A risk model connects exposures, assumptions and data to a risk estimate. A two-asset example shows how correlation changes portfolio volatility and why the output is not a loss guarantee.
Risk model
A risk model is a quantitative representation of exposures and uncertain outcomes used to estimate a particular kind of financial risk. It connects a portfolio or activity to risk drivers, assumptions and data, then reports something a decision-maker can use: for example, forecast volatility, a loss quantile, or losses under specified scenarios. The practical consequence is that a risk number is conditional on how the model represents the world.
A volatility model answers a different question from a credit-risk model that estimates defaults and recoveries. Neither automatically answers whether positions can be liquidated during market disruption. That requires attention to liquidity risk, including trading costs and funding needs.
The estimate starts before the formula
Consider a model intended to estimate the spread of a portfolio's return over one period. Before calculating anything, specify:
Component
What must be decided
Exposure
Which holdings, valuations, currencies and portfolio weights enter the calculation?
Horizon and return convention
What period is forecast, and are returns measured consistently, including income?
Risk drivers
Which market movements matter, and how are they translated into portfolio returns?
Estimation
Which observations, sampling frequency, lookback window and estimation method produce the parameters?
Output and use
Is the result a standard deviation, a tail estimate or a scenario loss, and what decision will it inform?
These choices are part of the model. A perfectly coded formula fed stale prices, mismatched return horizons or inappropriate dependence estimates can produce a misleading report.
One portfolio, two dependence inputs
Take a synthetic, fully invested portfolio with initial weights wA=wB=0.5. Assume no trades, costs or external cash flows during the period. Both assets' simple total returns refer to that same period. The model's volatility inputs are σA=20% and σB=10%. They are stipulated inputs, not live estimates, and nothing is annualised.
For this linear portfolio return, Rp=wARA+wBRB, portfolio variance is
σp2=wA2σA2+wB2σB2+2wAwBρσAσB.
Here ρ is the correlation between the asset returns; ρσAσB is their covariance. The identity requires finite second moments, but no normal-distribution assumption.
For fixed 50/50 holdings and 20%/10% asset volatilities, the variance contributions from the assets are 0.0100 and 0.0025. Correlation zero adds no covariance contribution, giving variance 0.0125 and volatility 11.1803%; correlation one adds 0.0100, giving variance 0.0225 and volatility 15%.
The same synthetic one-period portfolio produces different risk estimates when its dependence input changes. Variances use squared decimal returns; volatility is expressed as a percentage return over the period.
The second result is a sensitivity scenario, not evidence that correlation will become one. Nor does zero correlation generally mean independent returns. What the comparison establishes is narrower: the estimated diversification benefit depends on the dependence input, even when holdings and individual volatilities stay fixed.
For a worked derivation of the covariance term, Expectation, Variance and Covariance develops the underlying probability calculation. Its full lesson requires a free signed-in account; the public page provides a preview.
What the number does not promise
Volatility is spread around an expected return. It is not a guaranteed loss bound and does not specify a complete return distribution. Even a correctly specified model can assign positive probability to losses greater than one standard deviation.
For a concrete counterexample consistent with the zero-correlation inputs above, let asset A return either +20% or −20%, equally likely, and asset B independently return either +10% or −10%, equally likely. Both means are zero and the portfolio volatility is still about 11.1803%. Yet both assets fall together with probability 1/4, producing a portfolio return of −15%. These are constructed outcomes, not historical observations or a simulation.
To estimate a value at risk or expected shortfall, additional information about the loss distribution is needed. Assuming a convenient distribution is itself a modelling choice. A volatility estimate alone cannot establish a tail-loss estimate, a maximum drawdown or the cash required to survive a funding squeeze.
Validate the model against its purpose
Useful checks go beyond reproducing the arithmetic:
Challenge the representation. Does the chosen return mapping capture material exposures? A linear approximation can miss option convexity or changing positions.
Check inputs and implementation. Verify data provenance, units, timing, portfolio coverage and calculations independently.
Test estimates against subsequent outcomes. Compare forecasts with realised behaviour over matching horizons and examine performance across market conditions. Backtesting supplies evidence; a favourable sample does not establish universal reliability.
Vary assumptions and test adverse scenarios. Sensitivity analysis identifies dependence on parameters; stress testing asks what adverse conditions could do to the portfolio. Scenarios need not carry estimated probabilities.
The US agencies' 2026 model-risk guidance discusses conceptual soundness, outcomes analysis and ongoing monitoring, with review tailored to the model's use and materiality. It also recognises that material risk can remain after rigorous validation. These are useful governance principles; they do not certify a model as correct. Revised Guidance on Model Risk Management, sections V–VI.
A usable risk report therefore names its horizon, assumptions, estimation date, intended use and limitations alongside the number. Changes in exposures or market behaviour can make an old estimate less relevant. Model risk is the risk of adverse consequences from the model or its use, whereas the risk model is the analytical representation being used.
Sources
Elton, E. J., Gruber, M. J., Brown, S. J., and Goetzmann, W. N. (2014). Modern Portfolio Theory and Investment Analysis, 9th edition, Chapter 4, pp. 51–53. Portfolio-return variance, covariance and correlation; the examples here are original.
