Mock Exams

Calibrated mock exams for the May 18, 2026 TMA4268 final. Each mock is constructed from the prof’s exam-review lecture, the exam analysis, the 2023—2025 finals, and the in-scope concept atoms. Per-problem weights add to $100\%=100$ points and map directly onto NTNU’s prosentvurderingsmetoden grade boundaries (A: 89—100, B: 77—88, C: 65—76, D: 53—64, E: 41—52, F: 0—40).

Workflow

Tackle the mock cold under exam conditions (4 hours, open book: ISLP + your A5 sheet + a calculator). Then compare against the solution PDF. The solution sketches use the same partial-credit grading conventions Stefanie and Sara used on the official 2024 / 2025 keys.

Available mocks

#	Exam	Solution proposal	Weighting (per problem, %)
1	Mock Exam 1 (PDF)	Solution (PDF)	10 / 28 / 16 / 22 / 24
2	Mock Exam 2 (PDF)	Solution (PDF)	10 / 28 / 16 / 22 / 24
3	Mock Exam 3 (PDF)	Solution (PDF)	10 / 28 / 16 / 22 / 24
4	Mock Exam 4 (PDF)	Solution (PDF)	10 / 28 / 16 / 22 / 24
5	Mock Exam 5 (PDF)	Solution (PDF)	10 / 28 / 16 / 22 / 24
6	Mock Exam 6 (PDF)	Solution (PDF)	10 / 28 / 16 / 18 / 28
7	Mock Exam 7 (PDF)	Solution (PDF)	10 / 28 / 16 / 20 / 26
8	Mock Exam 8 (PDF)	Solution (PDF)	10 / 28 / 16 / 22 / 24
9	Mock Exam 9 (PDF)	Solution (PDF)	10 / 28 / 16 / 22 / 24
10	Mock Exam 10 (PDF)	Solution (PDF)	10 / 28 / 16 / 20 / 26

If you prefer HTML

LaTeX-rendered web versions of every mock (and its solution), built from the same .tex sources with pandoc + MathJax. Math is live and selectable; the layout mimics the compiled PDF rather than the wiki. TikZ figures are replaced by a note pointing back to the PDF.

#	Exam (HTML)	Solution (HTML)
1	Mock Exam 1	Solution
2	Mock Exam 2	Solution
3	Mock Exam 3	Solution
4	Mock Exam 4	Solution
5	Mock Exam 5	Solution
6	Mock Exam 6	Solution
7	Mock Exam 7	Solution
8	Mock Exam 8	Solution
9	Mock Exam 9	Solution
10	Mock Exam 10	Solution

Regenerate with web/scripts/build_exam_html.sh after any .tex edit.

Per-mock topic rotation

LaTeX sources live at mock-exams/mock-exam-N.tex and mock-exams/mock-exam-N-solution.tex for mocks 1—7, and under mock-exams/mock-N/ for mocks 8—10. Each mock deliberately rotates topics relative to the others so they cover complementary slices of the prof’s scope:

• Mock 1 --- MLE,=,LS mathy derivation; complete-linkage hierarchical clustering on a 4$\times$4 matrix; lasso (with $\lambda$ via 10-fold CV) as the regularizer; the logistic-regression interaction trap (\texttt{balance},$\times$,\texttt{sex}) sits in the classification problem.
• Mock 2 --- variance of an average of correlated predictors as the mathy derivation (the random-forest decorrelation argument); single-linkage hierarchical clustering on a 5$\times$5 matrix; ridge \emph{and} PCR side by side; subset-selection MC; the logistic interaction trap again.
• Mock 3 --- LDA decision-boundary derivation as the mathy slot (the prof flagged this twice in lecture as “a typical exam question”); single-linkage hierarchical clustering on a different 4$\times$4 matrix; ridge in regression alongside a GAM with smoothing-spline + cubic-spline terms and gradient-boosted trees; LDA vs.\ QDA confusion-matrix comparison in the classification problem; the interaction trap (\texttt{bmi},$\times$,\texttt{smoker}) is on the OLS side, not the logistic side; bootstrap, K-means, and OOB-error MC slots that mock 1 didn’t have.
• Mock 4 (calibration pass) --- the bias—variance decomposition derivation as the mathy slot (the prof’s guaranteed exam topic, re-flagged three times); two pseudocode problems (nested $k$-fold CV with the wrong-way trap, bootstrap SE for the sample median, gradient-boosting inner loop); harder linear regression with explicit (near-)collinearity between \texttt{area}, \texttt{rooms}, and \texttt{area_per_room} plus an \texttt{age}-quadratic minimum and a standardised \texttt{distance}$\times$\texttt{has_balcony} interaction; classification problem is the heaviest yet, with logistic + a regularized neural net (dropout, early stopping, label smoothing) + AdaBoost; XGBoost / mini-batch SGD / backprop / boosting-flavor T/F all show up in P2; hierarchical clustering and $K$-means kept to a 2-pt T/F so they don’t disappear entirely.
• Mock 5 (calibration pass) --- bias—variance decomposition as the mathy slot (the prof’s explicitly promised topic from L27); calibration pass weighting up classification, CV pseudocode + nested CV, bootstrap SE, collinearity, backprop + mini-batch SGD, label smoothing / early stopping / dropout, and the AdaBoost / gradient-boosting / XGBoost family; weighting down dendrograms (kept as a small 2-pt complete-linkage 4$\times$4) and k-means (dropped); hardened linear-regression problem (polynomial + interaction + 3-level categorical + standardized predictors), small backprop chain-rule derivation in the NN classifier; AdaBoost pseudocode plus a single-iteration numerical drill.
• Mock 6 (calibration pass) --- bias—variance decomposition as the mathy slot (the prof’s most-promised exam topic), with a shrinkage-estimator application; nested-CV pseudocode and a small backpropagation chain-rule derivation flesh out P3. Concrete-strength regression with deliberate \texttt{water}/\texttt{superplast} collinearity (and a continuous,$\times$,categorical interaction) makes the linear-regression problem the hardest of the series; telecom-churn classification ramps up classification share, with AdaBoost pseudocode + hand-calc of ${\alpha }_m$ and weight updates as the boosting question. Heavier weight on cross-validation (incl. the wrong-way / nested-CV pitfall), bootstrap SE, collinearity, mini-batch SGD, dropout / early stopping / label smoothing, and the boosting variants (AdaBoost, gradient boosting, XGBoost); dendrograms, ROC / AUC / sensitivity / specificity, and $K$-means deliberately down-weighted relative to mocks 1—3.
• Mock 7 (calibration pass) --- bias,—,variance \emph{decomposition derivation} as the mathy slot (the prof’s explicit promise in L27); CV and \emph{nested} CV pseudocode + an AdaBoost-by-hand round in Problem 3; collinearity in linear regression diagnosed via two compatible fingerprints (correlation $0.99$, inflated SEs) on a wine-quality dataset; classification weighted up to $26\%$ via a churn problem comparing logistic regression, LDA, and gradient boosting; dendrograms / AUC / $K$-means correspondingly downweighted; backprop, mini-batch SGD, dropout, label smoothing, and early stopping promoted into Problem 2.
• Mock 8 --- LDA decision-boundary derivation as the mathy slot, including equal-prior algebra and prior-shift interpretation; mini-batch SGD and $k$-fold CV pseudocode in Problem 3; regression problem on bike-sharing demand with polynomial terms, interactions, collinearity, ridge/lasso comparison, gradient boosting, and a spline-based GAM; classification problem on bank-loan default with logistic interactions, AdaBoost from exponential loss, and a neural-network classifier with a backprop derivation.
• Mock 9 --- LDA decision-boundary derivation plus mini-batch SGD pseudocode and an OOB-probability bootstrap calculation; fuel-economy regression problem with B-splines, categorical contrasts, collinearity diagnostics, principal-component regression, and GAM comparison; bank-loan default classification problem comparing logistic regression, LDA/QDA discriminants, XGBoost regularization knobs, ROC, sensitivity, specificity, and class imbalance.
• Mock 10 --- LDA-to-QDA bridge as the mathy slot, with explicit boundary algebra and the quadratic term that stops cancelling under QDA; $k$-fold CV one-SE pseudocode plus a small backprop chain-rule calculation; diabetes-progression regression with quadratic terms, interactions, lasso CV, gradient boosting, GAM, and random forest; South African heart-disease classification with logistic interactions, AdaBoost classifier-weight derivation, neural-network regularization, class imbalance, and KNN scaling.

How each mock is constructed

The construction recipe is the same every time:

1. Format mix from the prof’s stated signals. Mostly multiple choice + true/false + short interpretation, plus exactly one mathy / derivation problem, plus two data-analysis problems where the regression / GLM / CV / ROC output is given (no code expected).
2. Topic balance from the slide-and-exercise scope. Roughly equal weight across the twelve modules, biased toward what the prof flagged: bias—variance (guaranteed), interactions, regularization, classification metrics, PCA, hand-by-hand hierarchical clustering, NN parameter counting, boosting hyperparameters.
3. Past-exam reformulation rules. Wherever a 2023—2025 question used “fit this in R,” the mock replaces it with “here is the output, interpret.” Old paper coding tasks become interpretation tasks; old paper conceptual / theoretical / hand-calculation tasks stay intact.
4. Difficulty calibration. Aimed to spread strong, average, and weak students across the percentile bands. There is no “trick” problem; partial credit is generous; the mathy slot follows the prof’s bar of “mathy but not, you know, no weird spaces or fancy proofs.”
5. Internal audit pass. Each mock goes through a critic agent that scores every sub-part for likelihood-given-prof-signals and a solver agent that produces the official-style solution proposal. Disagreements get reconciled before the PDF is published.

Related

• Exam analysis --- canonical synthesis of the prof’s signals across all 27 lectures.
• L27 (Apr 28) --- the exam-review lecture, the highest-priority single source for “what’s actually on the test.”
• Past exams (2023—2025) --- with official solutions inline.
• Scope rule --- the authoritative “what’s in / what’s out” reference, derived from slides + lectures + exercises.