Regression and Forecasting for Risk
Regression estimates a conditional relationship between an outcome and explanatory variables. A fitted coefficient is not proof of causation, and a good historical fit is not evidence that a forecast will remain accurate after a regime change.
The method at a glance
| Concept | How it works | Watch for |
|---|---|---|
| Prediction | Substitute inputs using the same units as the estimated model. | Adding percentage points to decimal returns. |
| Residuals | Inspect unexplained variation, dependence and changing variance. | Assuming a large R² validates the assumptions. |
| Forecast test | Use a genuine holdout or time-ordered evaluation. | Leaking future information into model fitting. |
01Prediction
How it works: Substitute inputs using the same units as the estimated model.
Watch for: Adding percentage points to decimal returns.
02Residuals
How it works: Inspect unexplained variation, dependence and changing variance.
Watch for: Assuming a large R² validates the assumptions.
03Forecast test
How it works: Use a genuine holdout or time-ordered evaluation.
Watch for: Leaking future information into model fitting.
Original worked example
Prediction versus realized return
A model predicts asset return = 0.5% + 1.3 × market return. The market returns 4%; the asset actually returns 4.8%.
- STEP 1
Predicted return = 0.5% + 1.3 × 4% = 5.7%.
- STEP 2
Residual = actual - predicted = 4.8% - 5.7% = -0.9 percentage points.
- STEP 3
A single negative residual is an observation to assess, not proof that the model is biased.
The model overpredicted this return by 0.9 percentage points. Review many out-of-sample residuals before judging performance.
Illustrative assumptions and calculations for learning. Not observed market data or an official exam item.
Apply the method and check its limits
Before you start: Covariance, conditional expectation and the units of each variable.
Interpret coefficients before forecasting
ŷ = α̂ + β̂x; residual = y - ŷ
The slope measures the fitted change in the response per unit change in x. The intercept is the fitted response at x = 0, which may be outside the observed range. Ordinary least squares minimizes the sum of squared residuals. A coefficient can capture association with omitted factors, so a large or statistically significant slope alone does not establish causation.
Audit the forecast information set
For a prediction made on Monday, every input must have been available on Monday. A revised economic series or a volatility estimate using later returns introduces look-ahead bias. Use a time-ordered holdout, compare with a simple benchmark and examine error by regime. Serial dependence and changing residual variance can also undermine conventional standard errors even when fitted values look plausible.
Confidence intervals and prediction intervals differ
An interval for the conditional mean captures coefficient uncertainty. A prediction interval for a new observation also includes residual variability, so it is generally wider. For time series, preserve time order in validation and ask whether relationships are stable. A coefficient estimated in calm markets may not describe a stressed regime.
Work through the topic
Interpret a fitted relationship, calculate its prediction and residual correctly, and evaluate whether its assumptions and information set fit a risk forecast.
Read coefficients, units and residuals together
For a simple regression y = α + βx + ε, the intercept is the fitted value when x is zero and the slope is the fitted change in y for a one-unit change in x. The unit matters. A coefficient expressed per percentage point cannot be combined casually with an input entered as a decimal fraction. Write the equation in the same convention as the supplied estimates before doing arithmetic.
Suppose the fitted return relationship is asset return = 0.5% + 1.3 × market return. If the market return is -2%, the fitted asset return is -2.1%. If the actual asset return is -3%, the residual, defined as actual minus fitted, is -0.9 percentage points. A residual is not the same as the predicted return. Its sign indicates which direction the observation differs from the prediction.
The slope describes the fitted association within the model and data. It does not establish that changing the explanatory variable would cause the predicted change in every setting. An intercept may also have little economic meaning if zero lies far outside the relevant data range. Interpretation should consider the sample, omitted variables and whether the relationship is stable enough for the intended use.
| Quantity | Calculation | Meaning |
|---|---|---|
| Fitted return | 0.5% + 1.3 × (-2%) = -2.1% | Conditional point prediction |
| Observed return | -3.0% | Realized outcome in the example |
| Residual | -3.0% - (-2.1%) = -0.9 percentage points | Observation is below the fitted value |
Original single-factor forecast, using percentage units consistently
- Fitted return
- Calculation: 0.5% + 1.3 × (-2%) = -2.1%
- Meaning: Conditional point prediction
- Observed return
- Calculation: -3.0%
- Meaning: Realized outcome in the example
- Residual
- Calculation: -3.0% - (-2.1%) = -0.9 percentage points
- Meaning: Observation is below the fitted value
Ask which assumption a diagnostic challenges
Ordinary least squares chooses coefficients to minimize the sum of squared residuals. A good in-sample fit says that the fitted equation follows the observed data relatively closely under that criterion. It does not by itself establish suitable inference, stable forecasts or a complete risk model. Separate the estimation rule from the additional assumptions used to interpret standard errors and tests.
If residual variation grows with the level of a fitted value, constant-variance standard errors may be unsuitable. If residuals are serially dependent, observations do not supply the same information as independent errors. Strong relationships among explanatory variables can make individual coefficient estimates unstable. A diagnostic is useful when it identifies a specific weakness and its consequence, rather than simply producing a pass/fail label.
For an original example, a model appears accurate in calm months but repeatedly underpredicts losses during volatile periods. Possible explanations include changing residual variance, omitted nonlinear exposure or a regime change. Plot residuals against time and relevant state variables, compare a simple alternative, and assess out-of-sample behavior. Do not automatically add variables until the historical error disappears; that can hide the original problem through overfitting.
Evaluate forecasts on information that was genuinely unavailable
A forecasting exercise must recreate the information set at each forecast date. If the model uses revised economic data, final ratings or future returns that were not then available, the backtest can look better than a usable forecast. Record release dates and transformations. Fit scaling parameters and feature selection only with training information, then apply them to the held-out period.
A confidence interval for the conditional mean reflects uncertainty in the estimated relationship. A prediction interval for a new observation also includes the randomness of that observation, so it is ordinarily wider under the same model. For a simplified illustration, if parameter uncertainty contributes variance 1 and new-outcome noise contributes variance 4, the prediction standard deviation is √5, not 1. Omitting the latter term understates uncertainty even when the mean estimate is precise.
Choose a holdout strategy that matches the use case, often chronological or rolling for time-dependent forecasts. Compare the model with an explicit benchmark and evaluate errors relevant to the decision. After deployment, monitor whether the relationship deteriorates. Statistical diagnostics tell you what is wrong; governance determines who can restrict use, approve a change or require remediation. Keep those responsibilities distinct from fitting the equation.
Does a high R-squared establish a good risk forecast?
No. It measures in-sample fit under a defined model and response variable. It does not rule out leakage, unstable coefficients, poor tail behavior or a wrong causal interpretation. The relevant question is whether the model improves the intended decision on information available at the time.
Should an unusual residual always be removed?
First determine whether it is a data error, a genuine event or evidence of a misspecified model. Removing a valid extreme observation can make a risk model look more stable while deleting the event it most needs to represent. Document any correction or exclusion and assess its effect on results.
Try a changed assumption
For the same 0.5% + 1.3 × market-return model, what is the prediction when the market loses 2%?
Show answer and reasoning
0.5% + 1.3 × (-2%) = -2.1%. The coefficient is dimensionless when both returns use matching units. This is a conditional prediction, not a guaranteed realized return.
Original learning check. No email or account needed.
Can I use R² to choose the best forecast?
Not alone. Compare forecast error on unseen data, inspect assumptions and consider whether added variables improve stability and interpretation.
Sources and scope
Links checked 2026-10-10. GARP establishes exam scope; the technical references support the methods identified below. These are selected study notes and original examples, not complete curriculum coverage or official exam questions.
- GARP FRM study materials and 2026 learning objectives
Exam-domain scope and the official route to the annual learning objectives. This page does not certify coverage of every reading.
- MIT: Regression Analysis, lecture 6
Least squares, coefficient interpretation, model assumptions and inference.
- Federal Reserve: revised model-risk guidance, SR 26-2
Current US supervisory background. Issued April 2026; supersedes SR 11-7. This does not change the prescribed FRM exam-year readings.
Prepared by MiloPrep editorial team. MiloPrep is independent of GARP. References do not imply endorsement or independent review of these notes.