Develop the calculations and interpretations central to CAS MAS-II: credibility weighting, mixed-model covariance, statistical learning, and constant-variance time series. This guide pairs 20 original single-answer practice questions with 60 technical lessons, worked examples, and misconception checks. The question allocation is 4/3/9/4 and the concept allocation is 12/9/27/12 across the four domains. These are representative learning allocations, not published examination totals. Coverage follows the supplied 2026 outline for use as of October 16, 2026; announced 2027 scheduling does not replace that scope. Citations establish syllabus relevance; technical explanations draw on established subject knowledge. The actual assessment may use additional listed item formats.
What this guide covers
Build MAS-II skills with 20 original questions and 60 worked concepts across the 2026 syllabus.
A. Introduction to Credibility
1. EPV and VHM separate noise from heterogeneity
Let Θ identify a risk's persistent characteristics, μ(Θ) its conditional mean, and v(Θ) its conditional observation variance. Expected process variance is EPV = E[v(Θ)], while variance of hypothetical means is VHM = Var[μ(Θ)]. The total-variance identity gives Var(X) = EPV + VHM. For repeated observations conditionally independent given Θ, their unconditional covariance is VHM. This decomposition distinguishes fluctuations within a risk from genuine differences across risks. Credibility increases when persistent heterogeneity is large relative to temporary noise, provided the assumed conditional structure is appropriate.
Worked example: Suppose two equally common risk types have conditional means 80 and 120, and each has conditional variance 300. The collective mean is 100, VHM is 400, and EPV is 300. An individual observation therefore has unconditional variance 700.
Common mistake: A large overall sample variance does not establish high credibility; much of that variance may be temporary within-risk process noise.
Official syllabus reference: [2]
2. Limited-fluctuation frequency standards
For a Poisson count N with mean λ, a normal approximation converts P(|N − λ| ≤ rλ) ≥ p into λ ≥ (z/r)², where z is the standard-normal quantile for central probability p. The standard controls relative sampling fluctuation around an unknown underlying mean; it does not directly minimize prediction error. Exposure requirements follow by dividing the required expected count by an assumed frequency. The approximation is less dependable for small expected counts, and the probability level, error tolerance, and frequency assumption must be explicitly specified.
Worked example: An exercise requires central probability 0.95 and relative error 0.10, supplying z = 1.96. The required expected count is 384.16. At an assumed frequency of 0.08 claims per exposure, this corresponds to 4,802 exposure units before any required rounding.
Common mistake: The full-credibility count is not a universal actuarial threshold; it changes with the selected probability, error tolerance, and distributional assumptions.
Official syllabus reference: [2]
3. Severity credibility depends on squared coefficient of variation
For n independent claim severities with mean μ and variance σ², the sample mean has variance σ²/n. A normal-approximation relative-error requirement gives n ≥ z²σ²/(r²μ²) = (z/r)²CV². Thus severity volatility raises the number of claims required for full credibility. The formula concerns a fixed number of independent severities and a finite variance. It should not be applied unchanged to dependent claims or infinite-variance severity models. Estimating the coefficient of variation introduces additional practical uncertainty beyond the idealized standard itself.
Worked example: Two severity portfolios both average $2,000 per claim. Their coefficients of variation are 1 and 2. With identical probability and relative-error requirements, the second portfolio needs four times as many claims because its squared coefficient of variation is four times larger.
Common mistake: Equal mean severity does not imply equal credibility requirements; dispersion relative to the mean determines the sample size in this approximation.
Official syllabus reference: [2]
4. Compound-Poisson aggregate loss credibility
For aggregate loss S = ΣXi with Poisson claim count of mean λ, independent identically distributed severities, and independence between count and severities, E[S] = λμ and Var(S) = λE[X²]. A normal relative-error calculation yields λ ≥ (z/r)²E[X²]/μ² = (z/r)²(1 + CV²). Aggregate uncertainty contains both count fluctuation and severity fluctuation. This differs from credibility for average severity with a fixed claim count. The result depends on the Poisson frequency assumption; alternative count variances change the aggregate variance and the resulting standard.
Worked example: Suppose the supplied frequency-only standard is 400 expected claims and severity has coefficient of variation 1.5. Under the compound-Poisson assumptions, the aggregate-loss standard becomes 400(1 + 1.5²) = 1,300 expected claims, reflecting both sources of loss variation.
Common mistake: Using only severity variance omits frequency uncertainty; aggregate-loss credibility uses the severity second moment under the compound-Poisson model.
Official syllabus reference: [2]
5. Square-root partial credibility and the complement
A common limited-fluctuation partial-credibility rule is Z = min(1, sqrt(V/Vfull)), where V measures information volume consistently with the full-credibility standard. The estimate is ZX + (1 − Z)M, combining subject experience X with complement M. Square-root scaling reflects the inverse-square-root behavior of relative sampling error under the motivating assumptions. It is a prescribed partial-credibility convention rather than the Bühlmann least-squares solution. Counts, exposures, and premium cannot be substituted indiscriminately; the information measure must match the standard and the quantity being estimated.
Worked example: A severity standard requires 900 claims, while a segment supplies 225 comparable claims. Its credibility is one-half. If observed average severity is $1,400 and the complement is $1,000, the resulting credibility estimate is $1,200 per claim.
Common mistake: Applying the experience-volume ratio directly produces a different rule; square-root partial credibility requires taking its square root before weighting.
Official syllabus reference: [2]
6. Bühlmann credibility as an optimal linear predictor
With n exchangeable, conditionally independent observations per risk, Bühlmann credibility predicts the risk mean by Z Xbar + (1 − Z)m. Here m is the collective mean, K = EPV/VHM, and Z = n/(n + K). This predictor minimizes mean squared error within the relevant class of linear predictors under the model's moment assumptions. It does not require specifying the entire latent-risk distribution. Greater process variance increases K and shrinkage; greater between-risk variance decreases K. The result is linear credibility, which need not equal an unrestricted Bayesian posterior mean.
Worked example: A risk has five observations averaging 150, while the collective mean is 110. If EPV is 1,200 and VHM is 300, K equals 4 and Z equals 5/9. The prediction is 132.22, retaining substantial shrinkage despite above-average experience.
Common mistake: The number of observations alone does not determine credibility; the ratio of within-risk noise to between-risk variation also matters.
Official syllabus reference: [2]
7. Balanced-data estimation of Bühlmann parameters
For r independent risks with n comparable observations each, estimate EPV by pooling the within-risk sample variances. The variance of risk sample means estimates VHM + EPV/n, so subtract estimated EPV/n to estimate VHM. This correction prevents sampling noise in observed group means from being mistaken for persistent heterogeneity. The formulas assume a balanced, exchangeable structure with suitable independence. A negative estimated VHM can occur through sampling variation; constraining it to zero is a practical boundary treatment that should be identified rather than silently interpreted as negative true variance.
Worked example: Six risks each have four annual observations. The pooled within-risk variance is 800, and the sample variance of their annual averages is 350. Estimated VHM is 350 − 800/4 = 150, giving an estimated K of 800/150.
Common mistake: Using the raw variance of group averages as VHM overstates persistent differences because each average still contains process noise.
Official syllabus reference: [2]
8. Bühlmann–Straub exposure weighting
Bühlmann–Straub credibility allows observations with differing information weights. For a rate Xij and weight wij, assume E[Xij|Θi] = μ(Θi) and Var(Xij|Θi) = v(Θi)/wij. The risk's weighted mean is ΣwijXij/Wi, with total weight Wi = Σwij. Credibility becomes Zi = Wi/(Wi + K), where K = E[v(Θ)]/Var[μ(Θ)]. Exposure changes precision, not the conditional mean rate under this formulation. If larger exposures systematically correspond to different risk means, the simple structure requires reconsideration rather than automatic weighting.
Worked example: A risk has loss costs 90 and 130 over exposures 1 and 3. Its weighted mean is 120, not 110. With total exposure 4, K = 6, and collective mean 100, credibility is 0.4 and the predicted rate is 108.
Common mistake: Treating unequal-exposure annual rates as equally informative discards the inverse-exposure variance structure built into the Bühlmann–Straub model.
Official syllabus reference: [2]
9. Unbalanced exposure and structural variance correction
Under Bühlmann–Straub assumptions, let Wi be risk totals, W their sum, and Xbarw the overall exposure-weighted mean. For r risks, Q = ΣWi(Xbari − Xbarw)² has expectation (r − 1)v + a[W − ΣWi²/W], where v is EPV and a is VHM. Subtracting the process component and dividing by the bracket estimates a. This denominator corrects for unequal risk weights. Estimating v separately requires appropriate within-risk information; one weighted mean per risk generally cannot identify both components without additional assumptions.
Worked example: Three risks have exposure totals 1, 2, and 3, giving W = 6 and W − ΣWi²/W = 11/3. If Q = 42 and estimated v = 10, estimated a is (42 − 20)/(11/3) = 6.
Common mistake: An ordinary unweighted variance of risk averages does not supply the correct unbalanced-exposure correction for Bühlmann–Straub structural estimation.
Official syllabus reference: [2]
10. Normal–normal Bayesian credibility
Suppose a latent mean Θ has prior N(m, τ²), and n observations are conditionally independent N(Θ, σ²). The posterior mean is Z Xbar + (1 − Z)m, where Z = nτ²/(σ² + nτ²). Posterior variance is (1/τ² + n/σ²)⁻¹. This is an exact Bayesian posterior calculation and matches Bühlmann credibility for this model. Predicting a new observation adds σ² to posterior uncertainty about Θ. Distinguish uncertainty about the persistent mean from the larger uncertainty of an individual future outcome.
Worked example: A prior mean is 100 with variance 25. Four observations average 120, and each has conditional variance 100. Posterior mean is 110 and posterior variance is 12.5. A new observation has predictive variance 112.5 because its own noise remains.
Common mistake: Posterior variance of the latent mean is not the predictive variance of a future observation; future process noise must also be included.
Official syllabus reference: [2]
11. Beta–binomial updating for event probabilities
For a persistent event probability P with Beta(α, β) prior and binomial observations containing x events in n trials, the posterior is Beta(α + x, β + n − x). Its mean is (α + x)/(α + β + n), a weighted average of x/n and α/(α + β). The experience weight is n/(n + α + β). Conditional independence applies given P; future trials become dependent after averaging over shared probability uncertainty. For multiple future trials, the predictive distribution is beta-binomial rather than an ordinary binomial with known probability.
Worked example: A Beta(2, 8) prior represents a mean event probability of 0.20. Observing six events in twenty trials gives Beta(8, 22) and posterior mean 8/30. Ten future trials therefore have expected event count 10(8/30), approximately 2.67.
Common mistake: Replacing the posterior predictive distribution by a binomial with the posterior mean preserves its expectation but understates uncertainty from the shared unknown probability.
Official syllabus reference: [2]
12. Gamma–Poisson rates and predictive overdispersion
Using shape–rate notation, a Gamma(α, β) prior for Poisson rate Θ updates to Gamma(α + N, β + E) after N claims over exposure E. The posterior mean rate is (α + N)/(β + E). For future exposure e, predictive mean is e times that rate, and predictive variance is eE[Θ|data] + e²Var(Θ|data). The additional term reflects parameter uncertainty and produces a negative-binomial predictive distribution. Rate and scale parameterizations must be distinguished because their numerical updating formulas differ. Exposure is added to the rate parameter, not to the shape.
Worked example: After updating, suppose Θ has Gamma shape 12 and rate 4. Its mean is 3 and variance is 0.75. Over two future exposure units, expected claims are 6 and predictive variance is 6 + 4(0.75) = 9.
Common mistake: A Poisson predictive variance equal to its mean omits posterior uncertainty in the latent rate and understates the mixed-model predictive variance.
Official syllabus reference: [2]
B. Linear Mixed Models
13. The mixed-model matrices X, Z, G, and R
A linear mixed model is y = Xβ + Zb + ε, with E[b] = 0, E[ε] = 0, Var(b) = G, Var(ε) = R, and independence between b and ε. Conditional on random effects, the mean is Xβ + Zb; marginally it is Xβ. The marginal covariance is V = ZGZ′ + R. X specifies population effects, Z maps group-specific deviations to observations, and G and R describe separate covariance sources. Gaussian assumptions support the usual likelihood machinery, while the matrix structure explains how correlated observations arise.
Worked example: For two observations from one group, a random-intercept Z is a column of two ones. If G = [4] and R = 9I, the marginal covariance matrix has diagonal entries 13 and off-diagonal entries 4, despite conditionally independent residuals.
Common mistake: A non-diagonal marginal covariance matrix does not imply correlated conditional residuals; shared random effects can generate all observed within-group dependence.
Official syllabus reference: [2]
14. Random intercepts impose compound symmetry
With a random intercept variance τ² and independent equal-variance residuals σ², every observation has marginal variance τ² + σ² and every distinct within-group pair has covariance τ². Their common correlation is τ²/(τ² + σ²), producing compound symmetry. This structure treats nearby and distant measurements as equally correlated within a group. It is useful when a persistent group-level shift plausibly explains dependence, but it cannot represent correlation that decays with time separation unless additional residual or random-effect structure is introduced. Correlations across independent groups remain zero.
Worked example: Repeated observations have random-intercept variance 6 and residual variance 14. Their intraclass correlation is 0.30, whether separated by one month or twelve months. If empirical residual dependence decays sharply with lag, a random intercept alone may be inadequate.
Common mistake: A random intercept does not automatically produce stronger correlation for nearby measurements; its shared contribution is identical across all within-group pairs.
Official syllabus reference: [2]
15. Random slopes create time-dependent covariance
For random intercept b0 and slope b1, covariance between random contributions at times s and t is g00 + (s + t)g01 + stg11. Marginal variance at time t is g00 + 2tg01 + t²g11 plus residual variance. Thus variability and correlation can change across measurement times even when residual variance is constant. Centering time changes the interpretation and covariance of random intercepts and slopes while preserving fitted trajectories under an equivalent transformation. The random-effect covariance matrix must remain positive semidefinite for a valid model.
Worked example: Let g00 = 4, g01 = −1, and g11 = 2. The random trajectory variance at time zero is 4, while at time two it is 4 − 4 + 8 = 8. Residual variance would be added separately.
Common mistake: The random-slope variance is multiplied by squared time in an observation's variance; treating it as a constant misses changing trajectory dispersion.
Official syllabus reference: [2]
16. Residual covariance: AR(1), heterogeneity, and unstructured forms
Residual covariance R describes remaining dependence after fixed and random effects. For equally spaced observations, an AR(1) residual structure uses σ²ρ^|t−s|, making correlation decline geometrically with lag. Heterogeneous versions allow time-specific residual variances, while an unstructured matrix estimates each variance and covariance separately. Greater flexibility consumes parameters and can destabilize estimation with limited groups. Random effects and residual structures can both explain dependence, so their roles must be assessed jointly. A residual model should match measurement spacing and design rather than being selected solely for having the most parameters.
Worked example: With residual variance 25 and AR(1) correlation 0.6, observations one interval apart have residual covariance 15; two intervals apart have covariance 9. Adding a random intercept contributes an additional shared covariance to both pairs.
Common mistake: An unstructured residual matrix is not automatically preferable; limited replication can make its many covariance estimates unstable or poorly identified.
Official syllabus reference: [2]
17. Nested and crossed grouping structures
Grouping factors are nested when each level of one factor belongs to exactly one level of another, such as branches within regions. Factors are crossed when their combinations overlap, such as adjusters handling claims from multiple branches. Independent random effects for crossed factors create covariance whenever observations share either factor. A nested model creates a different sharing pattern and therefore a different marginal covariance matrix. Determine grouping from the actual observational design before interpreting variance components. Group labels must uniquely identify the intended units, especially when local labels repeat within higher-level groups.
Worked example: Two claims handled by the same adjuster in different branches share an adjuster effect under a crossed model. They need not share a branch effect. Treating adjusters as nested within branches would incorrectly remove their shared adjuster contribution across branches.
Common mistake: Repeated use of a factor name does not establish nesting; nesting requires each lower-level unit to belong to only one higher-level unit.
Official syllabus reference: [2]
18. BLUP shrinkage depends on group information
For known covariance parameters, the best linear unbiased predictor of random effects is GZ′V⁻¹(y − Xβ), with estimated fixed effects substituted in practice. In a simple random-intercept model, the group mean deviation is multiplied by nτ²/(σ² + nτ²). Small or noisy groups shrink more strongly toward the population trajectory; large, precise groups shrink less. Estimated covariance parameters produce empirical BLUPs rather than predictors using known parameters. Shrinkage concerns group-specific deviations and should not be confused with automatically shrinking every fixed-effect coefficient toward zero.
Worked example: A group's residual mean is 12 above the fixed-effect prediction. With n = 3, τ² = 4, and σ² = 24, the shrinkage factor is 12/36 = 1/3. Its predicted random intercept is therefore 4.
Common mistake: A predicted random effect smaller than the raw group deviation is an intended precision adjustment, not evidence that the model ignored that group.
Official syllabus reference: [2]
19. ML and REML answer different estimation needs
Maximum likelihood estimates fixed effects and covariance parameters through the joint Gaussian likelihood. Restricted maximum likelihood estimates covariance parameters using combinations of observations that remove fixed effects, accounting for degrees of freedom used in estimating them. REML often reduces small-sample variance-component bias. For likelihood-based comparison of different fixed-effect designs, use ML because restricted likelihoods depend on the design being removed. REML comparisons of covariance structures require the same fixed-effect design and suitable likelihood conventions. Refit a selected fixed-effect specification by REML when appropriate for final covariance estimation.
Worked example: An analyst compares models with and without a treatment-by-time interaction while retaining the same random effects. ML likelihoods or ML AIC support that comparison. After choosing the fixed effects, the analyst can use a REML fit to report covariance estimates.
Common mistake: Comparing the smallest REML AIC across different fixed-effect designs does not provide the usual valid likelihood-based fixed-effect selection.
Official syllabus reference: [2]
20. Variance-component tests have boundary complications
A fixed-effect null such as βtreatment = 0 typically lies inside the parameter space, but a null variance τ² = 0 lies on its boundary because variances cannot be negative. Ordinary likelihood-ratio chi-square reference distributions may therefore be inappropriate for testing random effects. In simple settings, mixture reference distributions arise; more complex covariance hypotheses require their own justification or simulation methods. Fixed-effect inference also depends on how denominator degrees of freedom are approximated. Interpret reported tests according to the tested parameter, estimation method, and reference distribution rather than treating all p-values identically.
Worked example: Testing whether a random intercept is needed compares τ² = 0 with τ² > 0. A likelihood-ratio statistic of 2 cannot automatically be assigned the ordinary one-degree-of-freedom chi-square p-value without checking the applicable boundary-test assumptions.
Common mistake: A variance-component test is not an ordinary unrestricted coefficient test; its null value lies at the edge of the allowed parameter space.
Official syllabus reference: [2]
21. Conditional residuals diagnose different features than marginal residuals
Marginal residuals subtract Xβhat, while conditional residuals additionally subtract predicted random effects. Conditional residual plots help investigate the remaining R structure, including unequal variance, residual autocorrelation, and outliers. Marginal residual patterns can reflect legitimate group effects rather than model failure. Normal quantile plots of residuals and predicted random effects address different distributional assumptions, and predicted random effects are themselves shrunken estimates. Diagnostics should distinguish mean misspecification from covariance misspecification. A time pattern in residual means suggests missing systematic structure, whereas lag dependence around zero suggests remaining correlation.
Worked example: Marginal residuals cluster by branch, but conditional residuals no longer do after fitting random intercepts. Conditional residual spread still increases with time. The group clustering is explained, while the changing spread suggests considering a heterogeneous residual variance structure.
Common mistake: Group patterns in marginal residuals are not automatically defects when the model explicitly assigns those patterns to random effects.
Official syllabus reference: [2]
C. Statistical Learning
22. Tree depth balances squared bias against prediction variance
For squared-error prediction at a fixed predictor value, expected test error decomposes into irreducible noise variance, squared bias, and variance of the fitted prediction across training samples. Increasing tree depth permits finer partitions, potentially reducing approximation bias. It also leaves fewer observations in each leaf and can make predictions more sensitive to the training sample. A deeper tree improves expected prediction only when its reduction in squared bias exceeds its increase in variance. Validation estimates the resulting predictive performance when these population components are unknown.
Worked example: At one predictor value, a shallow tree has squared bias 9 and prediction variance 3; a deeper tree has squared bias 4 and variance 10. With noise variance 4, expected test errors are 16 and 18. The shallow tree is preferable at that point.
Common mistake: Lower training error does not establish lower expected test error; the reduction in squared bias may be outweighed by greater prediction variance.
Official syllabus reference: [2]
23. Validation must reproduce the full fitting pipeline
Validation estimates performance of a procedure, including preprocessing and tuning, rather than only a final fitted equation. Scaling, imputation, PCA, feature selection, and hyperparameter selection must be learned within the training portion of each split when they use sample information. Otherwise, validation observations influence the model being evaluated and produce optimistic results. Grouped or temporally ordered data require splits matching the intended prediction setting. After choosing settings with validation, an untouched test set provides a separate assessment. Repeatedly selecting models against that test set turns it into another tuning resource.
Worked example: When evaluating KNN, calculate predictor means and standard deviations separately inside each training fold, then apply them to its validation fold. If all years from one policy share information, keep that policy together when predicting entirely new policies.
Common mistake: An unsupervised preprocessing step can still leak information when it is fitted using validation observations before evaluating the full prediction procedure.
Official syllabus reference: [2]
24. Predictor scaling changes KNN neighborhoods
K-nearest neighbors uses predictor distances to select the observations contributing to a prediction. Under Euclidean distance, each squared coordinate difference contributes directly, so changing a predictor's scale changes its influence on neighborhood membership. For unweighted binary classification, the estimated positive probability is the fraction of positive observations among the selected neighbors. Scaling can therefore change both that probability and the assigned class without changing K or any labels. Suitable scaling depends on the intended meaning of similarity; standardization gives predictors comparable scale but does not establish equal predictive relevance.
Worked example: For query (0,0), use positives (6,0),(8,0) and negatives (0,2),(0,3),(0,10). With K=3, raw distances are 6,8,2,3,10, giving positive probability 1/3. Dividing the first coordinate by 10 gives distances 0.6,0.8,2,3,10. The nearest three now contain both positives, so probability becomes 2/3.
Common mistake: Changing measurement units can change Euclidean KNN predictions unless the distance calculation applies a compensating transformation.
Official syllabus reference: [2]
25. Gini impurity evaluates class mixing within nodes
For class proportions pk in a node, Gini impurity is Σpk(1 − pk) = 1 − Σpk². It is zero for a pure node and larger when classes are mixed. A split is evaluated through the observation-weighted average impurity of its child nodes, not an unweighted sum unless child sizes match. The improvement is parent impurity minus that weighted child value. This tree-building Gini measure concerns local class composition and is distinct from the normalized Gini associated with ROC ranking or the concentration Gini used for insurance prediction.
Worked example: A parent contains equal class proportions, so its Gini is 0.50. A split creates equally sized children with positive proportions 0.80 and 0.20. Each child's Gini is 0.32, making weighted child impurity 0.32 and improvement 0.18.
Common mistake: Averaging child impurities without their observation weights can favor a split that isolates a tiny pure node while leaving most observations mixed.
Official syllabus reference: [2]
26. Entropy and information gain
Classification entropy is H = −Σpk log(pk), using a consistent logarithm base and defining 0 log 0 as zero. A pure node has zero entropy, while uncertainty grows as class probabilities become more balanced. Information gain subtracts the size-weighted child entropy from parent entropy. Changing the logarithm base rescales all gains consistently without changing split rankings calculated on the same data. Entropy and Gini often select similar splits but need not agree. Misclassification error is less sensitive to probability changes and is generally less effective for growing classification trees.
Worked example: Using base-two logarithms, a balanced binary parent has entropy one bit. A split into equally sized children with positive proportions 0.90 and 0.10 gives each entropy about 0.469 bits. The split's information gain is therefore approximately 0.531 bits.
Common mistake: A node can become less uncertain even when its majority class stays unchanged; classification error alone may miss that probability improvement.
Official syllabus reference: [2]
27. Regression-tree splits minimize within-leaf RSS
A regression tree predicts a constant within each terminal region. Under squared-error loss, the optimal constant is the region's sample mean, and its contribution to RSS is the sum of squared deviations from that mean. Candidate splits are compared by adding the RSS values of their two children. Greedy splitting chooses the best immediate improvement rather than globally searching every possible tree. Minimum leaf sizes and later pruning control complexity. A split that separates predictor values is useful only to the extent that it improves the modeled response distribution under the selected loss.
Worked example: A parent contains responses 2, 4, 8, and 10, with mean 6 and RSS 40. A split grouping 2 with 4 and 8 with 10 gives child means 3 and 9. Total child RSS is 4, reducing RSS by 36.
Common mistake: Child predictions under squared-error loss are their response means, not predictor midpoints or the midpoint between the two child means.
Official syllabus reference: [2]
28. Cost-complexity pruning selects a subtree sequence
Cost-complexity pruning balances fit and size through a criterion such as RSS(T) + α|T| for regression trees. Here |T| counts terminal nodes and α controls the penalty per leaf. As α increases, smaller subtrees become preferable. A large initial tree supports a nested candidate sequence, and validation or cross-validation selects the tuning level. The penalty's units and numerical scale depend on whether the loss is a sum, an average, or another defined quantity. Never transfer an α unchanged between differently normalized criteria without accounting for that definition.
Worked example: Two candidate subtrees have RSS values 60 and 75 with five and two leaves. Their criteria are equal when 60 + 5α = 75 + 2α, giving α = 5. Above that value, the two-leaf subtree is preferred.
Common mistake: Pruning does not mean selecting the tree with smallest training RSS; that criterion alone systematically favors unnecessary additional leaves.
Official syllabus reference: [2]
29. Bagging and out-of-bag prediction
Bagging fits models to bootstrap samples and averages regression predictions or combines classification votes. It is especially useful for unstable learners such as deep trees because averaging reduces variance. A bootstrap sample of size n drawn with replacement omits approximately e⁻¹ of training observations when n is large. For each observation, out-of-bag prediction uses only trees whose bootstrap samples excluded it. These predictions support internal performance assessment without a separate validation split, although repeated tuning against out-of-bag error still involves selection. Bagging does not inherently eliminate bias or guarantee calibrated probabilities.
Worked example: A training observation appears in some bootstrap samples but is omitted from 120 fitted trees. Its out-of-bag regression prediction averages only those 120 trees. Averaging predictions from every tree would mix genuine out-of-bag predictions with predictions from models trained on that observation.
Common mistake: An observation is not out-of-bag for the entire ensemble; its eligible prediction trees depend on which bootstrap samples omitted it.
Official syllabus reference: [2]
30. Random forests reduce correlation among trees
Random forests extend bagging by restricting the predictor candidates considered at each split. This prevents a dominant predictor from forcing nearly identical upper splits in every tree and can reduce correlation among tree predictions. For exchangeable tree predictions, ensemble variance is σ²[ρ + (1 − ρ)/B]. Increasing tree count B reduces the uncorrelated component, while reducing correlation ρ lowers the limiting variance. Restricting predictors too aggressively can weaken individual trees, so the candidate-predictor count is a tuning choice. Random forests do not require predictors themselves to be statistically independent.
Worked example: If each tree has variance 64 and pairwise correlation 0.25, an ensemble of 16 trees has variance 64[0.25 + 0.75/16] = 19. Increasing tree count cannot reduce variance below 16 unless correlation or individual-tree variance also changes.
Common mistake: Random predictor selection does not assume independent predictors; its purpose is to diversify fitted trees and manage the strength–correlation trade-off.
Official syllabus reference: [2]
31. Boosting builds an additive correction sequence
Boosting constructs a model sequentially, adding weak learners that target deficiencies in the current prediction. Under squared-error gradient boosting, each new tree fits current residuals and enters with a learning-rate multiplier. Tree depth controls interaction complexity, while learning rate and iteration count jointly control how gradually the model adapts. Unlike bagging, the learners are not fitted independently to provide interchangeable averages. Excessive iterations can overfit, so stopping is selected using validation. Other losses replace raw residuals with the appropriate negative-gradient or related working-response calculations.
Worked example: Current predictions for three observations are 10, 20, and 30, while responses are 14, 18, and 35. Residuals are 4, −2, and 5. A fitted residual tree predicts 3 for the first observation; learning rate 0.1 updates its prediction to 10.3.
Common mistake: Boosting does not replace the current model with the newest tree; each learner contributes an incremental correction to the accumulated prediction.
Official syllabus reference: [2]
32. PCA loadings are covariance eigenvectors
For centered predictors with covariance matrix S, the first principal-component loading vector maximizes a′Sa subject to a′a = 1. It is the eigenvector associated with the largest eigenvalue. Subsequent loading vectors maximize remaining variance subject to orthogonality to earlier vectors. Each eigenvalue equals its component's variance under the chosen covariance convention. The loading-vector sign is arbitrary: reversing all its entries reverses scores without changing explained variance or reconstruction. PCA uses predictor variation rather than outcome information, so a high-variance direction is not automatically the most predictive direction for a response.
Worked example: For covariance matrix with diagonal entries 5 and off-diagonal entries 3, the leading unit loading is (1, 1)/sqrt(2), with eigenvalue 8. The orthogonal contrast (1, −1)/sqrt(2) has eigenvalue 2, capturing variation in the difference between predictors.
Common mistake: A loading vector and its negative represent the same principal-component axis; opposite reported signs do not establish conflicting PCA results.
Official syllabus reference: [2]
33. PCA scores and low-rank reconstruction
A component score projects a centered observation onto a unit loading vector: zik = Σj(xij − xbarj)ajk. Retaining K orthonormal components reconstructs the centered observation as Σk zikak, after which the original means are added back. If variables were standardized, reconstruction must also undo their scale transformation. PCA's truncated reconstruction minimizes squared reconstruction error among rank-K linear approximations under the chosen data geometry. Scores locate observations along axes, while loadings describe those axes in predictor space. Confusing the two can reverse the interpretation of software output.
Worked example: An observation centered at (3, 4) has score 5 on loading (0.6, 0.8). Keeping only that component reconstructs it as 5(0.6, 0.8) = (3, 4). Another point not lying on that axis would retain a nonzero reconstruction residual.
Common mistake: A score belongs to an observation, while a loading belongs to a predictor direction; they are not interchangeable entries in PCA output.
Official syllabus reference: [2]
34. Standardization changes the variance objective of PCA
Covariance PCA finds directions maximizing variance in the original centered predictor units. A predictor with a large marginal variance can therefore dominate the first component. Correlation PCA first divides each centered predictor by its standard deviation, giving every predictor unit marginal variance. It maximizes variance in these standardized coordinates, which can change both component directions and explained-variance proportions. Loadings from the two analyses refer to different coordinate systems. Choose between them according to whether original measurement scales should determine relative influence, rather than assuming standardization is universally preferable.
Worked example: For covariance matrix [[9,3],[3,4]], eigenvalues are (13±√61)/2. Covariance PCA gives first loading approximately (0.906,0.424), explaining 80.0%. Standardization yields correlation matrix [[1,0.5],[0.5,1]]. Its first loading, in standardized coordinates, is (0.707,0.707), explaining 1.5/2 = 75%.
Common mistake: Equal loadings after standardization mean equal weights per standard deviation, rather than equal coefficients on the original centered measurements.
Official syllabus reference: [2]
35. Explained variance uses squared singular values
The proportion of variance explained by component k is λk/Σλj. If software reports component standard deviations, square them to obtain eigenvalues before calculating proportions. For a centered data matrix with singular values dk, covariance eigenvalues are dk²/(n − 1) under the usual sample covariance convention. Cumulative explained variance sums component proportions in descending eigenvalue order. A variance target selects a representation size, but it does not guarantee preservation of response-predictive information or meaningful clusters. Scree plots and reconstruction goals provide additional context without changing the calculation.
Worked example: Software reports component standard deviations 3, 2, and 1. The variances are 9, 4, and 1, totaling 14. Retaining the first two explains 13/14, approximately 92.9%, rather than the 5/6 obtained by incorrectly adding standard deviations.
Common mistake: Component standard deviations must be squared before forming explained-variance proportions; summing them does not measure the retained variance.
Official syllabus reference: [2]
36. K-means alternates assignments and centroid updates
K-means minimizes within-cluster squared Euclidean distances to cluster centroids. Its standard algorithm alternates assigning each observation to its nearest centroid and recomputing each centroid as the mean of assigned observations. Each step cannot increase the objective, but convergence may reach a local rather than global minimum. Multiple initializations reduce dependence on starting centers. Scaling affects the objective because squared distances weight large-scale coordinates heavily. K must be chosen separately; raw within-cluster sum of squares decreases as K increases and cannot alone justify ever more clusters.
Worked example: With one-dimensional observations 1, 2, 8, and 9, initial centers 1 and 8 assign the first pair together and the second pair together. Updating produces centers 1.5 and 8.5, with total within-cluster squared distance equal to 1.
Common mistake: K-means centroids are coordinate-wise means, not necessarily actual observations, and a converged solution need not be the global optimum.
Official syllabus reference: [2]
37. Linkage rules define cluster-to-cluster dissimilarity
Agglomerative hierarchical clustering repeatedly merges the closest clusters under a linkage rule. Single linkage uses the smallest cross-cluster dissimilarity, complete linkage the largest, and average linkage the mean of all cross-cluster dissimilarities. Single linkage can form long chains through nearby intermediate observations, while complete linkage tends to favor tighter groups. These are different definitions, not alternative approximations to the same quantity. Ward's method instead assesses the increase in within-cluster sum of squares under its appropriate Euclidean formulation. Always identify the linkage before interpreting merge heights or comparing dendrograms.
Worked example: Two clusters have cross-distances 1, 3, 6, and 10. Their single-, average-, and complete-linkage distances are 1, 5, and 10. A nearby bridge point can therefore trigger an early single-linkage merge despite substantial separation elsewhere.
Common mistake: The distance between cluster centroids is not the general definition of average linkage, which averages all cross-cluster pairwise dissimilarities.
Official syllabus reference: [2]
38. Dendrogram cuts and dissimilarity choices
A dendrogram records successive merges, with heights representing the selected linkage dissimilarities. Cutting at a height partitions observations according to merges completed below that cut. Horizontal leaf order is largely arbitrary and should not be read as a numerical distance scale. The chosen observation-level dissimilarity and preprocessing can change the hierarchy before linkage is even applied. Euclidean distance emphasizes magnitude, while correlation-based dissimilarity can emphasize profile shape. A hierarchy provides nested candidate partitions; it does not by itself identify a uniquely correct number of substantive groups.
Worked example: Observations A and B merge at height 2, C and D at height 3, and the two pairs merge at height 8. A horizontal cut at height 5 yields two clusters, regardless of whether the plotted leaves appear as A-B-C-D or B-A-D-C.
Common mistake: Adjacent leaves are not necessarily the closest observations; merge heights and branch structure carry the relevant hierarchical information.
Official syllabus reference: [2]
39. Neural-network activations create nonlinear features
A feedforward hidden unit applies an activation function to an affine combination of inputs. ReLU uses max(0, z), while sigmoid maps a logit to a number between zero and one. Multiple nonlinear hidden units allow complex decision boundaries and interactions. Without nonlinear activations, stacked affine layers collapse to a single affine transformation, regardless of depth. Output activation and loss must match the response task: linear outputs commonly accompany squared-error regression, and sigmoid outputs commonly accompany binary cross-entropy. Bias terms shift activation thresholds and must be included during forward calculations.
Worked example: For inputs (2, −1), weights (1, 2), and bias 1, a hidden preactivation is 2 − 2 + 1 = 1. ReLU outputs 1. Removing the bias would instead place this observation exactly at the activation boundary.
Common mistake: Adding many linear layers without nonlinear activations does not create a more expressive nonlinear predictor; their transformations combine into one affine map.
Official syllabus reference: [2]
40. Backpropagation uses the chain rule
Backpropagation calculates loss derivatives through a network by applying the chain rule from output toward earlier layers. For a sigmoid output with binary cross-entropy, the derivative with respect to its logit simplifies to p − y. A weight gradient multiplies the incoming activation by the relevant downstream error signal. Gradient descent subtracts learning rate times gradient. Hidden-unit derivatives also include the activation derivative; an inactive ReLU has zero derivative away from its boundary. Backpropagation computes gradients, while the optimizer determines how those gradients update parameters.
Worked example: A sigmoid output predicts p = 0.8 for y = 1 and receives hidden activation h = 2. Its output-weight gradient is (0.8 − 1)2 = −0.4. With learning rate 0.1, gradient descent increases that weight by 0.04.
Common mistake: A negative gradient does not mean a weight should decrease; gradient descent subtracts the gradient, so a negative value increases the weight.
Official syllabus reference: [2]
41. Weight decay, dropout, and early stopping
Neural-network regularization controls flexibility through different mechanisms. Weight decay penalizes large weights, commonly using an L2 term. Dropout randomly removes selected activations during training, discouraging reliance on particular hidden units. Under inverted dropout, retained activations are divided by their retention probability during training, with no corresponding scaling required at inference. Early stopping chooses an iteration based on validation performance rather than minimizing training loss indefinitely. These methods can be combined but are not identical. Their strengths should be tuned using a validation procedure that remains separate from final performance assessment.
Worked example: An activation of 6 is retained with probability 0.75. In inverted dropout it becomes 8 when retained and zero otherwise, preserving expected activation 6. During inference, the full network uses the original activation without this random masking.
Common mistake: Dropout is generally a training regularizer; randomly discarding units during ordinary deterministic inference is not the standard prediction procedure.
Official syllabus reference: [2]
42. Convolution shares local feature detectors
A convolutional layer applies the same small collection of weights across input locations, exploiting local structure and reducing parameter count relative to a fully connected layer. Shared filters detect patterns wherever they occur, while multiple channels learn different features. Stride, padding, and filter width determine output dimensions. Pooling can reduce spatial resolution and sensitivity to small shifts, although it may discard location detail. Weight sharing is useful when neighboring positions have a meaningful relationship, as in images or ordered signals. It is not automatically appropriate for arbitrarily ordered tabular predictors.
Worked example: A one-dimensional input of length 10, filter width 3, stride 1, and no padding produces length 8. One filter uses three weights plus a bias across every position, rather than fitting a separate three-weight detector at each location.
Common mistake: A convolutional filter reuses weights across positions; assigning unrelated weights at each position removes the defining parameter-sharing mechanism.
Official syllabus reference: [2]
43. Recurrent hidden states retain information about sequence order
A recurrent neural network updates a hidden state using both the current input and the preceding state, with parameters shared across time steps. Earlier observations can therefore influence later predictions through successive state updates. In a simple linear recurrence, that influence can be traced through powers of the recurrent coefficient; nonlinear activations allow more flexible sequence representations. Reordering inputs generally changes the intermediate states and final output, even when their total and average remain unchanged. Interpreting a recurrent prediction requires knowing the initial state, update rule, input order, and output rule.
Worked example: Let h0=0, ht=0.5ht−1+xt, and predict with h3. Monthly counts (2,0,4) produce states 2,1,4.5 and prediction 4.5. Reversing them to (4,0,2) produces states 4,2,3 and prediction 3. Both sequences total six, but their final predictions differ.
Common mistake: A recurrent model does not generally reduce a sequence to its average; the state update can assign different influence to different positions.
Official syllabus reference: [2]
44. Confusion-matrix metrics condition on different populations
Precision is TP/(TP + FP), recall or sensitivity is TP/(TP + FN), and specificity is TN/(TN + FP). Accuracy is (TP + TN)/N. These denominators answer different questions: predicted positives, actual positives, actual negatives, and all observations. Class prevalence can make accuracy misleading when rare outcomes matter. A classification threshold changes the confusion matrix, so metrics should be interpreted at the stated threshold and on suitable evaluation data. F1 is the harmonic mean of precision and recall; it does not use true negatives and does not encode every possible decision cost.
Worked example: A model detects 18 of 30 actual positive cases and makes 24 positive predictions. Recall is 18/30 = 0.60, while precision is 18/24 = 0.75. The six false positives affect precision, while the twelve missed positives affect recall.
Common mistake: Precision and recall cannot be exchanged because one conditions on model predictions and the other on the observed positive class.
Official syllabus reference: [2]
45. AUROC measures ranking across class thresholds
The ROC curve plots true-positive rate against false-positive rate as the score threshold changes. AUROC equals the probability that a randomly selected positive observation receives a higher score than a randomly selected negative, with half credit for ties under the usual empirical definition. It measures discrimination rather than probability calibration. Strictly increasing transformations preserve rankings and therefore AUROC. A high AUROC does not guarantee high precision at a chosen threshold, especially when positives are rare. Evaluation still requires representative, independent data and a clearly defined positive class.
Worked example: Positive scores are 0.9 and 0.4; negative scores are 0.7 and 0.2. Three of the four positive–negative pairs are correctly ordered, so AUROC is 0.75. Replacing every score by its cube preserves the same ranking and AUROC.
Common mistake: AUROC does not verify that a score of 0.8 represents an 80% event probability; that is a calibration question.
Official syllabus reference: [2]
46. Lift must specify the denominator and population
Lift compares the outcome rate or loss cost in a selected score band with a defined baseline. For binary targeting, top-band lift commonly equals the band's event rate divided by the overall event rate. Cumulative gains instead measures the fraction of all events captured within a cumulative fraction of observations. In insurance applications, exposure-weighted loss costs may replace unweighted event rates, so the metric definition must be explicit. Lift depends on ranking direction and the chosen band size. It should be calculated on independent evaluation data rather than on observations used to fit the model.
Worked example: A test set contains 1,000 observations and 100 events. The highest-scored 100 observations contain 30 events. Their event rate is 30%, compared with 10% overall, so top-decile lift is 3 and cumulative event capture is 30%.
Common mistake: Capturing 30% of events is not itself lift three; the selected population share or baseline rate is needed to compute lift.
Official syllabus reference: [2]
47. Ranking Gini and concentration Gini require stated conventions
For binary classification with the standard ROC-based convention, Gini = 2AUROC − 1. Insurance predictive analysis also uses concentration curves that order exposures by predicted risk and accumulate observed losses. Their areas, direction, weighting, and normalization must be defined before calculating a Gini statistic. A normalized concentration Gini may divide a model's Gini by a corresponding outcome-ordered benchmark. These measures are distinct from classification-tree node impurity. Because several metrics share the name Gini, identify the mathematical definition and ordering convention instead of transferring a formula solely from the label.
Worked example: If a binary classifier's AUROC is 0.82, its ROC-based Gini is 0.64. That calculation does not automatically determine an exposure-weighted loss-concentration Gini for the same portfolio, because the latter needs the relevant ordering, losses, weights, and area convention.
Common mistake: The identity Gini = 2AUROC − 1 should not be applied indiscriminately to tree impurity or every insurance concentration statistic.
Official syllabus reference: [2]
48. Double-lift comparisons expose model disagreement
A double-lift analysis often groups observations by the ratio of two models' predictions, then compares observed and predicted outcomes across those disagreement bands. This focuses attention on where models assign different relative risks, which can be obscured by separate single-model lift charts. Exposure-weighted averages are appropriate when evaluating loss costs with unequal exposure. Overall calibration and relative segmentation are separate: multiplying all predictions by one constant changes levels but not band relativities. Use holdout observations and inspect band credibility, since noisy small bands can make apparent model advantages unstable.
Worked example: In a low A/B-ratio band, observed loss cost is 70; in a high-ratio band it is 140. Model A predicts a twofold increase while B predicts only 1.1-fold. This supports A's relative differentiation on those holdout bands, subject to their uncertainty.
Common mistake: A favorable double-lift pattern does not establish universal superiority; it evaluates the supplied disagreement bands and does not replace all other accuracy checks.
Official syllabus reference: [2]
D. Time Series with Constant Variance
49. Weak stationarity constrains means and covariances
A weakly stationary series has a constant finite mean, constant finite variance, and covariance that depends only on lag rather than calendar time. These properties justify using common autocovariances across the observed history. Constant variance alone is insufficient: a deterministic trend violates constant mean even if disturbances are homoscedastic. Weak stationarity is weaker than strict stationarity, which concerns the entire joint distribution under time shifts. For Gaussian processes, mean and covariance determine the distribution, linking the two notions more closely. Transformation, detrending, or differencing may be needed before fitting stationary dynamics.
Worked example: The series Yt = 5 + 0.2t + εt has constant disturbance variance but changing mean, so it is not weakly stationary. Subtracting the known deterministic trend leaves εt, which is stationary if it is identically distributed white noise with finite variance.
Common mistake: A flat residual spread does not prove stationarity; the mean and covariance structure must also remain invariant in the required sense.
Official syllabus reference: [2]
50. ACF and PACF measure different dependence
The autocorrelation function at lag k is Corr(Yt,Yt−k). Partial autocorrelation removes linear effects of intervening lags before assessing the remaining relationship. An AR(p) process has a theoretical PACF cutoff after p, while an MA(q) process has a theoretical ACF cutoff after q, under their standard stationary formulations. Sample functions fluctuate and rarely display exact zeros. Approximate individual-lag bounds are exploratory rather than simultaneous guarantees across many lags. Seasonal peaks suggest dependence at repeating intervals, while slow ACF decay can indicate persistence or nonstationarity requiring further investigation.
Worked example: A stationary sample has a gradually decaying ACF and a material PACF only at lag one. AR(1) is a reasonable initial candidate. If fitted residuals retain strong lag-two correlation, the initial identification was incomplete despite the suggestive original pattern.
Common mistake: A single sample correlation outside an approximate bound does not conclusively identify an ARIMA order or prove a fitted model invalid.
Official syllabus reference: [2]
51. White noise and innovations are related but distinct
Weak white noise has zero mean, constant finite variance, and zero autocovariance at nonzero lags. It need not be independent unless independence or Gaussian assumptions are added. Innovations are one-step prediction errors relative to the specified information set, so they are orthogonal to the available past under linear prediction and have conditional mean zero under stronger conditional-mean formulations. Residuals estimate unobserved innovations after fitting a model. Lack of residual autocorrelation supports the modeled linear dependence structure but does not alone establish independence, normality, or constant conditional variance.
Worked example: A fitted AR model produces residuals with negligible sample autocorrelation but unusually large squared residuals clustered together. The mean dynamics may be adequate, yet the residuals can still exhibit changing conditional volatility that ordinary residual ACF values do not reveal.
Common mistake: Uncorrelated observations are not necessarily independent; calling residuals white noise does not automatically establish every distributional assumption used in forecasting.
Official syllabus reference: [2]
52. AR(1) mean, variance, and mean reversion
For Yt = c + φYt−1 + εt with innovation variance σ², the causal stationary AR(1) requires |φ| < 1. Its mean is μ = c/(1 − φ), variance is σ²/(1 − φ²), and autocorrelation at lag k is φ^k. The h-step forecast is μ + φ^h(YT − μ). Positive φ produces persistent deviations, while negative φ produces alternating signs around the mean. The intercept c is the expected one-period increment contribution within the recursion, not generally the stationary mean. Innovations must satisfy the assumed relationship with past information.
Worked example: With c = 3, φ = 0.5, and innovation variance 9, the stationary mean is 6 and variance is 12. If the current value is 14, the two-step forecast is 6 + 0.25(14 − 6) = 8.
Common mistake: Using the AR intercept as the long-run mean ignores repeated propagation through the autoregressive term and generally produces incorrect forecasts.
Official syllabus reference: [2]
53. AR stationarity is determined by polynomial roots
Write an AR(p) model as φ(B)(Yt − μ) = εt, where φ(B) = 1 − φ1B − ... − φpB^p. A causal stationary solution requires roots of φ(z) = 0 outside the unit circle. Checking each coefficient separately is insufficient beyond AR(1). Stable roots produce decaying dependence, potentially with oscillation when complex roots occur. The PACF's theoretical cutoff at p supports identification but does not replace the root condition after fitting. Software may report roots or inverse roots, so determine which convention is being displayed before interpreting stability.
Worked example: For φ(B) = (1 − 0.5B)(1 − 0.25B), roots are 2 and 4, both outside the unit circle. Expanding gives an AR(2) with coefficients 0.75 and −0.125, illustrating that a negative second coefficient can coexist with stationarity.
Common mistake: Requiring every AR coefficient to lie between minus one and one is not a sufficient general stationarity test for higher-order models.
Official syllabus reference: [2]
54. MA invertibility differs from stationarity
A finite moving-average process Yt = μ + εt + θ1εt−1 + ... + θqεt−q is stationary for finite innovation variance, regardless of invertibility. Invertibility requires roots of 1 + θ1z + ... + θqz^q outside the unit circle and permits stable recovery of innovations from observed history. For MA(1), invertibility is |θ| < 1, and lag-one autocorrelation is θ/(1 + θ²) under the plus-sign convention. The theoretical ACF vanishes beyond lag q. Sign conventions differ across presentations, so read the actual model equation.
Worked example: An MA(1) with θ = 0.5 has lag-one autocorrelation 0.5/1.25 = 0.4. A noninvertible representation with θ = 2 has the same lag-one correlation, illustrating why an invertibility convention helps select a unique innovation representation.
Common mistake: A noninvertible finite MA process is not automatically nonstationary; invertibility concerns recovering innovations, while stationarity concerns the process distribution over time.
Official syllabus reference: [2]
55. ARMA forecasts propagate known and future innovations
An ARMA model combines autoregressive dependence on past observations with moving-average dependence on past innovations. Forecasting conditions on observed values and estimated available innovations, while future innovations have expectation zero. For ARMA(1,1) written with plus MA coefficient θ, the infinite-MA coefficients are ψ0 = 1 and ψj = (φ + θ)φ^(j−1) for j ≥ 1. These coefficients determine how future shocks enter multistep forecast errors. Known current innovations can affect near-term forecasts even though they contribute no conditional uncertainty when treated as observed and parameters are fixed.
Worked example: With φ = 0.4, θ = 0.3, current value 10, and current innovation 2, the next forecast is 4.6. The following forecast is 1.84. The second future value's error includes the next innovation multiplied by 0.7.
Common mistake: Setting all innovations to zero discards known recent innovation information; only innovations not yet observed are replaced by their conditional mean.
Official syllabus reference: [2]
56. ARIMA differencing changes both model and forecast
An ARIMA(p,d,q) model applies an ARMA(p,q) structure to the d-times differenced series. First differencing is ΔYt = Yt − Yt−1, while second differencing removes changes in first differences. Forecasts of differenced values must be accumulated back to the original scale using the latest observed levels and differences. Differencing targets stochastic nonstationarity but can be excessive; over-differencing may introduce unnecessary moving-average behavior and increase forecast uncertainty. The d parameter describes transformation order, not the number of lagged predictors or a deterministic polynomial degree selected automatically from a graph.
Worked example: The latest level is 100. Forecasted first differences for the next three periods are 2, 1, and 1. Original-scale forecasts are 102, 103, and 104 because each predicted difference updates the preceding forecasted level rather than independently updating 100.
Common mistake: A forecast for a differenced series is not yet a level forecast; it must be integrated using the appropriate observed starting values.
Official syllabus reference: [2]
57. Deterministic and stochastic trends respond differently to shocks
A deterministic trend model writes Yt = a + bt + ut, with stationary deviations ut. Removing the fitted trend can yield stationarity, and temporary shocks eventually dissipate according to the deviation dynamics. A stochastic trend, such as a random walk with drift, accumulates innovations: Yt = Yt−1 + δ + εt. Its shocks permanently alter future expected levels, and differencing rather than simple detrending produces a stationary increment process. Both models can produce upward-looking sample paths, so visual trend alone does not identify the appropriate mechanism or forecast uncertainty.
Worked example: A one-time shock of 10 in a trend-stationary AR(1) deviation with coefficient 0.5 contributes only 1.25 after three periods. The same innovation in a random walk shifts every subsequent expected level by 10, even though no further shocks occur.
Common mistake: Subtracting a fitted straight line does not generally remove a stochastic trend, because accumulated innovations remain embedded in the residual level.
Official syllabus reference: [2]
58. Seasonal regression needs an identifiable parameterization
Deterministic seasonality can be modeled using seasonal indicators or harmonic sine and cosine terms. With an intercept, include one fewer indicator than seasons or impose a constraint such as seasonal effects summing to zero. Otherwise the design is rank deficient. Forecasting combines the trend at the future time with the corresponding future seasonal effect. Harmonic terms provide smoother periodic patterns and can use fewer parameters than a full set of indicators. If regression disturbances are serially correlated, ordinary independent-error uncertainty calculations are inadequate even when the seasonal mean specification is correct.
Worked example: Quarterly seasonal effects constrained to sum to zero are −4, 1, 5, and −2. If the future trend level is 120 in quarter 3, the mean forecast is 125. The seasonal effect follows the future quarter, not the latest observed quarter.
Common mistake: An intercept plus an unrestricted indicator for every season duplicates one column combination and prevents unique estimation of all listed coefficients.
Official syllabus reference: [2]
59. Seasonal differencing removes repeating stochastic levels
For seasonal period s, seasonal differencing is (1 − B^s)Yt = Yt − Yt−s. It targets persistence between corresponding seasons rather than merely subtracting a fixed seasonal average. Combining ordinary and seasonal differencing gives (1 − B)(1 − B^s)Yt = Yt − Yt−1 − Yt−s + Yt−s−1. Deterministic seasonal indicators and seasonal differencing represent different assumptions about how seasonal patterns evolve. Forecasts must reverse all applied differences using the necessary historical values. Seasonal differencing should be justified by the series structure rather than automatically applied whenever a seasonal plot shows variation.
Worked example: For quarterly data, the current level is 120 and the same quarter last year was 111, so the seasonal difference is 9. If the next quarter's seasonal difference is forecast as 7 and its prior-year level was 115, the level forecast is 122.
Common mistake: Seasonal differencing is not equivalent to adding seasonal indicators; one models evolving seasonal levels while the other represents fixed seasonal mean effects.
Official syllabus reference: [2]
60. Forecast variance and residual checks complete an ARIMA analysis
For a causal model with innovation variance σ² and infinite-MA coefficients ψj, the h-step forecast-error variance with known parameters is σ²Σj=0 to h−1 ψj². Gaussian assumptions turn this variance into normal prediction intervals; estimated parameters can add uncertainty beyond that expression. Residual ACF and portmanteau tests assess remaining serial correlation, with reference degrees of freedom adjusted appropriately for fitted dynamics. A large diagnostic p-value is lack of evidence against the checked property, not proof of correctness. Prediction intervals describe future observations rather than uncertainty only in the conditional mean.
Worked example: An AR(1) has φ = 0.5 and innovation variance 4. Its three-step forecast-error variance is 4(1 + 0.25 + 0.0625) = 5.25. Under Gaussian known-parameter assumptions, a supplied 1.96 multiplier gives an interval extending about 4.49 around the forecast.
Common mistake: Passing a residual autocorrelation test does not establish normality, calibration, or parameter certainty, and a mean confidence interval is not a prediction interval.
Official syllabus reference: [2]