Why do we model operational loss severity with a lognormal distribution?

Question

AcadiFi · Accepted Answer

The lognormal distribution is preferred for loss severity modeling because it captures several empirical features of operational losses that the normal distribution cannot.

**Why Lognormal Over Normal?**

1. **Always positive:** Loss amounts can't be negative. The normal distribution allows negative values, which is nonsensical for dollar losses. The lognormal is defined only for positive values.

2. **Right-skewed:** Most operational losses are small (a few thousand dollars), but occasionally there are massive outliers (millions or billions). The lognormal has a long right tail that captures this.

3. **Multiplicative nature:** Losses often result from multiplicative processes (a fraud grows exponentially as it goes undetected). If the log of losses is normally distributed, the losses themselves are lognormal.

**Parameter Estimation**

If loss amounts L follow a lognormal distribution, then ln(L) follows a normal distribution with mean mu and standard deviation sigma.

Suppose Thornfield Trust collected 200 operational loss observations. You take the natural log of each loss amount and compute:
- Mean of ln(losses) = mu = 10.3 (this corresponds to a median loss of e^10.3 = $29,674)
- Std dev of ln(losses) = sigma = 1.8

The lognormal parameters are:
- **Mean loss** = e^(mu + sigma^2/2) = e^(10.3 + 1.62) = e^11.92 = **$149,896**
- **Variance** = [e^(sigma^2) − 1] x e^(2*mu + sigma^2) = a very large number reflecting the heavy tail

Notice the mean ($149,896) is much higher than the median ($29,674) — this is the right-skew effect.

**Limitations:**
The lognormal may underestimate the probability of extreme losses. For fat-tailed operational losses, practitioners often use:
- **Generalized Pareto Distribution (GPD)** for losses above a high threshold
- **Weibull** or **Burr** distributions as alternatives
- **Extreme Value Theory (EVT)** approaches that focus specifically on the tail

**Exam Tip:** The FRM exam may give you mu and sigma and ask you to compute the mean, median, or a specific quantile of the severity distribution. Remember: median = e^mu, mean = e^(mu + sigma^2/2), and the p-th quantile = e^(mu + sigma x z_p).

Explore our FRM quantitative analysis section for more on loss distribution modeling.

Why do we model operational loss severity with a lognormal distribution?

Master Part I with our FRM Course

Related Questions

Practice Questions