Wehrspohn Risk Management

How does risk assessment work in enterprise risk management?

Case study

By Dr. Uwe Wehrspohn.

Using a risk model in Excel, we show how to assess risks gross and net with suitable distributions, aggregate them with a Monte Carlo simulation and compare them with the risk-bearing capacity. Using a cyber risk as an example, we also show how to model a risk from its drivers and transfer the simulated result into the aggregation.

How do you assess the individual risks in enterprise risk management, that is, in company-wide risk management? And how do you combine them into an overall picture? These questions were the subject of a talk on risk quantification that we gave to a working group of the Risk Management & Rating Association.

We show in Excel how to assess risks with distributions and aggregate them with a Monte Carlo simulation. We also show how to assess an individual risk with a model built from its drivers.

What information does a risk assessment need?

In short: Risk Kit is an add-in for Microsoft Excel that lets you build and analyse Monte Carlo simulations directly in the spreadsheet. What we show here applies in general, though: assessing and aggregating risks with a Monte Carlo simulation works on the same principle in any simulation tool.

The first example is a risk aggregation as required by IDW PS 340. The template for it is included in Risk Kit. In it, each risk is assessed twice, before measures and after measures, that is, gross and net. Figure 1 shows the block before measures. Each row describes one risk.

Figure 1: The template for the risk aggregation with the block before measures. All ten example risks are assessed with the PERT distribution using the minimum, the most likely and the maximum loss (German interface).

For each risk, we first record the usual data: the risk ID, the name of the risk in plain text and the risk field, for example market risk, climate risk or cyber risk. The risk field is an intermediate aggregate between the individual risk and the company as a whole. This later shows whether the company takes large risks in areas that do not fit its business model and are therefore undesirable.

The assessment starts with the question of whether a risk can occur only once or several times per period. Most risks can occur several times, but some are so singular that they can occur only once. This property applies regardless of the measures taken. The template therefore records it only in the block before measures, in the column “Risiko kann mehrfach auftreten” (risk can occur several times).

For the occurrence of a risk there is a probability or a frequency of occurrence. For the impact on occurrence, we use a distribution with certain parameters. The distribution is named clearly so that it is obvious what exactly is being assessed. A simple way to start is to assess all risks with a three-point distribution. We chose the PERT distribution, which represents the minimum, the most likely and the maximum loss. For the recession, risk 1, these are €0, €1,750,000 and €5,000,000 with a probability of occurrence of 0.25. These parameters are easy to understand and to explain, and the assessment remains consistent.

Which distributions fit?

In football, we know that you can play the ball in many ways: dribble, pass, stop, volley, curl, head and so on. Even as fourteen-year-olds we wanted to react flexibly to the situation in a match and play the ball accordingly. A player who can only play with his right foot and only shoot curled balls would be a risk to his team. If all players do the same thing, the team may be entertaining, but it will not be successful. It is the same in risk management. The point is to choose the right and useful methods. We do not want to do everything, but what matters.

Risk Kit offers a manageable selection of the most important distributions for modelling in companies. There are many more, but most of us will hardly ever need distributions such as Bradford or Burr, unless we are actuaries at an insurance company. The selection consists of three groups:

  • For the occurrence of a risk there are distributions such as Bernoulli, binomial, multinomial and Poisson.
  • For the impacts we use classic distributions such as PERT, triangular, modified PERT, LogNormal or Weibull.
  • In addition, there are innovative distributions such as Expert, Flex and Poly.

If we use several distributions, we always state the name of the distribution and the matching parameters. Figure 2 shows this for the assessment after measures. The column “Verteilung” (distribution) here contains Normal, Pert, LogNormal3, Pert2P, Trapezoid, Weibull, Expert, Dreieck (triangular), Flex and ModPert. The columns “Param 1” to “Param 6” hold a different number of parameters depending on the distribution.

Figure 2: The block after measures of the same template. Each risk is assessed with its own distribution, whose name is in the column “Verteilung” (distribution) and whose parameters follow in the columns “Param 1” to “Param 6” (German interface).

Good case and bad case instead of worst case

We made sure to provide distributions that experts can estimate well. That is better than the worst case, which is often deceptive. Nobody wants to claim that every risk has a maximum and that this maximum is known exactly.

Proven methods can also be used in a new way. For the PERT distribution, for example, we make it possible to enter a good case and a bad case instead of a best case and a worst case. Experts find these easier to estimate. Here the good case is the 10% quantile and the bad case the 90% quantile, that is, a value that is exceeded in only 10% of cases. Risk Kit then extends these points to a PERT distribution.

Figure 3 shows this for the flooding of the production site, which is assessed after measures with Pert2P in Figure 2. The entries are the good case of €1,250,000 as the 10% quantile, the most likely value of €1,800,000 and the bad case of €2,750,000 as the 90% quantile. The two blue lines in the chart mark the two quantiles. The minimum and maximum follow from the extension: according to the statistics on the right, the smallest value is around €783,000 and the largest around €3.84 million. The expert does not have to state them.

Figure 3: The PERT distribution for the flooding after measures, defined by the good case (10% quantile), the most likely value and the bad case (90% quantile). The blue lines mark the two quantiles, and the statistics on the right show the minimum and maximum of the extended distribution (German interface).

The distributions shown can also be fitted to data and to simulation results. Poly and Flex in particular are so flexible that they even cover the other continuous distributions as special cases. In this way, a few distributions open up many possibilities, like a Swiss army knife.

How do you aggregate the risks and check the risk-bearing capacity?

For the aggregation, we add up the simulated losses of all risks and obtain the total loss before measures and the total loss after measures. In the example, Risk Kit runs 10,000 simulation runs for this. The results can be saved and archived. For review, they can be passed on to auditors or colleagues without handing over the model with all its formulas.

A results report shows the gross and net results side by side as a histogram and as a distribution function. This is the view from above on the company’s risk, so to speak. Figure 4 shows the distribution functions. They give the probability that the total loss does not exceed a given amount. The orange curve after measures rises much earlier than the blue one before measures. At €5,000,000 it is already close to 1, the blue one only at about 0.3. So after measures a total loss of at most €5,000,000 is almost certain, before measures only in about 30% of cases.

Figure 4: The distribution functions of the total loss before measures (blue) and after measures (orange) from the results report (German interface).

In the report, the sensitivities show in a drill-down how strongly the aggregated risk fluctuates when an individual risk occurs. Risk maps are often criticised. In Risk Kit they are shown as a general scatter plot that plots two dimensions of a risk against each other, for example the frequency of occurrence and the contribution to the 99% value at risk. A categorical risk map plots the reputational damage against the financial loss.

Below this, the report contains a table with the overall risk figures before and after measures and with a risk tolerance curve, shown in Figure 5. With it, we can assess the risk-bearing capacity in escalation levels and see whether the risk is smaller than the available capital.

Figure 5: Risk aggregates, risk-bearing capacity and rating in the results report. On the right the required risk coverage potential for different target PDs and the base data of the rating procedure (German interface).

The block “Risikoaggregate” (risk aggregates) contains the expected loss and the value at risk (VaR) of the total loss at different confidence levels. The 99.9% VaR is the loss that is not exceeded in 99.9% of the simulation runs. The measures reduce the expected loss from €6,846,905 to €2,493,313 and the 99.9% VaR from €17,990,681 to €10,934,422.

In the block “Risikotragfähigkeit” (risk-bearing capacity), each confidence level is matched by its own risk capital, from €3,000,000 for the expected loss to €13,000,000 for the 99.9% VaR. These levels form the risk tolerance curve. The column “Verwendetes Kapital” (capital used) shows how much of it the total loss after measures takes up. At the 50% VaR, it is €2,115,166 of €4,000,000, that is, 53%. The share is highest at the 99.9% VaR with 84%. At all levels, the risk after measures therefore stays below the risk capital.

More on this: Our case study “How does risk aggregation work under IDW PS 340 and StaRUG?” shows how to aggregate risks with this template and compare them with the risk-bearing capacity.

How the risks change the rating

Risk Kit contains ready-made rating procedures of the Deutsche Bundesbank, the Oesterreichische Nationalbank and a number of authors. The procedure in Figure 5 was developed by Marco Wolfrum and Werner Gleißner to calculate the effect of the value at risk on the rating. Here the rating is expressed as a probability of default (PD). On the right are the company’s base data: equity of €50,000,000, capital employed of €210,000,000, EBT of €35,000,000 and provisions of €3,000,000. The PD today is 0.77%. The columns “PD vor Maßnahmen” and “PD nach Maßnahmen” (PD before and after measures) show how the PD changes with the respective loss. At the 99.9% VaR it rises to 1.93% before and 1.36% after measures.

A rating procedure can also be used the other way round. You set a target PD and derive the required risk capital from it. In the block “Notwendiges Risikodeckungspotential” (required risk coverage potential), for example, a target PD of 0.90% goes with required risk capital of €2,900,000 and a target PD of 1.50% with €12,900,000. All these calculations, however, assume that the rating procedure is representative of your own company and therefore suitable.

How do you assess a risk with a model?

Until now, risk assessment has mostly been based on expert judgement. This is a major obstacle to effective risk management and one reason why many boards do not take the topic seriously enough. We therefore need a new approach: assessing risks with the help of models.

One example is cyber risk, which the working group had discussed before. IT departments have a great many cyber risks because they have to cover many systems and different threat scenarios. There are often ten times as many cyber risks as general corporate risks. It is therefore important to bundle these risks and deliver only a few clusters to enterprise risk management.

Cyber risk is a case-based risk. Such risks also exist in other areas. A hospital performs many operations, and some go wrong. A manufacturer sells products, and some are defective. An IT department is attacked, and some attacks succeed.

Breaking the risk down into its drivers

In risk aggregation we combine risks. To assess an individual risk, we go the opposite way and break it down into its drivers. What drives the risk, and how does the mechanism work by which it is triggered and by which its impacts arise?

For cyber risk, we first look at the frequency of attacks. For a particular type of attack there is then a certain vulnerability: most attacks are fended off, and only a few succeed. Successful attacks cause direct costs, for example because the server environment has to be set up again. Some successful attacks also cause secondary losses, for example through the theft and leaking of confidential data. This can lead to reputational damage, lawsuits, ransom payments and so on. IT routinely carries out business impact analyses for this. You could continue the list and also look at tertiary losses. The approach remains the same.

A distribution that is open to the right for the attack frequency

In the next step, we describe the drivers as distributions. The very first one, the attack frequency, is a challenge. Can we say exactly how often we will be attacked at most? Usually not. IT departments do collect a lot of data, but it has to be qualified, because not every attack is recorded and the threat situation can change quickly. So we need a distribution that is open to the right, integrates data and can be adjusted by expert judgement.

The Expert distribution is ideal for this. It uses three quantiles: the good case, the median and the bad case. Optionally we specify a minimum, while the maximum remains open. What counts as a good and a bad case is set by a probability. Here the good case is a 5% case and the bad case a 95% case.

Figure 6 shows the model with the Expert distribution for the attack frequency. In the table “Eingaben” (inputs), it has the minimum 0, the good case 100, the median 300 and the bad case 1,000 attacks. The statistics on the right show these values as the 5% quantile, median and 95% quantile. With 99.9% probability we are at no more than about 5,000 attacks, 5,122 to be exact. The preview chart shows the distribution down to one case in a million, and its axis extends to 90,000 attacks. So the distribution has a thin, long tail that also covers extreme values. That would not be possible with a PERT distribution. It is always short and stout, so to speak, and could never have such a long tail.

Figure 6: The cyber risk model with the inputs on the left and the Expert distribution for the attack frequency on the right. The distribution has a minimum of 0 and is open to the right; its maximum is infinite (German interface).

Each attack now becomes a coin toss. The successful attacks are the coin tosses in which the coin came up heads, in Figure 6 with a vulnerability of 0.075%. Whether a successful attack leads to a secondary loss is again a coin toss, one per successful attack, here with a probability of 10%. The direct and secondary losses are then the sum of the impacts of the respective successful attacks. Their size is also described in the table with a minimum, a good case, a median and a bad case.

More on this: Our case study “How do you assess cyber risks with a Monte Carlo simulation?” shows how to assess the factors of a cyber risk one by one and combine many cyber risks into a cluster.

What the simulation shows

We run the simulation and obtain the probability of occurrence and the quantiles. Nothing happens in about 76% of cases. Sometimes only smaller, primary losses occur, in rarer cases a first secondary loss and in very rare cases several secondary losses.

We then combine a portfolio of 30 cyber risks. Its distribution has a special shape with several humps that stand for one, two or more secondary losses. In the portfolio, the probability of occurrence is almost 100%, because with so many risks something always happens.

To possibly transfer losses to an external party, we also assess a cyber risk insurance policy. The expected loss is split exactly between the deductible and the insurance; in total it remains the same with and without insurance. In the extreme case, the insurance saves us a multiple of the premium. Often, however, the premium is higher than the insurance payout, because the losses naturally often stay below the deductible. From here, the insurance can be optimised further.

Taking the result into the risk aggregation

Suppose we take out the insurance and want to include the residual risk in the risk aggregation from Figure 1. To do this, we calibrate a distribution to the simulated self-retained losses and check which distribution fits best.

Figure 7 shows the result. On the left are the distributions, sorted by the Kolmogorov-Smirnov statistic. The smaller it is, the better the distribution fits. At the very top are Flex with 0.0034 and Poly with 0.0082. The modified PERT distribution ModPert with 0.0144 is also selected. In the middle, the simulated data are shown as orange bars, with the three selected distributions on top.

Figure 7: Calibrating distributions to the simulated self-retained losses. Flex (blue) follows the peaks of the data, Poly (green) and ModPert (red) smooth them (German interface).

The Flex distribution is so adaptable that it reproduces the peaks of the uncovered secondary losses very precisely. If you are worried about overfitting here, you take a Poly distribution. It smooths the peaks and shows the essential shape of the losses. Other distributions such as the modified PERT distribution also fit well. It is very well suited to being calibrated to data.

More on this: Our article “Custom distributions in a Monte Carlo simulation: the Flex distribution” shows how to calibrate a Flex distribution to data.

We choose a suitable distribution, copy it and paste it into the risk aggregation. There we set the probability of occurrence of this risk to 1, because everything that happens is already summarised in this distribution. In this way, a risk assessed with a model is transferred into enterprise risk management in a few steps.

What changes with model-based assessment

In this new kind of risk assessment, expert judgement has changed its character. In the past, experts had to assess abstract risks for the company as a whole. Today they define which concrete factors influence a risk and how these interact. The factors are tangible and measurable. If developments are possible that are not contained in the data, you can, for example, adjust the attack frequency manually and override the data at that point. Assessing an individual factor is more tangible and much easier to discuss. This allows more control and feedback.

One important keyword is validation. In conventional approaches it is often missing entirely. Forecasts are made, but it is rarely checked how realistic they were and what actually happened. With model-based assessments, validation becomes easier because there are now many concrete, observable factors. For each factor, you can check whether it stayed within the range of the forecast. You no longer have to wait for a major incident. Risk management thus gains a much clearer view of its situation.