Gains:
- Ability to establish deterministic scenario, stochastic simulation (Monte Carlo) and stress test concepts with the support of artificial intelligence and interpret the results in risk language
- Ability to design shock scenarios such as inflation, interest, catastrophic damage and pandemic with artificial intelligence and estimate liability and capital impact
- Being able to distinguish that the scenarios produced by artificial intelligence are based on historical data and that actuarial judgment comes to the fore in tail risks and structural breaks.
Actuarialism is the profession of predicting the future; But the future is not a single number, it is a spectrum of possibilities. It is not enough to say "Our expected damage is 50 million"; The real question is: "If a bad year comes, how many millions will it cost, can the company handle it?" Tools that answer this question are scenario analysis and stress testing. Measuring in advance the impact of an earthquake, high inflation, interest rate shock or pandemic on the company's liabilities and capital is a fundamental requirement of both internal risk management and legislation such as Solvency II. In this unit, we will cover deterministic scenario, stochastic simulation (Monte Carlo) and stress testing, and see how to use AI in these analyses.
Let's remind you from the beginning: scenarios are based on historical data and assumptions. AI designs scenarios, writes simulation code, and interprets results; But which shocks are realistic, how to handle tail risk, and how to include structural breaks (events that have never happened in the past) are left to actuarial judgment.
Deterministic scenario, stochastic simulation and stress testing
It is important to distinguish three concepts. The deterministic scenario is to choose a single set of assumptions and calculate the outcome: “What would our claims cost be if inflation was 40 percent?” Simple, clear and easy to communicate; But it only shows one way, not a possibility. Stochastic simulation (Monte Carlo) derives a probability distribution of the outcome by repeatedly generating thousands of random scenarios. Thus, tail information such as "50 on average, but 78 million at the 99.5 percent level" is obtained. A stress test is a special type of scenario that measures the impact of certain severe but plausible events (earthquake, market crash); usually defined by the regulator.
Monte Carlo is one of actuarial's most powerful tools. The logic is this: you extract random values from the frequency and intensity distributions and calculate the total damage; you repeat this 10,000 times; Accumulates 10,000 possible total damage in your hand. From this accumulation, you can read the average, percentiles such as 75 percent, 99.5 percent and worst-case scenarios. This is exactly how the 99.5 percent level required by Solvency II (a bad year that will occur once in 200 years) is found. AI quickly produces the code for this simulation.
The following table compares the three approaches:
Approach
what gives
strong point
limit
Deterministic scenario
single result
Simple, easy to communicate
Does not give probability
Monte Carlo
Probability distribution
Indicates tail risk
Dependent on assumption and data
stress test
Result under shock
Tests durability
Shock selection is judicial
Tail risk: the most important and most dangerous area
In actuarial theory, the real danger lies not in the mean but in the tail (the extreme, low-probability but high-impact region of the distribution). What sinks companies is not an average year, but a rare but devastating event: a major earthquake, a lot of major damage coming at the same time, a sudden interest rate shock. To measure this region, VaR (Value at Risk — the maximum loss at a given confidence level) and TVaR/CTE (average loss beyond the tail) are used. Monte Carlo is the natural way to calculate these measurements.
But there is another danger in the tail: historical data does not represent the tail well. Rare events, by definition, have been rare in the past; so the model tends to underestimate them. AI easily fits a distribution based on historical data, but it doesn't spontaneously consider "what if there's a pandemic or a 100-year earthquake that this portfolio has never experienced?" It is the actuary's job to include structural breaks, correlation increases (everything getting worse at the same time during a crisis) and trends such as climate change into the model.
Caution: The sentence "Our model handles 99.9 percent scenarios" is comforting, but remember that the model learns from data whose tail is past. An event that never happened is the blind spot of even the best model. This is where expert judgment comes into play.
How to use AI in scenario and stress testing
1) Monte Carlo simulation code:
Write a compound damage Monte Carlo simulation with Python (with comments). Assumptions (I give): number of damage ~ Poisson(lambda=1600), amount of damage ~ Gamma(shape=k, scale=theta) [I will give k and theta].1) 10,000 times: draw quantity, draw that amount, add up.2) Average of total damages, 75%, 99%, 99.5% percentages print.3) Calculate and explain VaR(99.5%) and TVaR(99.5%).FITTING parameters; I will give it. Library: numpy.
2) Deterministic stress scenario:
Your role: risk assistant. My base expected loss is 50 million TL. Set up the following 3 deterministic scenarios: A) Inflation shock: severity +30%. B) Frequency shock: number of damages +15%. C) Both together. Calculate the new expected damage in each scenario, show the percentage impact based on the base and comment which one is the most critical. Just use the base number I gave you.
3) Scenario design consultancy:
I am designing stress test scenarios for a health insurance portfolio. - Propose a draft 'severe but plausible' scenario each for the pandemic, high inflation, new expensive treatment technology and regulatory change. - State whether each scenario affects frequency, severity or both. - Point out tail risks that cannot be captured with historical data. I will make the decision; You provide the framework and justification.
4) Translating the result into management:
My Monte Carlo result: expected damage 50M, VaR(99.5%) 78M, TVaR(99.5%) 89M. Explain this to a board of directors in 5 sentences:- What does 99.5% and TVaR mean, in plain language.- Why is 'planning by the average' inadequate?- What figure should they look at in terms of capital?
Weak prompt / Strong prompt
Weak prompt:
Do a stress test for my portfolio.
There is no base value, no shock definition, no distribution. AI produces a general text, not a feasible number.
Powerful prompt:
Your role: actuarial risk assistant. Base assumptions (anonymous): expected frequency 0.12, average severity 30,000 TL, exposure 20,000 policy-years. Technical interest is 4%.Task:1) Calculate the base expected total loss.2) Apply the following stresses and show the percentage impact of each: - Inflation: severity +25% - Bad winter: frequency +10% - Interest rate decline: 4% -> 3% (impact on the present value of the liability)3) Comment on the 2 most critical scenarios and why.4) Do you recommend Monte Carlo for tail risk, why?Just use the numbers I gave; If it's missing, ask.
three mini cases
Case 1 — The deceitfulness of the average. A company had planned its capital based on expected damage (50 million). When Monte Carlo was done, it was seen that the 99.5 percent level was 82 million: in a bad year, there was an additional need of 32 million and the company would not be able to meet it. The capital plan has been revised. AI ran simulation of 10,000 scenarios in seconds; The actuary evaluated the interpretation and distribution.
Case 2 — Correlation blindness. Earthquake and business interruption coverage separately appeared low risk in a portfolio. But an earthquake triggers both at the same time; The model assuming independence underestimated tail risk by half. When the actuary added correlation to the scenario, the real capital need increased significantly. Lesson: in a crisis, risks are not independent; Question AI's assumed acceptance of independence.
Case 3 — An event that did not happen. A healthcare portfolio looked very "solid" based on data from the past 10 years. When a pandemic arrived, both frequency and severity skyrocketed simultaneously; The model never predicted this because there was nothing like it in the data. In the next period, the deterministic pandemic scenario was added. AI produced the draft script; The actuary decided the magnitude of the shock. Lesson: historical data queue is not complete.
Common mistakes
- Planning around the average. It is the tail, not the average, that sinks the company; Look at extreme metrics like VaR/TVaR.
- Assuming risks are independent. During a crisis, correlations increase; Acceptance of independence makes light of the tail.
- Full confidence in historical data. Rare and structural events are absent from the data; Add scenario with expert judgment.
- Mixing deterministic and stochastic. A single scenario does not give probability; Probability requires Monte Carlo.
- Presenting the simulation result without verifying it. Check percentages and VaR independently; Parameter error distorts the entire distribution.
Tip: Before presenting Monte Carlo results, do a simple consistency check: is the simulation mean close to the expected value (frequency × intensity × exposure) that you calculated by hand? If not, there is a parameter or code error.
In summary
The future is not a single number, but a spectrum of possibilities. Deterministic scenarios show a single "what if" path, Monte Carlo shows the entire distribution and tail risk, stress tests show the impact of severe events. The real danger is in the tail, not the average, and historical data does not represent the tail well; Increased correlation and events that have not occurred require actuarial judgment. AI designs the scenario, codes the simulation and interprets the result; but shock selection, correlation assumption, tail judgment, and final risk decision rest with the actuary. Always verify results with a simple consistency check.
Application task
Set anonymous base assumptions (frequency, intensity, exposure, interest). Ask the AI to (a) base expected damage, (b) 3 deterministic scenarios including inflation, frequency and interest rate shocks, (c) calculate VaR and TVaR(99.5%) with a Monte Carlo simulation. Verify that the simulation mean is close to the expected value you calculated manually. Then discuss with the AI how assuming the two collaterals are dependent (correlated) would change the tail.
checklist
- [ ] Have I also looked at the tail measures (VaR, TVaR) rather than the average?
- [ ] Have I taken into account the correlation of risks at the time of the crisis?
- [ ] Have I added structural/rare scenarios not found in historical data with expert judgment?
- [ ] Have I consciously separated the deterministic and stochastic approach?
- [ ] Have I manually checked the Monte Carlo mean for consistency with the expected value?
- [ ] Have I written the justification for scenario choices and shock magnitudes?