WHY MATHEMATICS?

Why Mathematics? | L1-Penalised Quantile Regression: Sparse Coefficients, Lambda Paths, Variable Selection and Tail Prediction

How check loss targets a percentile while an absolute-coefficient penalty creates sparsity, why scaling changes selection, and how cross-validation prevents leakage.

Why mathematics is important becomes easier to see when a real decision depends on it. This guide follows the mechanism from definitions and dimensions to a worked calculation, validation, practical learning and responsible use. It keeps factual evidence separate from interpretation and treats every illustrative scenario as an example rather than a promise of results.

Choose a reading route

Write quantile check loss plus the L1 coefficient penalty
Standardise predictors within every training fold
Trace the piecewise-linear coefficient path over lambda
Validate sparse selection, stability and post-selection uncertainty

Chapter map

1. Target
2. Data
3. Check loss
4. Penalty
5. Slope candidate
6. Intercept candidate
7. Threshold
8. Small lambda
9. Large lambda
10. Scaling
11. Correlated predictors
12. Solution path
13. Cross-validation
14. Post-selection
15. Transfer
16. Claim

Visit the How Mathematics Works hub for the wider series and its connections to education, reasoning, science, computing, design and careers.

1. Target

This is where modelling choices begin to affect evidence. Choose tau and response orientation.

A concrete situation

Use tau=.5.

What the mathematics does not license

Another tau selects another sparse model. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Record level. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

2. Data

The central mathematical idea is concrete. Use (1,1),(2,2).

A concrete situation

A perfect slope exists.

What the mathematics does not license

Two points are only a mechanism example. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Label illustration. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

3. Check loss

Here the geometry becomes operational. Use half absolute residual at median.

A concrete situation

Candidate slope one has zero loss.

What the mathematics does not license

Software may omit the half factor. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Match lambda scale. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

4. Penalty

This step deserves a careful pause. Add lambda times absolute slopes.

A concrete situation

Do not penalise intercept in teaching case.

What the mathematics does not license

Penalty conventions vary. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

State factors. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

Did You Know? In this chapter, the most important number is useful only together with its construction rule. Changing scaling, sample membership or validation design can change the numerical answer without changing the arithmetic.

5. Slope candidate

The useful question is not only how to calculate, but what the calculation assumes. Set alpha=0,beta=1.

A concrete situation

Objective=lambda.

What the mathematics does not license

One candidate is not full path. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Solve globally. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

6. Intercept candidate

A small example makes the mechanism visible. Set alpha=1.5,beta=0.

A concrete situation

Loss=.5, penalty zero.

What the mathematics does not license

Even-sample medians form an interval. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

State chosen alpha. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

7. Threshold

This is where modelling choices begin to affect evidence. Solve lambda=.5.

A concrete situation

Candidates tie at .5.

What the mathematics does not license

Other coefficients may also tie. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Inspect optimum set. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

8. Small lambda

The central mathematical idea is concrete. Take lambda below .5.

A concrete situation

Slope candidate wins this comparison.

What the mathematics does not license

Low penalty can overfit. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Validate. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

9. Large lambda

Here the geometry becomes operational. Take lambda above .5.

A concrete situation

Intercept candidate wins this comparison.

What the mathematics does not license

High penalty can underfit. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Check bias. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

10. Scaling

This step deserves a careful pause. Multiply x by ten.

A concrete situation

Coefficient shrinks by ten for same fit.

What the mathematics does not license

Unadjusted penalty changes unfairly. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Standardise in folds. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

11. Correlated predictors

The useful question is not only how to calculate, but what the calculation assumes. Add a near-duplicate x.

A concrete situation

Either may enter.

What the mathematics does not license

Selection instability is expected. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Use stability analysis. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

12. Solution path

A small example makes the mechanism visible. Track coefficient knots over lambda.

A concrete situation

Piecewise-linear changes occur.

What the mathematics does not license

A coarse grid can miss transitions. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Report grid. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

13. Cross-validation

This is where modelling choices begin to affect evidence. Score held-out check loss.

A concrete situation

Keep preprocessing inside training folds.

What the mathematics does not license

Leakage makes lambda look better. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Audit pipeline. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

14. Post-selection

The central mathematical idea is concrete. Treat chosen model as data-dependent.

A concrete situation

Use appropriate uncertainty methods.

What the mathematics does not license

Naive fixed-model intervals understate uncertainty. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Disclose selection. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

15. Transfer

Here the geometry becomes operational. Explain sparse tail predictors in original units.

A concrete situation

Keep useful variables open to review.

What the mathematics does not license

Zero is not proof of no cause. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Keep options open. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

16. Claim

This step deserves a careful pause. State tau, scaling, penalty and lambda rule.

A concrete situation

Teaching threshold=.5.

What the mathematics does not license

Do not universalise it. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Bound conclusion. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

A seven-day practice lab

Day 1: define the decision, observation unit, X variables and Y or reference condition.
Day 2: reproduce the smallest worked example and check every dimension, unit and sign.
Day 3: build a transparent baseline and write what it assumes.
Day 4: fit the main method inside a training-only pipeline; record every parameter.
Day 5: challenge one assumption with a deliberately difficult dataset or operating regime.
Day 6: compare validation metrics with the practical workload created by errors.
Day 7: present one supported conclusion, one unresolved uncertainty and one safe next experiment.

Questions for a student, parent or mentor

Which part of the model is a mathematical definition, and which part is a choice?
What small calculation would catch a sign, scale or dimension error?
What evidence comes from training, validation and the untouched test?
What benign condition could look like a fault or prediction failure?
How would the answer change for another response, asset, school project or time period?
What should a human review before acting on the score?

A method-specific audit before deployment

L1-penalised quantile regression adds the sum of absolute slope coefficients to the usual check-loss objective. The quantile loss targets a conditional percentile; the coefficient penalty controls model complexity and can set some slopes exactly to zero. This creates a sparse model that may be easier to interpret when predictors are numerous, while keeping the asymmetric decision meaning of the chosen quantile.

For points (x,y)=(1,1),(2,2) at tau=.5, candidate intercept zero and slope one has check loss zero. If the intercept is unpenalised and the slope penalty is lambda|beta|, its objective is lambda. An intercept-only candidate at 1.5 has two residual magnitudes .5, median check-loss total .5 and zero slope penalty. Between these two candidates, the slope wins when lambda<.5, they tie at .5 and the intercept-only candidate wins when lambda>.5.

A full solution path can change at discrete lambda values because both check loss and the L1 penalty are piecewise linear. Correlated predictors can substitute for one another, so variable selection may be unstable even when prediction is adequate. Scaling matters: multiplying a predictor changes its coefficient and therefore its penalty unless predictors are standardised or penalty factors compensate.

Select lambda using a declared criterion such as cross-validated check loss at the target quantile, and keep all preprocessing inside each training fold. Report whether the intercept is penalised, predictor scaling, penalty factors, lambda grid, solver tolerance, selected variables, stability and post-selection uncertainty. A zero coefficient is not proof that a variable is scientifically irrelevant or causally inactive.

Frequently asked questions

What is penalised?

Usually the absolute values of slope coefficients, with the intercept left unpenalised.

Why can coefficients become zero?

The L1 penalty has corners that support sparse optima.

Why does scaling matter?

The same fitted effect can have a different coefficient magnitude after rescaling.

How is lambda chosen?

By a declared criterion such as cross-validated check loss.

Is a zero coefficient proof of irrelevance?

No. Correlation, scale and selection instability affect the chosen model.

What must be reported?

Tau, scaling, penalty factors, lambda grid, validation, stability and uncertainty.

Primary source and fact-check note

Youjuan Li and Ji Zhu, “L1-Norm Quantile Regression,” Journal of Computational and Graphical Statistics 17(1), 163–185 (2008), DOI 10.1198/106186008X289155. Definitions and core claims in this article were checked against this primary publication. Calculations shown here were independently recomputed. Examples are educational illustrations; they are not reported industrial performance results.

Useful next reading

Keep the options open. Mathematics education supports clearer questions, transferable problem-solving skills and informed choices across science, engineering, computing, AI, statistics, design and many other careers. It does not guarantee intelligence, income, admission or employment; those outcomes depend on far more than one subject or model.

Extended laboratory: topic-specific validation and responsible use

Design audit

Freeze sampling units, predictor support, time order, response orientation, missingness and dependence.

Objective audit

Write likelihood, prior, forest weights, check losses, penalties or lag equations explicitly.

Tuning audit

Record scale, priors, sampler settings, tree settings, lambda grids, tau grids and lag order.

Computation audit

Check posterior convergence, weight normalisation, solution paths, tied optima and solver tolerances.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Coherence audit

Inspect quantile order, support, stability, effective neighbours, hit sequences and extrapolation.

Uncertainty audit

Use model-compatible posterior, sandwich or bootstrap methods and include tuning or repair.

Validation audit

Use untouched or rolling data for check loss, coverage, width and subgroup calibration.

Final claim

Translate the conditional tail result without causal, posterior or guarantee overclaims.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Mechanism workshop: calculate, challenge and transfer

Workshop 1: Target

Begin with this mechanism: Choose tau and response orientation. Recreate the concrete case: Use tau=.5. Write predictors, responses, quantile levels, residuals, check losses, scaled coefficients, penalty factors, lambda objectives, selected variables and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Another tau selects another sparse model. Complete the transfer step: Record level. Change exactly one observation, quantile level, ALD scale, prior, chain, tree, leaf size, forest weight, inverse-CDF convention, predictor scaling, penalty factor, lambda, selected variable, lag, structural break, solver vertex, forecast origin, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, tuning, computation, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving likelihood misspecification, sparse support, correlated predictors, time leakage, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 2: Data

Begin with this mechanism: Use (1,1),(2,2). Recreate the concrete case: A perfect slope exists. Write predictors, responses, quantile levels, residuals, check losses, scaled coefficients, penalty factors, lambda objectives, selected variables and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Two points are only a mechanism example. Complete the transfer step: Label illustration. Change exactly one observation, quantile level, ALD scale, prior, chain, tree, leaf size, forest weight, inverse-CDF convention, predictor scaling, penalty factor, lambda, selected variable, lag, structural break, solver vertex, forecast origin, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, tuning, computation, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving likelihood misspecification, sparse support, correlated predictors, time leakage, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 3: Check loss

Begin with this mechanism: Use half absolute residual at median. Recreate the concrete case: Candidate slope one has zero loss. Write predictors, responses, quantile levels, residuals, check losses, scaled coefficients, penalty factors, lambda objectives, selected variables and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Software may omit the half factor. Complete the transfer step: Match lambda scale. Change exactly one observation, quantile level, ALD scale, prior, chain, tree, leaf size, forest weight, inverse-CDF convention, predictor scaling, penalty factor, lambda, selected variable, lag, structural break, solver vertex, forecast origin, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, tuning, computation, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving likelihood misspecification, sparse support, correlated predictors, time leakage, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 4: Penalty

Begin with this mechanism: Add lambda times absolute slopes. Recreate the concrete case: Do not penalise intercept in teaching case. Write predictors, responses, quantile levels, residuals, check losses, scaled coefficients, penalty factors, lambda objectives, selected variables and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Penalty conventions vary. Complete the transfer step: State factors. Change exactly one observation, quantile level, ALD scale, prior, chain, tree, leaf size, forest weight, inverse-CDF convention, predictor scaling, penalty factor, lambda, selected variable, lag, structural break, solver vertex, forecast origin, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, tuning, computation, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving likelihood misspecification, sparse support, correlated predictors, time leakage, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Workshop 5: Slope candidate

Begin with this mechanism: Set alpha=0,beta=1. Recreate the concrete case: Objective=lambda. Write predictors, responses, quantile levels, residuals, check losses, scaled coefficients, penalty factors, lambda objectives, selected variables and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: One candidate is not full path. Complete the transfer step: Solve globally. Change exactly one observation, quantile level, ALD scale, prior, chain, tree, leaf size, forest weight, inverse-CDF convention, predictor scaling, penalty factor, lambda, selected variable, lag, structural break, solver vertex, forecast origin, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, tuning, computation, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving likelihood misspecification, sparse support, correlated predictors, time leakage, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 6: Intercept candidate

Begin with this mechanism: Set alpha=1.5,beta=0. Recreate the concrete case: Loss=.5, penalty zero. Write predictors, responses, quantile levels, residuals, check losses, scaled coefficients, penalty factors, lambda objectives, selected variables and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Even-sample medians form an interval. Complete the transfer step: State chosen alpha. Change exactly one observation, quantile level, ALD scale, prior, chain, tree, leaf size, forest weight, inverse-CDF convention, predictor scaling, penalty factor, lambda, selected variable, lag, structural break, solver vertex, forecast origin, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, tuning, computation, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving likelihood misspecification, sparse support, correlated predictors, time leakage, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 7: Threshold

Begin with this mechanism: Solve lambda=.5. Recreate the concrete case: Candidates tie at .5. Write predictors, responses, quantile levels, residuals, check losses, scaled coefficients, penalty factors, lambda objectives, selected variables and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Other coefficients may also tie. Complete the transfer step: Inspect optimum set. Change exactly one observation, quantile level, ALD scale, prior, chain, tree, leaf size, forest weight, inverse-CDF convention, predictor scaling, penalty factor, lambda, selected variable, lag, structural break, solver vertex, forecast origin, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, tuning, computation, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving likelihood misspecification, sparse support, correlated predictors, time leakage, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 8: Small lambda

Begin with this mechanism: Take lambda below .5. Recreate the concrete case: Slope candidate wins this comparison. Write predictors, responses, quantile levels, residuals, check losses, scaled coefficients, penalty factors, lambda objectives, selected variables and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Low penalty can overfit. Complete the transfer step: Validate. Change exactly one observation, quantile level, ALD scale, prior, chain, tree, leaf size, forest weight, inverse-CDF convention, predictor scaling, penalty factor, lambda, selected variable, lag, structural break, solver vertex, forecast origin, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, tuning, computation, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving likelihood misspecification, sparse support, correlated predictors, time leakage, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Workshop 9: Large lambda

Begin with this mechanism: Take lambda above .5. Recreate the concrete case: Intercept candidate wins this comparison. Write predictors, responses, quantile levels, residuals, check losses, scaled coefficients, penalty factors, lambda objectives, selected variables and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: High penalty can underfit. Complete the transfer step: Check bias. Change exactly one observation, quantile level, ALD scale, prior, chain, tree, leaf size, forest weight, inverse-CDF convention, predictor scaling, penalty factor, lambda, selected variable, lag, structural break, solver vertex, forecast origin, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, tuning, computation, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving likelihood misspecification, sparse support, correlated predictors, time leakage, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 10: Scaling

Begin with this mechanism: Multiply x by ten. Recreate the concrete case: Coefficient shrinks by ten for same fit. Write predictors, responses, quantile levels, residuals, check losses, scaled coefficients, penalty factors, lambda objectives, selected variables and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Unadjusted penalty changes unfairly. Complete the transfer step: Standardise in folds. Change exactly one observation, quantile level, ALD scale, prior, chain, tree, leaf size, forest weight, inverse-CDF convention, predictor scaling, penalty factor, lambda, selected variable, lag, structural break, solver vertex, forecast origin, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, tuning, computation, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving likelihood misspecification, sparse support, correlated predictors, time leakage, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 11: Correlated predictors

Begin with this mechanism: Add a near-duplicate x. Recreate the concrete case: Either may enter. Write predictors, responses, quantile levels, residuals, check losses, scaled coefficients, penalty factors, lambda objectives, selected variables and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Selection instability is expected. Complete the transfer step: Use stability analysis. Change exactly one observation, quantile level, ALD scale, prior, chain, tree, leaf size, forest weight, inverse-CDF convention, predictor scaling, penalty factor, lambda, selected variable, lag, structural break, solver vertex, forecast origin, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, tuning, computation, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving likelihood misspecification, sparse support, correlated predictors, time leakage, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 12: Solution path

Begin with this mechanism: Track coefficient knots over lambda. Recreate the concrete case: Piecewise-linear changes occur. Write predictors, responses, quantile levels, residuals, check losses, scaled coefficients, penalty factors, lambda objectives, selected variables and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: A coarse grid can miss transitions. Complete the transfer step: Report grid. Change exactly one observation, quantile level, ALD scale, prior, chain, tree, leaf size, forest weight, inverse-CDF convention, predictor scaling, penalty factor, lambda, selected variable, lag, structural break, solver vertex, forecast origin, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, tuning, computation, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving likelihood misspecification, sparse support, correlated predictors, time leakage, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Workshop 13: Cross-validation

Begin with this mechanism: Score held-out check loss. Recreate the concrete case: Keep preprocessing inside training folds. Write predictors, responses, quantile levels, residuals, check losses, scaled coefficients, penalty factors, lambda objectives, selected variables and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Leakage makes lambda look better. Complete the transfer step: Audit pipeline. Change exactly one observation, quantile level, ALD scale, prior, chain, tree, leaf size, forest weight, inverse-CDF convention, predictor scaling, penalty factor, lambda, selected variable, lag, structural break, solver vertex, forecast origin, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, tuning, computation, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving likelihood misspecification, sparse support, correlated predictors, time leakage, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 14: Post-selection

Begin with this mechanism: Treat chosen model as data-dependent. Recreate the concrete case: Use appropriate uncertainty methods. Write predictors, responses, quantile levels, residuals, check losses, scaled coefficients, penalty factors, lambda objectives, selected variables and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Naive fixed-model intervals understate uncertainty. Complete the transfer step: Disclose selection. Change exactly one observation, quantile level, ALD scale, prior, chain, tree, leaf size, forest weight, inverse-CDF convention, predictor scaling, penalty factor, lambda, selected variable, lag, structural break, solver vertex, forecast origin, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, tuning, computation, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving likelihood misspecification, sparse support, correlated predictors, time leakage, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 15: Transfer

Begin with this mechanism: Explain sparse tail predictors in original units. Recreate the concrete case: Keep useful variables open to review. Write predictors, responses, quantile levels, residuals, check losses, scaled coefficients, penalty factors, lambda objectives, selected variables and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Zero is not proof of no cause. Complete the transfer step: Keep options open. Change exactly one observation, quantile level, ALD scale, prior, chain, tree, leaf size, forest weight, inverse-CDF convention, predictor scaling, penalty factor, lambda, selected variable, lag, structural break, solver vertex, forecast origin, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, tuning, computation, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving likelihood misspecification, sparse support, correlated predictors, time leakage, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 16: Claim

Begin with this mechanism: State tau, scaling, penalty and lambda rule. Recreate the concrete case: Teaching threshold=.5. Write predictors, responses, quantile levels, residuals, check losses, scaled coefficients, penalty factors, lambda objectives, selected variables and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Do not universalise it. Complete the transfer step: Bound conclusion. Change exactly one observation, quantile level, ALD scale, prior, chain, tree, leaf size, forest weight, inverse-CDF convention, predictor scaling, penalty factor, lambda, selected variable, lag, structural break, solver vertex, forecast origin, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, tuning, computation, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving likelihood misspecification, sparse support, correlated predictors, time leakage, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.