Economic Modelling · Valuing benefit guarantees using simulation

Stochastic Asset Models and Scenario Generation for Guarantees

Updated 11 October 2026 · Fact-checked

A stochastic asset model describes how asset returns move randomly over time. To generate scenarios, you pick a model (such as lognormal), set its parameters, draw random numbers, and step the asset value forward period by period. Each path gives one guarantee payoff. The average of discounted payoffs estimates the cost.

Understand Stochastic Asset Models and Scenario Generation

A benefit guarantee, such as a minimum maturity value on a unit-linked policy, pays off depending on future asset values. Nobody knows those values. So we model them as random and simulate many possible paths. This is a scenario generator.

The simplest model is the lognormal model, from geometric Brownian motion. The log of the asset price grows by a normal amount each period. Prices stay positive, and returns over disjoint periods are independent. It needs only two parameters: a drift μ and a volatility σ.

The lognormal model is simple but has weaknesses. Real returns show fat tails, volatility that changes over time, and clustering of bad periods. A regime-switching model copes with this. The market is in one of a few states, for example calm or turbulent. Each state has its own μ and σ. A Markov chain with a transition matrix moves the market between states. Other options include jump models, GARCH-type models and mean-reverting models for interest rates.

Simulation needs random numbers. Computers produce pseudo-random uniform numbers on (0,1). You convert them to standard normals, for example by the inverse transform method Z = Φ⁻¹(U) or the Box-Muller method. Fixing a seed makes results repeatable, which matters for checking and audit.

Parameter choice depends on purpose. For pricing and hedging costs you use risk-neutral parameters: drift equals the risk-free rate. For projecting capital or real outcomes you use real-world parameters estimated from data and judgement. Using the wrong drift is a common and costly error.

Key rules to remember

Lognormal model (GBM) solution: S(t) = S(0) × exp[(μ − σ²/2)t + σ√t × Z], Z ~ N(0,1); Real-world uses μ as the expected return. Under risk-neutral valuation replace μ by r (net of any continuous dividend or charge).
Step-by-step simulation: S(t+h) = S(t) × exp[(μ − σ²/2)h + σ√h × Z]; Use a fresh independent Z for each step. h is the time step in years.
Mean and variance of log return: ln[S(t+h)/S(t)] ~ N[(μ − σ²/2)h, σ²h]; Expected price is E[S(t)] = S(0) × exp(μt).
Inverse transform for normals: Z = Φ⁻¹(U), U ~ U(0,1); Each uniform gives one standard normal.
Box-Muller: Z₁ = √(−2 ln U₁) × cos(2πU₂); Z₂ = √(−2 ln U₁) × sin(2πU₂); Two independent uniforms give two independent standard normals.
Regime-switching transition: P(next state = j | current state = i) = pᵢⱼ, with Σⱼ pᵢⱼ = 1; Each row of the transition matrix sums to 1. Draw a uniform to pick the next state.
Monte Carlo estimate and standard error: Estimate = (1/N) Σ v^T × payoffᵢ; SE = s ÷ √N; s is the sample standard deviation of discounted payoffs. To halve the error you need four times the paths.

How to solve Stochastic Asset Models and Scenario Generation questions

Use this order for any question on building scenarios or interpreting a generator.

1State the purpose: risk-neutral (pricing, hedging) or real-world (capital, projection). This fixes the drift.
2Choose the model and justify it briefly: lognormal for simplicity, regime-switching or jumps for fat tails and volatility clustering.
3List the parameters and how you would set them: μ or r, σ, and for regime models the state parameters and transition matrix.
4Set the time step h to match the guarantee, for example annual for a maturity guarantee or monthly if there are monthly charges or lapses.
5Generate uniform random numbers, convert to normals, and apply the update formula for each step. For regime models, first draw the state.
6Compute the guarantee payoff on each path and discount it, using the risk-free rate for risk-neutral work.
7Average over N paths, report the standard error, and comment on accuracy and model limitations.

Quickest way: Fast path calculation from given normals

When to use it: When the question gives you Z values and asks for a simulated asset value or payoff.

Compute the drift term (μ − σ²/2)h once.
Compute σ√h once.
For each step, compute exp(drift + σ√h × Z) and multiply it onto the running value.
For a single-step question, apply the formula once with t as the full horizon.
Apply the guarantee formula, for example max(G − S(T), 0), then discount.

Common mistakes in Stochastic Asset Models and Scenario Generation

Forgetting the −σ²/2 term in the exponent.
Students mix up the drift of the price with the drift of the log price.
Fix: Always write (μ − σ²/2) in the exponent. Check that E[S(t)] = S(0)e^(μt).
Using real-world μ when valuing a guarantee for pricing.
The data-estimated return feels more realistic.
Fix: For market-consistent cost, set drift to the risk-free rate and discount at the same rate. Use real-world drift only for real-world projections.
Using σ instead of σ√h for a shorter step.
Volatility is quoted per year and students forget to scale it.
Fix: Variance scales with time, so standard deviation scales with √h.
Reusing the same random number across steps or paths that should be independent.
Rushing, or misunderstanding independent increments.
Fix: Draw a fresh Z for each step and each path. Reuse only deliberately, as in antithetic variates.
Treating regime-switching states as permanent, or using a transition matrix whose rows do not sum to 1.
Confusing the matrix with state probabilities.
Fix: Check each row sums to 1. Draw the new state each step from the row of the current state.
Quoting the Monte Carlo estimate without any measure of error.
Students treat the simulation as exact.
Fix: Give the standard error s ÷ √N and note that more paths or variance reduction improves accuracy.

Worked examples

Example 1

An equity fund has S(0) = ₹100. Under a lognormal model with μ = 8% a year and σ = 20% a year, a simulated one-year standard normal draw is Z = 0.5. Find the simulated fund value after one year. Use e^0.08 ≈ 1.0833.

Show the solution

Drift term: μ − σ²/2 = 0.08 − 0.04/2 = 0.06.
Volatility term: σ√t × Z = 0.20 × 1 × 0.5 = 0.10.
Exponent = 0.06 + 0.10 = 0.16.
e^0.16 ≈ 1.1735.
S(1) = 100 × 1.1735 = ₹117.35.

Answer: About ₹117.35.

Example 2

A policy guarantees a maturity value of ₹100 on a fund starting at ₹100 with a 1-year term. Under risk-neutral valuation, r = 5% a year, σ = 20%. Two simulated paths use Z = −1 and Z = 1. Estimate the guarantee cost as the average discounted payoff. Use e^(−0.05) ≈ 0.9512, e^(−0.17) ≈ 0.8437 and e^(0.23) ≈ 1.2586.

Show the solution

Drift term: r − σ²/2 = 0.05 − 0.02 = 0.03.
Path 1 (Z = −1): exponent = 0.03 − 0.20 = −0.17. S(1) = 100 × 0.8437 = 84.37.
Path 2 (Z = 1): exponent = 0.03 + 0.20 = 0.23. S(1) = 100 × 1.2586 = 125.86.
Payoff = max(100 − S(1), 0). Path 1: 15.63. Path 2: 0.
Average payoff = (15.63 + 0) ÷ 2 = 7.815.
Discount at risk-free rate: 7.815 × 0.9512 = 7.43.

Answer: Estimated guarantee cost ≈ ₹7.43 per ₹100 of fund. With only two paths this is very imprecise, so the standard error would be large.

Exam tips

Always say first whether the question is risk-neutral or real-world. Examiners award marks for choosing the right drift.
Write the formula in standard notation, then substitute. Method marks are given even if arithmetic slips.
For regime-switching questions, list parameters per state and the transition matrix, and comment on why it captures fat tails and volatility clustering.
In computer-based questions, set a seed, generate normals with a vectorised function, and check the simulated mean of S(t) against S(0)e^(rt) as a sanity test.
Discuss limitations: parameter uncertainty, model risk and the number of paths needed for tail guarantees.

Practice questions from Valuing benefit guarantees using simulation

Stochastic Asset Models and Scenario Generation: frequently asked questions

Why use a lognormal model for equity returns?

Prices stay positive, returns over separate periods are independent, and it needs only two parameters. It also leads to closed-form results like Black-Scholes. Its weakness is that it understates fat tails and changing volatility.

What does a regime-switching model add?

It lets drift and volatility depend on a hidden state, such as calm or turbulent markets. A Markov chain moves between states. This produces fat tails and volatility clustering, which matter for guarantee costs.

How do I choose the parameters?

For risk-neutral valuation, set the drift to the risk-free rate and take volatility from implied or long-term estimates. For real-world work, estimate from historical data and adjust with judgement. Always state your assumptions.

How many simulations do I need?

Enough that the standard error is small relative to the answer. The error falls with the square root of N, so four times as many paths halves it. Tail guarantees need more paths.