4.10.5 Operational Risk Failure Modes

Basic Prerequisites: 04 · Aggregate Loss & LDA and EVT & Fat Tails.

---

1. Intuition & Practical Objective

The objective: face the reason operational risk is the hardest risk to quantify - the number you need is the one with the least data. Market risk has thousands of daily returns; credit risk has many borrowers; operational risk has, per bank, a handful of truly large losses per decade. LDA is mathematically sound; the failure modes are all statistical and institutional, and every one traces to a first principle.

This page names four failure modes and - because the discipline rewards demonstration - quantifies the two statistical ones (tail misestimation, scarcity) so you can feel their size:

1. Data scarcity & tail misestimation. The 99.9% loss is rarer than the data. Even a synthetic 5,000-year sample cannot pin it down: the VaR estimator’s standard error is ~26% of its value for a heavy tail, vs ~3% for lognormal.
2. Tail dependence of rare events. Poisson independence is the worst possible assumption exactly where it matters - in stress, losses cluster and co-move, so the true aggregate is fatter than the independent compound-Poisson convolution.
3. Frequency–severity conflation. Pooling event types breaks iid severity (02 · Loss Event Types) and corrupts the tail.
4. The measurement–incentive loop. Capital feeds on reported losses, so under-reporting lowers capital - the data pool is endogenously corrupted by the very model that uses it.

---

2. Mathematical Ground Truth & Derivations

2.1 Why the tail is statistically invisible

The estimator of ${\text{VaR}}_{0.999}$ from $T$ years of data uses the $\lceil 0.999T\rceil$-th order statistic. Its variance scales like

$$\text{Var}\left({\widehat{\text{VaR}}}_{0.999}\right)\propto \frac{\left(f_S^{-1}\text{-slope}\right)}{T}\approx \frac{(q_{0.999}{)}^2}{{\xi }^{2}T}\text{(heavy tail)},$$

because near the $99.9\%$ quantile the density is $f_S(q_{0.999})\sim \xi /q_{0.999}$ for a power tail $1-F_S(s)\sim s^{-\xi }$. Compared to a light tail, the relative standard error of the VaR estimator is inflated by roughly a factor $1/\xi +1$ relative to a light tail - about $1.7\times$ at $\xi =1.5$ in the relative-error sense (the simulated ratio is much larger because the tail density, not just the index, enters). That is the first principle: the fatter the tail, the less precisely any sample pins the quantile, and the more of your data lives where the answer lives least.

2.2 Why Poisson independence breaks in stress

The compound-Poisson variance $\text{Var}(S)=\lambda \mathbb{E}[X^2]$ assumes iid arrivals. Under clustering the true count has $\text{Var}(N)>\lambda$ (e.g. negative binomial), giving

$$\text{Var}(S)=\mathbb{E}[N]\mathbb{E}[X^2]+\text{Var}(N)(\mathbb{E}[X]{)}^2>\lambda \mathbb{E}[X^2].$$

The extra term - pure dependence contribution - is absent from the LDA formula and inflates the true tail exactly when capital is needed most. The failure mode is structural: the clean convolution of 04 · Aggregate Loss is built on an independence premise the real world violates in stress.

2.3 Tail-index sensitivity

The Hill estimator (03 · Frequency–Severity) turns a $\pm 0.1$ error in $\xi$ into a large capital swing, since ${\text{VaR}}_{0.999}$ of a Pareto-like tail scales as $\propto q_0^{1/\xi }$-type powers. The tail index is simultaneously the most influential and the least precisely estimated parameter in the whole model.

---

3. Computational Implementation - quantify the instability

A. VaR-estimator standard error, light vs heavy tail. Repeatedly estimate ${\text{VaR}}_{99.9}$ from 5,000-year samples; report the spread across 200 repetitions. Stdlib only.

The coefficient of variation jumps from 3.2% (lognormal) to 26% (Pareto). Even with 5,000 years of perfect simulated data, the heavy-tail VaR estimate is uncertain by a quarter of its value - the data-scarcity failure mode, quantified.

B. Hill tail-index instability under scarcity. Same true Pareto, growing sample sizes:

Even with 5,000 observations the Hill estimate is 0.550 against a true 0.5 (and the $k$ choice shifts it further). A bank with dozens of tail losses cannot be confident of its own tail index - which is why Basel mandates scenario analysis and external data to supplement the internal record.

---

4. Failure Modes & First-Principles Breakdowns

1. Data scarcity / tail misestimation (dominant). The parameter with the most leverage on capital is the least estimable. Mitigate with: external loss databases, scenario analysis, Bayesian pooling (Shevchenko 2011), and capital buffers around the estimator rather than point estimates.
2. Tail dependence of rare events. Independence (Poisson) is precisely wrong in stress: clustered arrivals add a $\text{Var}(N)(\mathbb{E}[X]{)}^2$ term LDA omits. Mitigate with negative-binomial/contagion frequency models and stress overlay.
3. Severity-family & $k$-selection risk. A lognormal-vs-Pareto choice changes 99.9% capital by ~5.8×; a Hill $k$ choice shifts the index. Any submission should report the capital range over defensible families and $k$ - never one number.
4. Incentive-corrupted data. Since capital falls when reported losses fall, under-reporting is economically attractive. Independent, audited loss capture (Basel SMA data standards) and de minimis thresholds are the guard - the model is only as honest as the data pipeline feeding it.

---

5. References

• Shevchenko, Modelling Operational Risk Using Bayesian Inference (2011)
• Embrechts, Klüppelberg & Mikosch, Modelling Extremal Events (1997)
• Glasserman, Monte Carlo Methods (2004)
• BCBS, Basel II (2006), ¶669(e)–(f)
• BCBS, Basel III d424 (2017)

---

6. Connected Graph Bridges

• Base: 04 · Aggregate Loss & LDA · EVT & Fat Tails
• Back: 03 · Frequency–Severity
• Forward: 06 · Advanced Extensions · Index Hub
• Sibling: Model Risk & Validation · Bayesian Statistics (the scarcity remedy)