Skip to content

Quantification Methods ​

What this is ​

This page gives the details of each risk quantification method in Modulos. Each method gives one monetary expected loss. Modulos uses this value for the totals, the limits, and the priorities.

For the model and the workflow, see Risk Quantification.

Where in Modulos ​

Select the method in the quantification wizard:

  • AI System → Risks → [Risk] → Threat Vectors → Quantify → Select Method

After you select a method, the Configure Parameters step shows the fields of that method. The result shows on the Current Analysis tab of the Threat Quantification Detail page.

Who can do what ​

Only users who can quantify threats can open the wizard. See Risk Quantification.

At a glance ​

Each method, what it does, and when to use it:

MethodWhat it doesWhen to use it
Risk AgentAn AI agent makes a Fermi estimate from the data of the AI System, and records its assumptions. You enter nothing.A first estimate that a person then examines
Manual EntryRecords a monetary value and an optional descriptionFast: you have a number that you can defend
Scenario AnalysisCalculates the expected loss from three scenarios, each with a probability and an impactRecommended default: you start without incident data and want clear assumptions
Monte Carlo SimulationSimulates many years of losses from a loss range and a frequency, and gives the expected loss and the tailAdvanced: you want a probabilistic model and can defend its inputs
Risk MatrixChanges a likelihood score and an impact score into money with a scaling factor and a frequency multiplierOnly when you have a calibrated mapping from scores to money

Manual entry ​

Use Manual Entry when you have a monetary estimate that you can defend, and you want to record it quickly.

Inputs

  • value: the monetary expected loss in the currency of your organization. The API accepts 0, or a number from 0.01 to 10,000,000,000,000.
  • description: context, assumptions, or the source of the estimate, as a string or null. The request must contain this key.

UI fields

  • Value: necessary, from 0.01 to 10,000,000,000,000.
  • Description: optional.

API payload

json
{
  "method": "manual_entry",
  "value": 2400000,
  "description": "Expected annual loss based on incident history and remediation cost."
}

Result: the Manual Entry Details panel shows the Entered Value and the Description.

Notes

  • Use the number as a model output, not as a rating. Record the assumptions in the description. Then other persons can examine them.
  • Use the same time basis as your risk appetite. In most organizations, this basis is one year.

Scenario analysis ​

Scenario Analysis is the recommended default method when you want a clear structure without a complex model. It makes you write clear assumptions. It calculates a monetary expected loss from a set of scenarios.

Formula

text
Expected loss = Σ (probability × impact × (1 - mitigation_effectiveness) × weight)

This method is deterministic: the same parameters always give the same output.

Inputs

  • scenarios: a set of scenarios.
    • name: a necessary identifier for reports.
    • probability: the likelihood of the scenario, in [0, 1].
    • impact: the monetary loss when the scenario occurs, >= 0.
    • mitigation_effectiveness: the fraction of the impact that mitigation removes, in [0, 1].
    • weight: a positive multiplier for the scenario, > 0.
  • normalize_weights: when true, Modulos scales the weights to a sum of 1.
  • strict_probabilities: when true, the method makes sure that the scenario probabilities have a sum of 1, in the tolerance.
  • probability_tolerance: the permitted difference from 1 when the strict check is on. The default is 0.05.

UI fields

In the wizard, Scenario Analysis has three rows: Best Case, Worst Case, and Base Case. Each row has these columns:

  • The Probability column: a number from 0 to 1, with a slider in steps of 0.1.
  • The Impact column: the monetary loss of the scenario. The minimum is 1.
  • The Contribution column: the Probability multiplied by the Impact, and its percentage of the total. Modulos calculates this column.

The wizard does not send mitigation_effectiveness or weight. The method then uses 0 and 1.

Result: the Scenario Analysis Results panel shows each scenario with its contribution, the Weighted Expected Value, and the Range. When a scenario has a mitigation, the panel also shows the Residual Risk, the Mitigation, and the value Before Mitigation. A Key Assumptions section follows.

API payload

json
{
  "method": "scenario_analysis",
  "scenarios": [
    { "name": "best_case", "probability": 0.2, "impact": 200000 },
    { "name": "base_case", "probability": 0.6, "impact": 1200000 },
    { "name": "worst_case", "probability": 0.2, "impact": 7000000 }
  ]
}

Probability hygiene

By default, Scenario Analysis runs also when the sum of the probabilities is not 1. If the difference from 1 is more than the tolerance, the method records a warning in the output. Turn on strict probabilities only when you want Modulos to refuse inputs that do not agree.

When to use

  • You start without incident data and want a clear model that other persons can examine.
  • You want leaders to see the assumptions: what can occur, how likely it is, and how much it costs.
  • You want a stable baseline that you can make better when you get more Evidence.

Frequent errors

  • Probabilities that are only feelings, and not clear claims that other persons can examine.
  • Best and worst cases that are only "optimistic" and "pessimistic", without impacts from actual cost drivers.

Monte Carlo frequency and severity ​

Monte Carlo Simulation simulates many years of losses from your estimates of loss and frequency. It gives the expected annual loss and the tail metrics: quantiles, VaR, and CVaR.

Modulos has two models:

  • probability_impact: one incident can occur in a period, with the probability p and an uncertain severity.
  • frequency_severity: many incidents can occur in a period, with an uncertain frequency and an uncertain severity.

Model: probability × impact ​

Use this model when the threat is one type of incident for each period. The primary uncertainty is if the incident occurs and how much it costs.

Inputs

  • model: probability_impact
  • p: the probability of the incident in the period, in [0, 1]
  • severity: the loss of one incident, as a family with its params
  • trials: the number of simulation trials
  • rng_seed: an optional seed to get the same result again

API payload

json
{
  "method": "montecarlo_frequency_severity",
  "model": "probability_impact",
  "p": 0.15,
  "severity": {
    "family": "lognormal",
    "params": { "family": "lognormal", "mu_log": 12.0, "sigma_log": 0.9 }
  },
  "trials": 10000
}

Model: frequency × severity ​

Use this model when the threat can occur more than one time in each period. Examples are policy violations or safety incidents that occur again.

Inputs

  • model: frequency_severity
  • frequency: the number of incidents in one period, as a family with its params
  • severity, trials, and rng_seed: as in the probability_impact model

API payload

json
{
  "method": "montecarlo_frequency_severity",
  "model": "frequency_severity",
  "frequency": {
    "family": "poisson",
    "params": { "family": "poisson", "lambda": 2.0 }
  },
  "severity": {
    "family": "lognormal",
    "params": { "family": "lognormal", "mu_log": 11.7, "sigma_log": 1.1 }
  },
  "trials": 50000,
  "rng_seed": 42
}

Distributions available in the platform ​

Each severity and frequency object in the API has this structure:

json
{
  "family": "lognormal",
  "params": {
    "family": "lognormal",
    "mu_log": 12.0,
    "sigma_log": 0.9
  }
}

Supported families:

  • deterministic: value
  • bernoulli: p
  • poisson: lambda
  • lognormal: mu_log, sigma_log

The wizard uses a lognormal severity, and a Bernoulli or a Poisson frequency. For the other families, use the API.

UI fields ​

The Configure Parameters step of Monte Carlo Simulation has three panels and one section of settings:

  • Loss magnitude: Credible low and Credible high, the loss of one incident. Modulos fits a lognormal severity distribution with these two values at the 5th and the 95th percentile. The two values must be more than 0, and the low value must be less than the high value.
  • Severity preview: the loss curve of one incident from your two values, with Implied median, Implied mean, Severity p95, and Severity p99. A message shows when the tail is heavy (sigma_log more than 1.5).
  • Frequency: a switch with Once at most and Multiple times.
    • For Once at most, type the Annual probability, from 0% to 100%. Modulos uses the probability_impact model with a Bernoulli frequency.
    • For Multiple times, type the Expected events per year, more than 0. For example, 0.5 means one incident in two years. Modulos uses the frequency_severity model with a Poisson frequency.
  • Advanced simulation settings, closed by default: Number of trials from 1 to 100,000, with 10,000 as the default, and Random Seed. Keep Random Seed empty to get a random seed.

The wizard always requests the metrics EL, STD, QUANTILES, VAR, and CVAR. The other parameters below are available only through the API. A banner shows when you open the wizard again for a run with a different severity family or frequency family. Then type the values again.

Configure Parameters step for Monte Carlo Simulation with the Loss magnitude, Severity preview, and Frequency panels and the open Advanced simulation settings.

Monte Carlo Simulation asks for a loss range, a frequency, and simulation settings.

  1. Loss magnitude The loss of one incident, as a range from a credible low to a credible high value.
  2. Frequency Once at most asks for an annual probability. Multiple times asks for the expected events per year.
  3. Number of trials From 1 to 100,000 simulated years.

Result view ​

On the Current Analysis tab, a Monte Carlo run has four sections:

  1. The Simulation Results section: the Expected annual loss, the frequency, the standard error, and the run settings. A message shows when the expected annual loss is more than the calculated risk limit of the AI System. For this message, Modulos does not use a manual override of the limit.
  2. The Loss distribution section: a Risk threshold field and the probability that the annual loss is more than that threshold. A chart has the Distribution and Exceedance views. Three tiles show VaR and CVaR at the confidence level of the run, and Loss at p99. With the default confidence level of 95%, the tiles show VaR at 95% and CVaR at 95%. The default threshold is the calculated risk limit of the AI System. Without a limit, it is the 95th percentile of the annual loss.
  3. The Model assumptions section: what the run assumes about the incidents and the losses.
  4. The Advanced details section: the percentiles of the annual loss, the diagnostics of the run, and the raw parameters.
Monte Carlo result with the Expected annual loss card, the Risk threshold field, the loss histogram, and the VaR, CVaR, and Loss at p99 tiles.

The result of a Monte Carlo Simulation shows the expected annual loss, the loss distribution, and the tail values.

  1. Expected annual loss The mean annual loss of all simulated years.
  2. Risk threshold Type a loss to see the probability that the annual loss is more than it.
  3. VaR at 95% The annual loss that only 5% of the simulated years are above.
Advanced Monte Carlo parameters

You can send these parameters in the montecarlo_frequency_severity request:

  • metrics: the metrics to calculate: EL, STD, QUANTILES, VAR, CVAR
  • alpha: the confidence level for VaR and CVaR, in (0, 1), default 0.95
  • collect_samples: include the samples in the output, default false
  • max_samples_output: the maximum number of samples in the output, default 0
  • antithetic: use antithetic sampling, default false
  • stratified: use stratified sampling, default false
  • histogram_bins: the number of histogram bins, from 16 to 65,536, default 2048
  • allow_negative_severity: let the severity give negative values, default false
  • allow_infinite_moments: let the simulation use very large draws above the standard caps, default false
  • chunk_size: the number of trials in each simulation chunk, optional
  • sensitivity: the sensitivity configuration. This parameter is for experts and can change.

Risk Matrix ​

The Risk Matrix method changes ordinal inputs into money with a multiplication:

text
risk_score = likelihood_score × impact_score
quantified_value = risk_score × scaling_factor × frequency_multiplier

Use only with a calibrated mapping

Use Risk Matrix only when you have a calibration from your ordinal scores to monetary loss. A 5×5 matrix hides assumptions and causes category thinking. Use the output as a first estimate, and replace it with a better method when you can.

Inputs

  • likelihood_score: a number from 1 to 5.
  • impact_score: a number from 1 to 5.
  • scaling_factor: a multiplier that changes the score into money.
  • frequency_multiplier: a multiplier for the number of times that the scenario occurs in your time basis.

UI fields

  • A heat map. Select a cell to set Likelihood and Impact.
  • Likelihood and Impact, from 1 to 5.
  • Scaling Factor and Frequency Multiplier, more than 0.

The Risk Calculation Formula box shows the formula, the Risk Level, and the Score.

Result: the result shows the Likelihood, the Impact, and the Risk Score. The Analysis Breakdown section shows the inputs, the Risk Level, the Score, and the Financial Risk.

API payload

json
{
  "method": "risk_matrix",
  "likelihood_score": 3,
  "impact_score": 4,
  "scaling_factor": 100000,
  "frequency_multiplier": 1
}