger absolute score => stronger driver.
  • Positive score => increasing input tends to increase output.
  • Negative score => increasing input tends to decrease output.
  • Agreement between SRC and PRCC strengthens confidence.

Practical reading pattern

  1. Check trial quality - Review Trials used vs Trials failed. - High failure share means rankings may be unstable or biased by invalid regions.

  2. Check target definition - Ensure Target is the exact metric you intend to optimize/control. - In multi-output models, wrong target_column means wrong conclusions.

  3. Read top drivers first - Focus on top 3-5 ranks for action prioritization. - Use Direction to decide whether mitigation should increase, decrease, hedge, or stabilize a driver.

  4. Cross-check methods - If SRC and PRCC disagree strongly, relationship may be nonlinear or non-monotonic. - Increase trials and inspect model structure before acting.

  5. Treat near-zero scores carefully - Very small absolute values usually indicate weak influence within current uncertainty ranges. - They are not proof of irrelevance outside those ranges.

Decision-oriented examples

  • Cost variable top-ranked negative: prioritize procurement hedging, supplier contracts, or process efficiency.
  • Demand variable top-ranked positive: prioritize forecast accuracy, demand sensing, and commercial planning.
  • Fixed cost low-ranked: avoid over-investing effort on fixed-cost precision if other drivers dominate.

Good practice

  1. Start with method='both'.
  2. Explicitly set target_column for multi-output expressions.
  3. Use realistic distributions/units.
  4. Re-run with more trials to check rank stability.
  5. Re-test after major assumption changes (ranges/distributions) before comparing with earlier rankings.