What it means
Every business runs on models, whether or not anyone calls them that. A pricing spreadsheet, a credit scorecard, a demand forecast and a valuation built on discounted cash flows are all simplified representations of reality, and each one carries assumptions that may stop holding.
Model risk is the exposure created by relying on those simplifications. It matters because model output usually arrives stripped of its caveats.
A committee sees a single valuation, a loan limit or a churn prediction, and the uncertainty that the modeller understood perfectly well has been flattened into one number with decimal places that imply precision. Decisions then get made with more confidence than the underlying analysis supports.
Model risk comes in three broad flavours. There are mechanical errors such as a mis-typed cell reference or a broken range, data errors where the inputs are stale or mismeasured, and conceptual errors where the model's structure does not match the real world, which is by far the most dangerous because it survives every arithmetic check.
A model can be perfectly built and still be perfectly wrong. Managing it looks the same in a small business as in a bank, just lighter.
The core controls are independent review by someone who did not build the model, documented assumptions with named owners, sensitivity testing to show how the answer moves when key inputs change, and back-testing against what actually happened. Version control and locked input sheets prevent most of the mechanical problems.
The nuance worth carrying is that models built on calm historical data usually understate risk in unusual conditions. Relationships that held for a decade can break within weeks during a shock, which is why sensible modelling shows a range and a stress case rather than a single point estimate.
In practice
Real-world examples.
Example
A retailer's demand forecasting model is trained on three years of data that include an unusually warm summer. Buying decisions for outdoor furniture are set 30% too high, leaving $2,100,000 of stock to be cleared at a discount the following spring.
Example
An insurer discovers that a pricing model applies a regional adjustment factor using an outdated postcode file. Roughly 4% of policies have been priced against the wrong risk band for two years, and the remediation costs more in refunds and system work than the mispricing itself.
Example
A finance team values an acquisition target using a discounted cash flow model whose terminal growth rate was left at 4% from a previous deal. Because terminal value drives most of the answer, the error inflates the valuation by nearly 20%, and an independent review catches it days before the offer goes out.
Think of it
“Model risk is the danger that your model is wrong-errors in your mathematical assumptions.
Formula
Calculation
There is no single formula for model risk, but its cost can be quantified by rerunning a decision with corrected assumptions. A standard way to size it is through the expected credit loss calculation:
Expected credit loss = Exposure x Probability of default x Loss given default
A lender holds a $40,000,000 portfolio of small business loans. Its scoring model, calibrated on data from a strong economy, estimates a probability of default of 2% with loss given default of 60%.
Modelled expected loss = $40,000,000 x 2% x 60% = $480,000.
An independent review recalibrates the model using a full cycle of data and finds the true probability of default is closer to 3.5%.
Corrected expected loss = $40,000,000 x 3.5% x 60% = $840,000.
The model risk exposure is $840,000 - $480,000 = $360,000 of understated annual loss, equal to 0.9% of the portfolio. If the lender priced its loans on the modelled figure, its margin on this book was overstated by that amount every year.Case study
Seen in the real world.
Northgate Mutual Credit is a fictional lender used here as an illustrative example of how model risk builds up quietly. It ran a loan approval scorecard originally built by a consultant, and over six years the underlying customer mix shifted from established trading businesses to newer companies with shorter records, without the model being recalibrated.
The scorecard still returned tidy, confident scores, and default rates stayed low for years because the economy was benign. When conditions tightened, actual defaults ran at more than double the modelled rate, and a review found that two of the scorecard's most heavily weighted inputs had almost no predictive power for the newer customer base. Nothing in the spreadsheet was arithmetically wrong.
Northgate introduced annual back-testing, a written statement of assumptions with a named owner for each, and a rule that any model driving more than $10,000,000 of exposure must be reviewed by someone outside the team that built it. The illustrative point is that the failure was one of governance rather than mathematics.
Watch out
Common mistakes.
- Assuming a model is sound because its arithmetic checks out, when the most damaging errors are conceptual and leave every formula technically correct.
- Letting the person who built the model be the only person who reviews it, which removes the single most effective control available.
- Presenting a single point estimate to decision makers instead of a range, so the uncertainty known to the modeller never reaches the people taking the risk.
Questions
People also ask.
How do I test a model without specialist skills?
Change one key assumption at a time and see how far the answer moves, then compare past model predictions with what actually happened.
Does model risk apply to small businesses?
Yes, since a pricing spreadsheet or a cash flow forecast is a model, and a wrong assumption in either can cause serious harm relative to the size of the business.
Can machine learning models reduce model risk?
They can improve predictive accuracy but often increase model risk overall, because the logic is harder to inspect and the model can quietly learn patterns that do not hold in future conditions.
From the founder's library
Take it further with the book.
Build your financial confidence beyond this definition. Shihan's full-length guide, Accounting Fundamentals, takes the same plain-English approach and turns it into a complete, practical playbook for non-finance managers, business owners and students - with chapter-end quiz answers and presentation slides included.
25% off with code MMHQ25, applied at checkout. Priced in USD - checkout may show the equivalent in your local currency.
View the book and save 25%