Skip to content
Free previewYou are reading 1 of 111 lessons, in the same interface members use. Everything else unlocks with full access.Get full access · $150

1.1.1 - Multiple Regression Coefficient Interpretation

What you'll learn

  • Specify an investment regression and interpret a slope in the units of its exhibit.
  • Calculate a fitted level, a change between scenarios, and a signed residual.
  • Recover group intercepts and slopes from indicators and interactions.
  • Use residual evidence to distinguish a problem with the fitted mean from a problem with uncertainty.

Estimated reading time: 18 minutes, including practice.

WHY

Why it matters

A credit analyst observes that highly leveraged issuers pay wider spreads. A regression exhibit lets the analyst ask a narrower question: how much wider is the estimated spread for issuers with the same interest coverage and sector? In an item set, the same exhibit can support three different answers depending on whether the question asks for a spread, a spread difference, or an unexplained spread.

KEY TERMS

Key terms

Conditional mean
The average outcome in the population at specified values of all included predictors.
Partial slope
The rate at which the fitted outcome changes along one predictor, holding the others fixed in an additive model.
Disturbance
The unobservable difference between an outcome and its population regression function.
Residual
The observable difference between a recorded outcome and its fitted value.
Reference category
The group represented when every indicator for a categorical variable equals zero.
Interaction
A model term through which the slope or effect of one predictor depends on the value of another.
Homoskedasticity
Constant conditional variance of the disturbances across predictor values.
CORE

Core content

1.1.1.1 Specify the outcome and fit the conditional mean

Suppose an analyst wants to explain differences in option-adjusted spreads across corporate bonds observed on the same date. Each bond supplies one observation. Its spread is the dependent variable because that is the quantity to be explained. Leverage and interest coverage are explanatory variables because their measured values enter the model. Calling these variables independent variables describes their role in the equation. The two financial ratios may be correlated.

For observation \(i\), a two-predictor population model is \[Y_i=\beta_0+\beta_1X_{1i}+\beta_2X_{2i}+\varepsilon_i.\] Here \(Y_i\) is spread in basis points, \(X_{1i}\) is leverage in turns of debt to EBITDA, and \(X_{2i}\) is coverage in turns of earnings before interest and taxes to interest expense. EBITDA means earnings before interest, taxes, depreciation, and amortization. The intercept \(\beta_0\) and disturbance \(\varepsilon_i\) are in basis points. The slopes \(\beta_1\) and \(\beta_2\) are in basis points per turn of their respective ratios. Population coefficients are unknown. An estimated equation replaces them with sample estimates, commonly written \(b_0,b_1,b_2\).

Linear regression is linear in the coefficients. For example, an analyst can include leverage squared alongside leverage, assigning a separate coefficient to each. The fitted relationship can then bend as leverage changes. In that specification a one-turn increase changes both leverage terms, so interpreting the coefficient on leverage alone would miss part of the change.

Table 1 distinguishes the quantities an analyst can actually obtain from an estimation exhibit.

Table 1. Population and sample quantities
QuantityMeaning in the spread modelHow the analyst obtains it
Population slopeUnknown coefficient in the population relationshipEstimate it from observations
Estimated slopeWeight assigned to a measured predictorRead the coefficient column
Fitted spreadEstimated mean spread at the bond's inputsSubstitute its ratios into the estimated equation
DisturbanceDeviation from the population relationshipUnobservable because the population coefficients are unknown
ResidualDeviation from the estimated relationshipSubtract the fitted spread from the observed spread

Ordinary least squares (OLS) chooses the coefficients jointly to minimize \(\sum_{i=1}^{n}(Y_i-\hat Y_i)^2\), where \(n\) is the number of observations and \(\hat Y_i\) is the fitted outcome. The difference is measured vertically, in outcome units. Squaring prevents positive and negative misses from canceling and gives larger misses more weight. A 20-basis-point miss contributes four times as much to the objective as a 10-basis-point miss.

What a zero average residual tells you. With an intercept, unweighted OLS residuals sum to zero over the estimation sample, apart from rounding. If every fitted spread were too low on average, raising the intercept would improve the fit. The zero sum follows from estimation; it can coexist with systematic underprediction for low-leverage bonds and overprediction for another group. It need not hold in a new sample.

A Check an estimation exhibit against its sample

With an intercept, averaging the fitted equation gives \(\bar Y=b_0+\sum_{j=1}^{k}b_j\bar X_j\). Suppose mean spread is 150 basis points, mean leverage is three turns, mean coverage is four turns, and the reported slopes are 28 and −6. The missing intercept must be \(150-28(3)-(-6)(4)=90\) basis points. Take the mean of each column exactly as it entered estimation. For a squared predictor, use the mean of the squares, \(\overline{X^2}\), which generally differs from \((\bar X)^2\).

Count the estimated coefficients to recover the residual degrees of freedom. For \(n\) complete observations, an intercept, and \(k\) linearly independent predictor columns that are jointly full rank with the intercept, the count is \(n-k-1\). A model fitted to 60 loans with leverage, an unsecured indicator, and their interaction has \(60-3-1=56\) residual degrees of freedom. The product uses a coefficient of its own even though the analyst constructs it from two existing inputs.

B Inspect predictor pairs before interpreting slopes

A scatterplot matrix displays each pair of variables across the same observations. A tight diagonal band in the leverage and coverage panel signals little independent variation in those inputs. A bend in the spread and leverage panel suggests investigating functional form. Each point still represents a bond with its own coverage value. The trend through that panel therefore describes a pooled association. To read the leverage effect at fixed coverage, use the multiple-regression equation.

Figure 1 compares two possible leverage and coverage panels drawn on the same scales.

Fig 1. Independent variation in two predictor clouds
Key point: A narrow diagonal band leaves little leverage variation among bonds with similar coverage.
{"type":"normal_vs_abnormal","describe":"Each dot represents a bond, with leverage increasing to the right and coverage increasing upward. In the first panel bonds at a given coverage span a wide range of leverage. In the second they lie close to a descending line.","teaching_point":"A tight predictor relationship limits the information available to estimate separate slopes.","prompt":"Paired statistical scatterplots in clean textbook style on white. Front view, identical rectangular axes and scales, leverage horizontal and coverage vertical. Use small navy points, a broad oval cloud on the left and a narrow descending cloud with visible scatter on the right. MARKERS: 2 numbered circular markers, each a white digit on a solid dark circle, each joined to its target by a short thin leader line: 1 on the broad cloud; 2 on the tight descending cloud. Use each digit exactly once, large and easy to read. NO words, captions or text labels anywhere in the image: digits 1 to 2 only. Emphasize the different horizontal spread at a common height. Omit arrows, fitted lines and decoration.","caption":"Fig 4. Independent variation in two predictor clouds","size":"1536x1024"}
  1. Weak pairwise association. Bonds at similar coverage have substantial variation in leverage.
  2. Strong pairwise association. Leverage is largely predictable from coverage within this sample.

1.1.1.2 Interpret a slope using a controlled comparison

A slope answers a comparison with a specified starting point and change. In an additive spread model, a leverage coefficient of 28 means that two otherwise matched predictor profiles differing by one turn of leverage have fitted spreads differing by 28 basis points. The higher-leverage profile has the wider fitted spread. For a half-turn increase, the difference is 14 basis points. The two actual bonds can differ by more or less because each has its own residual.

Controlling for coverage can change the leverage coefficient. Suppose firms with high leverage also tend to have weak coverage, and weak coverage is associated with wider spreads. A simple regression on leverage can attribute some coverage-related variation to leverage. The multiple regression uses variation in leverage remaining after accounting for coverage. Its slope may be smaller, larger, or even opposite in sign to the simple slope, depending on the relationships in the data. A regression association by itself does not establish what would happen if management deliberately increased debt.

The intercept describes the fitted outcome when all numerical inputs equal zero and all indicators take their reference values. In a model of funds' monthly excess returns against market and sector excess returns, simultaneous zero factor returns can be a meaningful baseline. For a company with a debt-to-EBITDA ratio of zero and interest coverage of zero, the financial interpretation may be implausible. The intercept still belongs in the fitted equation. Centering a predictor at a typical value can make the baseline informative: using leverage minus three makes zero on that predictor represent three turns of leverage.

Table 2 shows how the same economic movement reaches the equation under different recording conventions.

Table 2. Units in a coefficient exhibit
Predictor conventionEconomic movementChange enteredIllustrative slope and fitted effect
Yield in percentage points4% to 5%112 basis points per point gives 12 basis points
Yield as a decimal4% to 5%0.011,200 basis points per decimal unit gives 12 basis points
Yield in basis points4% to 5%1000.12 basis points per basis point gives 12 basis points
Margin in percentage points10% rises by 10% relative to its starting value1Margin ends at 11%; a slope of 2 gives 2 output units

In general, multiplying every value of a predictor by 100 divides its additive slope by 100 to keep the fitted contribution unchanged. Changing the units of the dependent variable instead rescales every coefficient and residual in the same direction. Reporting fitted spreads in percentage points rather than basis points divides each coefficient and residual by 100. These transformations explain why the largest raw coefficient need not identify the economically most important predictor. Compare the fitted effects of economically relevant movements in common outcome units.

A Checkpoint

A portfolio model expresses monthly returns in percentage points and has a market slope of 1.20. The market return rises from 2.0% to 2.5%; the other factor is unchanged. What is the change in fitted portfolio return?

  1. 0.60 percentage points
  2. 1.20 percentage points
  3. 3.00 percentage points
  4. 0.006 percentage points
Answer: A. The input rises by 0.50 in the exhibit's units, so the fitted increase is \(1.20\times0.50=0.60\) percentage points. The 3.00 figure uses the ending market level and measures its contribution to the ending return.

B Compare standardized coefficients

If the exhibit standardizes both the outcome and the continuous predictors by subtracting their respective means and dividing by their sample standard deviations, the slope becomes \(b_j^*=b_js_{X_j}/s_Y\). A raw leverage slope of 28 basis points per turn, leverage standard deviation of 0.5 turns, and spread standard deviation of 35 basis points give \(28(0.5)/35=0.40\). At fixed coverage, one standard deviation more leverage corresponds to 0.40 standard deviations more fitted spread. If only leverage was standardized, the answer would instead be 14 basis points.

C Read coefficients when the exhibit uses natural logarithms

First identify which side of the equation is logged. Table 3 translates the coefficient \(b\) into a finite change, holding all other regressors fixed.

Table 3. Coefficient interpretations with natural logarithms
Outcome and predictorInput movementExact fitted comparison
Level outcome, log predictorMultiply the predictor by \(r\)Outcome changes by \(b\ln r\) outcome units.
Log outcome, level predictorAdd \(d\) predictor unitsExponentiated fitted outcome changes by \(100[\exp(bd)-1]\) percent.
Log outcome, log predictorMultiply the predictor by \(r\)Exponentiated fitted outcome changes by \(100(r^b-1)\) percent.

Level-log. Suppose spread in basis points has a slope of −20 on log issue size. A 10% larger issue gives \(r=1.10\), so the fitted spread changes by \(-20\ln(1.10)\approx-1.91\) basis points. The percentage movement belongs to issue size. The answer retains the spread units printed in the exhibit. For a small 1% increase, multiplying the slope by 0.01 gives a close approximation.

Log-level. Suppose log spread has a leverage slope of 0.08 per turn. One additional turn changes fitted log spread by 0.08. On the original scale the fitted ratio is \(\exp(0.08)\), an increase of about 8.33%. The familiar 8% answer is an approximation. Exponentiating a fitted mean of log spread gives a conditional geometric-mean estimate. Recovering an arithmetic mean spread generally requires an additional adjustment for the disturbance distribution.

Log-log. Suppose the slope on log issue size in a log spread model is −0.40. A 1% larger issue corresponds approximately to a 0.40% lower fitted spread. For a doubling, preserve the full ratio: \(2^{-0.40}\approx0.758\). The exponentiated fitted spread falls about 24.2%. Multiplying the elasticity by the 100% increase would substantially overstate the decline.

D Calculate a finite move with a squared predictor

Suppose the leverage terms in a spread equation are \(10L+2L^2\), in basis points. At three turns, the marginal slope is \(10+4(3)=22\) basis points per turn. For an actual rise from three to four turns, take the difference between the two fitted contributions: \(10(4-3)+2(4^2-3^2)=24\) basis points. The slope steepens during the move. Squaring the one-turn change would give the wrong contribution because the squared column moves from 9 to 16.

E Determine the direction of omitted-variable bias

Suppose the population spread equation contains leverage and coverage with zero conditional mean disturbance. Let \(\beta_C\) be the coverage slope and \(\delta\) the population slope from regressing coverage on leverage. Dropping coverage changes the population leverage slope by \(\beta_C\delta\). If each extra turn of coverage narrows spreads by six basis points and coverage tends to fall by half a turn per extra turn of leverage, the shift is \((-6)(-0.5)=+3\) basis points per turn. The pooled leverage slope is higher. Reversing either sign reverses the bias. A zero value for either factor removes this source of bias.

1.1.1.3 Calculate a level, a change, and a signed miss

A fitted level uses the complete estimated equation. A comparison between two fitted levels from that equation uses only terms that change. For an additive two-predictor model, \[\Delta\hat Y=b_1\Delta X_1+b_2\Delta X_2.\] Each delta means ending value minus starting value. An unchanged predictor contributes zero to the difference even when it contributes substantially to each level. A residual requires an actual observation, \(e_i=Y_i-\hat Y_i\). A positive residual places the outcome above the fitted mean; a negative residual places it below.

Table 4 gives the analyst's estimated equation and the inputs for two scenarios on the same issuer.

Table 4. Issuer spread model and scenario inputs
Model termEstimated coefficientCurrent inputRevised input
Intercept90 basis points11
Debt to EBITDA28 basis points per turn3.0 turns3.5 turns
Interest coverage−6 basis points per turn4.0 turns5.0 turns

Worked case

The issuer's observed spread is 162 basis points and its coupon is 5.25%. A credit committee asks how expensive the current borrowing spread looks relative to the model and how the revised fundamentals affect its fitted value.

Start with the current fitted benchmark. Leverage contributes \(28\times3=84\) basis points and coverage contributes \(-6\times4=-24\) basis points. Adding the intercept gives \(90+84-24=150\) basis points. The observed spread sits above that benchmark, so the residual must be positive: \(162-150=12\) basis points. The coupon is a contractual rate; the model has no coupon term. A wider observed spread describes the discrepancy but does not establish that the bond is mispriced.

Under the revised inputs the extra half turn of leverage adds 14 basis points, while the extra turn of coverage removes 6. The net increase is 8 basis points, giving a revised fitted spread of 158. Direct substitution checks both the magnitude and direction: \(90+28(3.5)-6(5)=158\). There is no revised observed spread, so a revised residual cannot yet be calculated.

Fig 2 shows the fitted spread \(\hat S=90+28L-6C\), where \(L\) and \(C\) denote leverage and coverage in turns and \(\hat S\) is measured in basis points.

Fig 2. Spread sensitivity at fixed coverage
Key point: Stronger coverage lowers the entire fitted line while the leverage slope stays constant.
{"x":{"label":"Debt to EBITDA (turns)","min":1,"max":6},"y":{"label":"Fitted spread (bp)","min":50,"max":260},"curves":[{"label":"Fitted issuer spread","expr":"90+28*x-6*c"}],"sliders":[{"name":"c","label":"Interest coverage (turns)","min":2,"max":8,"step":0.5,"value":4}],"readouts":[{"label":"Spread at three turns of leverage (bp)","expr":"90+28*3-6*c","digits":0}],"try":"Predict the spread at three turns of leverage if coverage improves by one turn, then raise coverage from 4 to 5 to check.","say":"The line rises by twenty-eight basis points for every additional turn of leverage. Increasing coverage shifts the line down by six basis points per extra turn.","prompt":"a live graph of fitted spread (bp) against debt to EBITDA (turns). The curve shows fitted issuer spread. The line rises by twenty-eight basis points for every additional turn of leverage. Increasing coverage shifts the line down by six basis points per extra turn. On the page, a slider changes the interest coverage (turns), and the curve redraws as it moves. It also displays spread at three turns of leverage (bp). Try this: Predict the spread at three turns of leverage if coverage improves by one turn, then raise coverage from 4 to 5 to check."}

A fitted value describes the estimated conditional mean. A future individual observation also includes an unknown disturbance. Even an accurate estimate of the mean can coexist with wide variation among individual bonds. Predictions become especially fragile outside the range of observed predictor combinations. A sample might contain both high-leverage firms and strong-coverage firms but almost no firms with both characteristics. Checking each predictor's separate minimum and maximum would miss that extrapolation.

A Choose the interval that matches the prediction

A committee asks first for the average spread of comparable issuers and then for one new issuer's realized spread at the same inputs. Both intervals share the same fitted center. The first uses the standard error of the estimated conditional mean. The second also allows for the new issuer's disturbance. Under classical assumptions, if the mean estimate's standard error is 4 basis points and the residual standard error is 12 basis points, the individual prediction standard error is \(\sqrt{4^2+12^2}\approx12.65\) basis points. Apply the same critical value to each standard error. Add the variances before taking the square root.

1.1.1.4 Recover the comparison encoded by an indicator

Consider an analyst modeling spreads across utilities, industrials, and retailers at the same duration. Sector names carry no numerical distance. Coding the sectors as one, two, and three in a single continuous term would impose equal steps and an arbitrary ordering. Indicators instead give each nonreference sector its own intercept adjustment. With an intercept and three exhaustive, mutually exclusive groups, use two indicators. The omitted sector remains in the data and supplies the baseline.

Table 5 translates each possible row of two sector indicators into a spread equation, with duration \(T\) measured in years and spreads in basis points.

Table 5. Sector coding with utilities as the reference
SectorIndustrial indicatorRetail indicatorFitted spread (bp)
Utility00\(80+4T\)
Industrial10\(95+4T\)
Retail01\(105+4T\)

The industrial coefficient is 15 basis points relative to utilities and the retail coefficient is 25. At a common duration, retail exceeds industrial by \(25-15=10\) basis points. At different durations the comparison includes the duration contribution as well. These are adjusted differences; the raw sector averages can also reflect differing duration distributions.

For \(g\) categories, the usual intercept specification has \(g-1\) indicators. Including every indicator together with the intercept makes their sum exactly equal the intercept column of ones. Multiple coefficient vectors then reproduce the same fitted values. An alternative coding drops the common intercept and includes all group indicators; each indicator then supplies its own group intercept. Its coefficient has a different interpretation from the coefficient in a reference-category exhibit.

Changing the reference changes the comparison. If retail becomes the baseline in Table 5, the intercept becomes 105 and the utility and industrial coefficients become −25 and −10, respectively. Every fitted spread is preserved. For a binary comparison, swapping the zero and one labels adds the old group coefficient to the intercept and reverses the coefficient's sign in an additive model.

A When the group also changes the slope

Suppose \(D\) is a binary indicator, \(X\) a continuous predictor, and \(b_3\) the coefficient on their product. The fitted equation is \[\hat Y=b_0+b_1X+b_2D+b_3XD.\] For the reference group, setting \(D=0\) gives intercept \(b_0\) and slope \(b_1\). For the comparison group, setting \(D=1\) gives intercept \(b_0+b_2\) and slope \(b_1+b_3\). Subtracting the two equations at the same \(X\) yields the group difference \(b_2+b_3X\). The indicator coefficient alone describes the gap at zero; the interaction coefficient describes how fast the gap changes.

Worked case

An analyst estimates monthly fund excess return, in percentage points, as \(\hat R=0.10+0.80M+0.20D+0.30MD\). Market excess return \(M\) is entered in percentage points. The indicator is one in an expansion and zero otherwise. The fund charges a 0.40% annual management fee. What is the fitted expansion return at a market excess return of 2%, and how far is it above the reference prediction?

In an expansion the intercept is \(0.10+0.20=0.30\) and the market slope is \(0.80+0.30=1.10\). The fitted return is \(0.30+1.10(2)=2.50\) percentage points. In the reference regime the return is \(0.10+0.80(2)=1.70\). The expansion gap is therefore 0.80 percentage points. The same gap follows directly from \(0.20+0.30(2)=0.80\). The annual fee has no separate term in this equation. Subtracting it would change the modeled quantity, and treating 0.30 as the expansion slope would confuse the increment with the total.

Fig 3 lets you change the interaction coefficient and observe where the two regimes give the same fitted return.

Fig 3. Market sensitivity across two regimes
Key point: Changing the interaction rotates the expansion line around its intercept.
{"x":{"label":"Market excess return (%)","min":-4,"max":4},"y":{"label":"Fitted monthly excess return (%)","min":-6,"max":6},"curves":[{"label":"Reference regime","expr":"0.10+0.80*x"},{"label":"Expansion","expr":"0.30+(0.80+h)*x"}],"sliders":[{"name":"h","label":"Additional market slope in expansion","min":0,"max":0.5,"step":0.05,"value":0.3}],"readouts":[{"label":"Expansion gap at a 2% market return (percentage points)","expr":"0.20+2*h","digits":2}],"try":"Predict whether the two lines will cross when their slopes become equal, then reduce the additional market slope from 0.30 to zero.","say":"Both lines rise with the market. Reducing the interaction flattens the expansion line until the two lines are parallel, with expansion two tenths of a percentage point higher.","prompt":"a live graph of fitted monthly excess return in percent against market excess return in percent. It has 2 curves: reference regime and expansion. Both lines rise with the market. Reducing the interaction flattens the expansion line until the two lines are parallel, with expansion two tenths of a percentage point higher. On the page, a slider changes the additional market slope in expansion, and the curve redraws as it moves. It also displays expansion gap at a 2% market return (percentage points). Try this: Predict whether the two lines will cross when their slopes become equal, then reduce the additional market slope from 0.30 to zero."}

B Checkpoint

Using Table 5, how much wider is a retail bond's fitted spread at six years of duration than an industrial bond's fitted spread at four years?

  1. 8 basis points
  2. 10 basis points
  3. 18 basis points
  4. 33 basis points
Answer: C. The retail estimate is 129 basis points and the industrial estimate is 111. The 10-basis-point sector gap would apply at equal durations; two additional years contribute another 8 basis points.

C Find the market return where the regime ranking reverses

Set the fitted group difference equal to zero. In the fund example, solve \(0.20+0.30M=0\) to obtain \(M=-0.20/0.30\approx-0.67\) percentage points. At a market excess return of −1%, the expansion prediction is lower by 0.10 percentage points. At a zero market excess return, it is higher by 0.20 points. In general the crossing occurs at \(X=-b_2/b_3\) when \(b_3\ne0\). Check the observed range in both groups before interpreting a calculated crossing as an empirically supported reversal.

1.1.1.5 Read residual evidence against the model assumptions

Consider a bond whose fitted spread is 140 basis points. A residual of 20 basis points is an individual miss. A cluster of positive residuals across many similarly leveraged bonds is evidence about the model. To interpret the cluster, distinguish where the residuals are centered, how widely they spread, and whether neighboring observations are related.

The population condition \(E(\varepsilon\mid X)=0\) says that disturbances average zero at each combination of regressors, where \(X\) represents the full set of included inputs. If omitted liquidity risk widens spreads and is correlated with included leverage after accounting for coverage, part of the estimated leverage slope can reflect liquidity. OLS residuals in a sample with an intercept are mechanically orthogonal to each included regressor. A zero sample linear correlation between residuals and leverage therefore cannot establish exogeneity. A curved pattern can still reveal a deficient conditional mean.

Table 6 contrasts the diagnostic evidence available from different displays.

Table 6. Reading regression diagnostics
DisplayObserved featureInterpretation
Residual versus fitted value or predictorCenter follows a U shapeA missing nonlinear term or other mean misspecification is plausible
Residual versus fitted valueCenter near zero; band widensConditional disturbance variance may increase with fitted level
Residual versus observation dateLong runs above or below zeroPositive serial dependence is plausible
Residual versus observation dateRepeated alternation of positive and negative deviationsNegative serial dependence is plausible
Normal quantile plotResidual quantiles depart markedly from a straight line in the tailsNormal errors may be a poor approximation
Residual versus fitted valueOne point lies far above the surrounding cloudInvestigate an unusual conditional outcome and assess its influence separately

Homoskedasticity concerns conditional variance. When variance changes but the mean equation and exogeneity assumptions hold, heteroskedasticity alone does not bias the OLS slopes. It can invalidate conventional standard errors. Serial correlation concerns covariance between disturbances on different observations, often successive dates. Reordering observations by fitted value erases the time sequence that makes a run visible. Either issue can undermine conventional inference; the evidence for one does not automatically establish the other.

Fig 4 separates the diagnostic questions according to what the analyst can observe.

Fig 4. Interpreting residual patterns
Key point: A change in the center, a change in spread, and a pattern over time point to different model problems.
flowchart TD
A[Residual pattern observed] --> B[Against fitted values]
A --> C[Against observation date]
B --> D[Curved local center]
B --> E[Widening vertical spread]
D --> F[Reconsider mean specification]
E --> G[Investigate nonconstant variance]
C --> H[Runs of similar signs]
H --> I[Investigate serial dependence]

Other assumptions require different evidence. Unique OLS coefficients require full column rank: no included regressor can be reproduced exactly from the other columns, including the intercept. Strong but imperfect collinearity can make individual estimates imprecise or unstable while fitted values for typical observations remain reasonably stable. Inspect the predictor definitions and their relationships. A random-looking residual plot cannot establish full rank or rule out high collinearity.

Conditional normality of disturbances supports exact conventional small-sample tests when the remaining classical assumptions hold. Normality is not required to compute OLS estimates or for unbiasedness under zero conditional mean and full rank. A normal quantile plot assesses distributional shape, while the predictor variables themselves may have skewed distributions. The coefficient's sign and economic magnitude are separate from statistical significance; assessing the latter requires uncertainty information.

A point with an unusual predictor combination has high leverage in the statistical sense. This use of leverage describes position in predictor space, even if financial leverage is absent from the model. A point can lie far from the predictor center yet close to the fitted surface. Influence asks how much estimates or fitted values change when the observation is removed. A large residual prompts investigation; it does not by itself determine leverage or influence.

Match each observation to the property it most directly describes.

A few extreme input combinations:
Potentially high statistical leverage.
A large vertical miss:
Unusual outcome given the fitted mean.
A slope changes materially on deleting one issuer:
Evidence of an influential observation.
One input equals twice another throughout the sample:
Failure of full column rank.
See also. Regression Fit and Statistical Inference develops coefficient tests. Regression Misspecification and Remedies covers responses to variance and dependence problems. Regression Influence Diagnostics examines deletion measures. Logistic Regression Interpretation changes the outcome model for binary events. For dated investment data, related applications include Stationary Time-Series Forecasting, Unit Roots and Cointegration, and Seasonality and Conditional Variance. Machine Learning Model Selection extends the comparison of predictive performance.

A Read a one-period residual-lag plot

Pair each month's residual with the preceding month's residual. Put the earlier value on the horizontal axis and the current value on the vertical axis. An upward tilt means large values tend to follow large values and small values tend to follow small values. A downward tilt suggests negative first-order dependence. With 36 consecutive monthly observations there are 35 adjacent pairs. Preserve the calendar spacing: across a missing month, the preceding available residual is two or more months old.

Figure 5 shows how three different residual patterns appear when the axes are fitted value and residual, followed by a comparison using lagged and current residuals.

Fig 5. Residual centers, widths, and lag relationships
Key point: Read the horizontal axis before assigning an explanation to a pattern of residuals.
{"type":"normal_vs_abnormal","describe":"Across the top, a level band of residuals is followed by a band that widens and one whose center bends into a U. Below, the horizontal coordinate changes to last month's residual. An upward-sloping cloud pairs similar successive misses, while a downward-sloping cloud pairs opposite misses.","teaching_point":"Curvature, changing dispersion, and serial association leave distinct patterns on the appropriate axes.","prompt":"Comparative statistical residual plots, clean textbook illustration on white. Straight-on view in two rows. Top row: three equal panels with fitted value horizontal, residual vertical and a thin gray horizontal zero line; navy points form a constant-width level band, a symmetric widening fan centered on zero, and a constant-width U-shaped band respectively. Bottom row: two panels with centered gray axes, lagged residual horizontal and current residual vertical; navy points form a rising cloud and a falling cloud respectively. MARKERS: 5 numbered circular markers, each a white digit on a solid dark circle, each joined to its target by a short thin leader line: 1 on the level band; 2 on the fan; 3 on the U-shaped band; 4 on the rising cloud; 5 on the falling cloud. Use each digit exactly once, large and easy to read. NO words, captions or text labels anywhere in the image: digits 1 to 5 only. Emphasize the common center of the fan and the contrasting slopes below. Omit arrows, grid lines and decoration.","caption":"Fig 5. Residual centers, widths, and lag relationships","size":"1536x1024"}
  1. Homoskedasticity. A level band of stable width is consistent with constant conditional variance.
  2. Heteroskedasticity. A fan centered on zero suggests changing conditional variance.
  3. Functional form misspecification. A curved local mean suggests an omitted nonlinear relationship.
  4. Positive serial correlation. Earlier and later residuals tend to move together.
  5. Negative serial correlation. A higher earlier residual tends to accompany a lower current residual.

B Recall check

Supply the quantity or condition each comparison requires.

Which OLS property makes an estimation-sample mean residual of zero unsurprising?
Fitting an intercept with unweighted OLS forces the residuals to sum to zero.
How many sector indicators accompany an intercept for five exhaustive sectors?
Four. The sector with all four indicators at zero supplies the reference.
What additional input is needed to turn a fitted value into a signed residual?
The observed value for the same outcome and observation.
Where is an indicator main effect measured when the model includes its product with duration?
At duration zero, or at the centering value if the model uses centered duration.
Which ordering preserves evidence of serial dependence?
Observation time order, so adjacent residuals correspond to successive periods.
COMMON MISTAKE

Common mistakes

Table 7 pairs plausible readings of an exhibit with the detail that resolves each one.

Table 7. Distinguishing similar answers
Tempting answerWhy it appealsWhat settles the comparison
Add the intercept to a scenario changeThe constant appears in both fitted equationsThe same constant cancels when the two levels are subtracted
Report the new reported yield as the input changeThe latest value is prominently displayedA change requires the ending input minus its starting value
Interpret the indicator coefficient as every group's total fitted spreadThe coefficient row carries the sector nameThe intercept and applicable continuous terms still contribute to the level
Read a fan as increasing underpredictionLarge positive misses appear at high fitted valuesLarge negative misses also appear there; check the local center
Infer slope significance from its sizeThe economic effect looks largePrecision depends on the standard error in the same coefficient units
ALGO

Approach

Use this sequence when an exhibit combines a group comparison with a reported outcome.

Set the target. Write down the requested level, difference between two profiles, or actual-minus-fitted residual, with its outcome unit.
Encode each profile. Convert percentage inputs to the exhibit's convention and assign zero or one to every relevant indicator.
Build the group equation. Combine the intercept with active group adjustments and the continuous slope with active interaction coefficients.
Evaluate and subtract. For a comparison, subtract the starting fitted value from the ending one. For a residual, subtract the fitted level from the actual outcome.
Check the sign against the vignette. A positive residual requires an above-model observation. A positive group gap requires the named comparison group to have the higher fitted level.
QUIZ

Self-check quiz

A bank analyst estimates loan spreads in basis points. The input \(L\) is borrower leverage in turns, and \(D\) equals one for an unsecured loan and zero for a secured loan. The estimated equation is \(\hat S=70+18L+12D+5LD\). An unsecured borrower has leverage of three turns and an observed spread of 160 basis points. The analyst also examines residuals from the estimation sample.

Question 1

What is the borrower's residual?

  1. −9 basis points
  2. 9 basis points
  3. 24 basis points
  4. 36 basis points

Question 2

How much would the unsecured borrower's fitted spread change if leverage fell from three turns to two?

  1. Decrease by 18 basis points
  2. Decrease by 5 basis points
  3. Increase by 23 basis points
  4. Decrease by 23 basis points

Question 3

The analyst recodes the indicator so that unsecured loans are the reference, retaining the interaction with leverage. What are the new intercept and the new coefficient on the secured indicator?

  1. 82 and −12 basis points
  2. 70 and −12 basis points
  3. 82 and 12 basis points
  4. 97 and −27 basis points

Question 4

The analyst groups the estimation-sample residuals by fitted spread. Using Table 8, which feature should the analyst investigate first?

Table 8. Residuals grouped by fitted spread
Fitted spread groupMean residual (bp)Residual standard deviation (bp)Loans
Low1550
Middle−21350
High13150
  1. Positive serial correlation
  2. Perfect multicollinearity
  3. Heteroskedasticity
  4. Systematic underprediction throughout the fitted range

Question 5

A reviewer estimates a simple regression on leverage alone and obtains a slope of 26 basis points per turn. Which interpretation is best supported?

  1. The difference from 18 measures the causal effect of collateral.
  2. The pooled slope and the secured-loan slope describe different comparisons.
  3. The bank should use 26 as the unsecured slope in the original equation.
  4. The original model violates full column rank because its slopes differ from 26.
ANSWERS

Answer key

Question 1 - Answer: B. The fitted spread is \(70+18(3)+12+5(3)=151\), leaving \(160-151=9\) basis points. A reverses the subtraction, C drops the interaction, and D uses the secured-loan benchmark.

Question 2 - Answer: D. The unsecured slope is \(18+5=23\) basis points per turn and leverage falls by one turn. A uses the secured slope, B uses only the incremental slope, and C reverses the input change.

Question 3 - Answer: A. At zero leverage, unsecured loans start at 82 basis points and secured loans at 70, so switching to secured subtracts 12. B retains the old intercept, C retains the old indicator sign, and D incorporates the interaction at three turns into what should be a zero-leverage comparison.

Question 4 - Answer: C. Residual dispersion rises from 5 to 31 basis points while the group means remain close to zero. A requires evidence about ordering, B requires an exact relationship among predictor columns, and D would require consistently positive local means.

Question 5 - Answer: B. The simple slope summarizes the pooled association, whereas 18 measures leverage sensitivity within secured loans. A assigns causation to a specification change, C overlooks the estimated unsecured slope of 23, and D confuses different fitted relationships with exact dependence among predictors.

RAPID

Rapid review

  • One additive slope → fitted outcome units per predictor unit with other inputs fixed.
  • Percentage as an input → check whether 5% enters as 5, 0.05, or 500 basis points.
  • Two fitted scenarios → multiply each slope by its input change and include any changed interaction term.
  • Observed outcome above fitted → positive residual.
  • All sector indicators zero → reference-sector equation.
  • Indicator times continuous input → group gap varies with that input.
  • Curved residual center → investigate the conditional mean.
  • Widening residual band → investigate conditional variance.
  • Runs across dates → investigate serial dependence.
  • Unusual predictor profile → examine leverage and the effect of deleting the observation.
  • Missing intercept → start from the mean outcome and subtract each slope times its regressor-column mean.
  • Log-log coefficient → use elasticity for a small move and raise the input ratio to the slope for a large move.
  • Squared leverage → subtract the starting squared level from the ending squared level.
  • One future issuer → include individual disturbance variance when calculating prediction uncertainty.
  • Equal fitted group outcomes → set the entire indicator-plus-interaction gap equal to zero.

References

  • 2027 Level II Topic Outlines. CFA Institute, 2027 examination cycle.
  • CFA Program Curriculum, Level II: Quantitative Methods. CFA Institute, 2027 examination cycle.
End of your free lesson

Keep going with the full CFA Level II Exam Prep Course

You just studied the way members do, in the same lesson interface. Full access unlocks every lesson, the question bank and flashcards for a full year.

Up next with full access1.1.2 - Regression Fit and Statistical Inference

One payment per program