WHY MATHEMATICS?

Why Mathematics? | Dawid–Sebastiani Score: Predictive Means, Variances, Log Dispersion and Standardised Errors

How a moment-based score judges both accuracy and confidence, why covariance matrices matter, and where first-and-second-moment evaluation reaches its limit.

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

Combine mean error with a declared predictive variance
Understand the balance between log dispersion and standardised error
Extend the calculation to vectors and covariance matrices
Add shape and tail checks that a moment score cannot provide

Chapter map

1. Observation
2. Forecast A mean
3. Forecast A spread
4. A dispersion term
5. A error term
6. A total
7. Forecast B
8. Forecast C
9. Repeated evaluation
10. Moment scope
11. Multivariate mean
12. Covariance
13. Correlation
14. Regularisation
15. Shape check
16. Claim

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

1. Observation

A small example makes the mechanism visible. Set realised y=13.

A concrete situation

Keep one response unit.

What the mathematics does not license

A single case is illustrative. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Aggregate later. 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. Forecast A mean

This is where modelling choices begin to affect evidence. Use mu=10.

A concrete situation

Error is 3.

What the mathematics does not license

Mean error alone ignores confidence. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Add variance. 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. Forecast A spread

The central mathematical idea is concrete. Use sigma=2.

A concrete situation

Variance is 4.

What the mathematics does not license

Standard deviation must be positive. A trustworthy workflow records the choice before seeing the final test result.

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.

4. A dispersion term

Here the geometry becomes operational. Compute log variance.

A concrete situation

log(4)≈1.3863.

What the mathematics does not license

Log base and constants must be consistent. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Use natural log. 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. A error term

This step deserves a careful pause. Standardise squared error.

A concrete situation

9/4=2.25.

What the mathematics does not license

Units cancel only after variance scaling. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Show units. 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. A total

The useful question is not only how to calculate, but what the calculation assumes. Add both terms.

A concrete situation

DSS≈3.6363.

What the mathematics does not license

Lower is better under loss orientation. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Label orientation. 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. Forecast B

A small example makes the mechanism visible. Use mu=11 and sigma=1.

A concrete situation

DSS=0+4=4.

What the mathematics does not license

A closer mean can lose if too confident. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Compare. 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. Forecast C

This is where modelling choices begin to affect evidence. Use mu=13 and sigma=5.

A concrete situation

DSS=log(25)≈3.2189.

What the mathematics does not license

An exact mean on one case may be luck. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Repeat cases. 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. Repeated evaluation

The central mathematical idea is concrete. Average over matched cases.

A concrete situation

Compare systems on identical targets.

What the mathematics does not license

Case selection changes results. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Freeze sample. 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. Moment scope

Here the geometry becomes operational. Recognise use of mean and variance only.

A concrete situation

Shape is not scored directly.

What the mathematics does not license

Equal moments can hide different tails. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Add diagnostics. 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. Multivariate mean

This step deserves a careful pause. Use a vector of target means.

A concrete situation

Align coordinates and units.

What the mathematics does not license

Mixing scales distorts covariance. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Standardise thoughtfully. 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. Covariance

The useful question is not only how to calculate, but what the calculation assumes. Estimate positive-definite Sigma.

A concrete situation

Use log determinant and inverse.

What the mathematics does not license

Singularity makes the score unstable. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Check eigenvalues. 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. Correlation

A small example makes the mechanism visible. Keep cross-target dependence.

A concrete situation

Off-diagonal terms affect Mahalanobis error.

What the mathematics does not license

Marginal scores miss dependence. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Report joint result. 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. Regularisation

This is where modelling choices begin to affect evidence. Stabilise covariance when dimension is high.

A concrete situation

Declare shrinkage or structure.

What the mathematics does not license

Regularisation changes the evaluated forecast. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Sensitivity check. 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. Shape check

The central mathematical idea is concrete. Inspect quantiles and tail events separately.

A concrete situation

Compare with a fuller distributional score.

What the mathematics does not license

DSS cannot certify multimodality or skewness. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Use companion evidence. 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

Here the geometry becomes operational. State moment-based scope and formula.

A concrete situation

Teaching scores are 3.6363,4 and 3.2189.

What the mathematics does not license

Do not claim full-distribution correctness. Evidence becomes stronger when the limitation is tested instead of hidden.

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

The Dawid–Sebastiani score evaluates a probabilistic forecast through its predictive mean and variance, or through a mean vector and covariance matrix in the multivariate case. In one dimension, a common negatively oriented form is log(sigma squared)+(y-mu) squared divided by sigma squared. The logarithmic term penalises excessive dispersion, while the standardised squared-error term penalises a mean that is far from the observation relative to the forecast variance. Lower is better when this loss convention is used.

Suppose the observation is 13. Forecast A has mean 10 and standard deviation 2, giving log(4)+9/4≈3.6363. Forecast B has mean 11 and standard deviation 1, giving log(1)+4=4, so A has the lower score despite the less accurate mean because B was too confident. Forecast C has mean 13 and standard deviation 5, giving log(25)+0≈3.2189. Its mean is exact in this one case, yet the broad variance still incurs a penalty. Repeated cases are needed before comparing systems.

The score deliberately uses only the first two moments. Two predictive distributions with the same mean and variance receive the same Dawid–Sebastiani score for the same observation even if their skewness, tail weight or number of modes differs. That can be a computational advantage when full densities are unavailable, but it is also a limitation. In more than one dimension, an estimated covariance matrix must be positive definite and well conditioned because the calculation uses a determinant and an inverse.

A responsible analysis declares the score orientation and any omitted constants, verifies units, reports how means and covariance matrices were estimated and checks numerical conditioning. It compares marginal and joint performance, examines tail and shape diagnostics separately and never calls a low moment score proof that the entire distribution is correct. Students can use the mechanism to see why forecast accuracy and calibrated confidence must be judged together.

Frequently asked questions

What does DSS use?

The predictive mean and variance, or a mean vector and covariance matrix.

Why is the error standardised?

A three-unit miss has different implications under a narrow and a broad forecast.

Why add log variance?

Without a dispersion cost, a forecast could evade error penalties by becoming arbitrarily wide.

What changes in several dimensions?

Variance becomes a covariance matrix, with a log determinant and Mahalanobis error.

Can DSS distinguish equal-moment distributions?

No. Distributions sharing the same mean and covariance receive the same score for the same observation.

What must be reported?

Formula, orientation, constants, moment estimation, covariance conditioning, case set, uncertainty and companion shape checks.

Primary source and fact-check note

A. Philip Dawid and Paola Sebastiani, “Coherent Dispersion Criteria for Optimal Experimental Design,” The Annals of Statistics 27(1), 65–81 (1999), DOI 10.1214/aos/1018031101. 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: score-specific validation and responsible use

Formula audit

State whether the score is log variance plus standardised squared error, include any constants and identify the preferred direction.

Moment audit

Verify predictive means, positive variances and units from the saved forecast object rather than reconstructing them after outcomes arrive.

Matrix audit

For multivariate forecasts, test covariance symmetry, positive definiteness, eigenvalues, determinant and numerical inverse stability.

Accuracy audit

Separate mean error from dispersion contribution so a final score cannot hide overconfidence or excessive width.

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.

Shape audit

Add quantile, tail, skewness or multimodality diagnostics because first two moments do not identify a full distribution.

Comparison audit

Evaluate identical target vectors and observation sets; treat dimension, ordering and standardisation as part of the score contract.

Sensitivity audit

Vary justified covariance regularisation and show whether the ranking survives.

Final claim

Call the result a moment-based probabilistic score; never claim it validates every feature of the predictive distribution.

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: Observation

Begin with this mechanism: Set realised y=13. Recreate the concrete case: Keep one response unit. Write outcome units, predictive mean, standard deviation, variance, log variance, squared error, standardised error, covariance determinant and score with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: A single case is illustrative. Complete the transfer step: Aggregate later. Change exactly one interval endpoint, alpha level, outcome, median, quantile grid, interval weight, normalisation, predictive mean, variance, covariance, threshold, weight function, ensemble value or case weight. Explain which mathematical object changes, which score term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving lucky coverage, excessive width, quantile crossing, unmatched grids, scale domination, overconfidence, singular covariance, identical moments with different tails, threshold cherry-picking or regional improvement with global deterioration. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 2: Forecast A mean

Begin with this mechanism: Use mu=10. Recreate the concrete case: Error is 3. Write outcome units, predictive mean, standard deviation, variance, log variance, squared error, standardised error, covariance determinant and score with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Mean error alone ignores confidence. Complete the transfer step: Add variance. Change exactly one interval endpoint, alpha level, outcome, median, quantile grid, interval weight, normalisation, predictive mean, variance, covariance, threshold, weight function, ensemble value or case weight. Explain which mathematical object changes, which score term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving lucky coverage, excessive width, quantile crossing, unmatched grids, scale domination, overconfidence, singular covariance, identical moments with different tails, threshold cherry-picking or regional improvement with global deterioration. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 3: Forecast A spread

Begin with this mechanism: Use sigma=2. Recreate the concrete case: Variance is 4. Write outcome units, predictive mean, standard deviation, variance, log variance, squared error, standardised error, covariance determinant and score with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Standard deviation must be positive. Complete the transfer step: Validate. Change exactly one interval endpoint, alpha level, outcome, median, quantile grid, interval weight, normalisation, predictive mean, variance, covariance, threshold, weight function, ensemble value or case weight. Explain which mathematical object changes, which score term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving lucky coverage, excessive width, quantile crossing, unmatched grids, scale domination, overconfidence, singular covariance, identical moments with different tails, threshold cherry-picking or regional improvement with global deterioration. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 4: A dispersion term

Begin with this mechanism: Compute log variance. Recreate the concrete case: log(4)≈1.3863. Write outcome units, predictive mean, standard deviation, variance, log variance, squared error, standardised error, covariance determinant and score with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Log base and constants must be consistent. Complete the transfer step: Use natural log. Change exactly one interval endpoint, alpha level, outcome, median, quantile grid, interval weight, normalisation, predictive mean, variance, covariance, threshold, weight function, ensemble value or case weight. Explain which mathematical object changes, which score term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving lucky coverage, excessive width, quantile crossing, unmatched grids, scale domination, overconfidence, singular covariance, identical moments with different tails, threshold cherry-picking or regional improvement with global deterioration. State the extra evidence that 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: A error term

Begin with this mechanism: Standardise squared error. Recreate the concrete case: 9/4=2.25. Write outcome units, predictive mean, standard deviation, variance, log variance, squared error, standardised error, covariance determinant and score with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Units cancel only after variance scaling. Complete the transfer step: Show units. Change exactly one interval endpoint, alpha level, outcome, median, quantile grid, interval weight, normalisation, predictive mean, variance, covariance, threshold, weight function, ensemble value or case weight. Explain which mathematical object changes, which score term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving lucky coverage, excessive width, quantile crossing, unmatched grids, scale domination, overconfidence, singular covariance, identical moments with different tails, threshold cherry-picking or regional improvement with global deterioration. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 6: A total

Begin with this mechanism: Add both terms. Recreate the concrete case: DSS≈3.6363. Write outcome units, predictive mean, standard deviation, variance, log variance, squared error, standardised error, covariance determinant and score with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Lower is better under loss orientation. Complete the transfer step: Label orientation. Change exactly one interval endpoint, alpha level, outcome, median, quantile grid, interval weight, normalisation, predictive mean, variance, covariance, threshold, weight function, ensemble value or case weight. Explain which mathematical object changes, which score term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving lucky coverage, excessive width, quantile crossing, unmatched grids, scale domination, overconfidence, singular covariance, identical moments with different tails, threshold cherry-picking or regional improvement with global deterioration. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 7: Forecast B

Begin with this mechanism: Use mu=11 and sigma=1. Recreate the concrete case: DSS=0+4=4. Write outcome units, predictive mean, standard deviation, variance, log variance, squared error, standardised error, covariance determinant and score with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: A closer mean can lose if too confident. Complete the transfer step: Compare. Change exactly one interval endpoint, alpha level, outcome, median, quantile grid, interval weight, normalisation, predictive mean, variance, covariance, threshold, weight function, ensemble value or case weight. Explain which mathematical object changes, which score term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving lucky coverage, excessive width, quantile crossing, unmatched grids, scale domination, overconfidence, singular covariance, identical moments with different tails, threshold cherry-picking or regional improvement with global deterioration. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 8: Forecast C

Begin with this mechanism: Use mu=13 and sigma=5. Recreate the concrete case: DSS=log(25)≈3.2189. Write outcome units, predictive mean, standard deviation, variance, log variance, squared error, standardised error, covariance determinant and score with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: An exact mean on one case may be luck. Complete the transfer step: Repeat cases. Change exactly one interval endpoint, alpha level, outcome, median, quantile grid, interval weight, normalisation, predictive mean, variance, covariance, threshold, weight function, ensemble value or case weight. Explain which mathematical object changes, which score term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving lucky coverage, excessive width, quantile crossing, unmatched grids, scale domination, overconfidence, singular covariance, identical moments with different tails, threshold cherry-picking or regional improvement with global deterioration. State the extra evidence that 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: Repeated evaluation

Begin with this mechanism: Average over matched cases. Recreate the concrete case: Compare systems on identical targets. Write outcome units, predictive mean, standard deviation, variance, log variance, squared error, standardised error, covariance determinant and score with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Case selection changes results. Complete the transfer step: Freeze sample. Change exactly one interval endpoint, alpha level, outcome, median, quantile grid, interval weight, normalisation, predictive mean, variance, covariance, threshold, weight function, ensemble value or case weight. Explain which mathematical object changes, which score term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving lucky coverage, excessive width, quantile crossing, unmatched grids, scale domination, overconfidence, singular covariance, identical moments with different tails, threshold cherry-picking or regional improvement with global deterioration. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 10: Moment scope

Begin with this mechanism: Recognise use of mean and variance only. Recreate the concrete case: Shape is not scored directly. Write outcome units, predictive mean, standard deviation, variance, log variance, squared error, standardised error, covariance determinant and score with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Equal moments can hide different tails. Complete the transfer step: Add diagnostics. Change exactly one interval endpoint, alpha level, outcome, median, quantile grid, interval weight, normalisation, predictive mean, variance, covariance, threshold, weight function, ensemble value or case weight. Explain which mathematical object changes, which score term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving lucky coverage, excessive width, quantile crossing, unmatched grids, scale domination, overconfidence, singular covariance, identical moments with different tails, threshold cherry-picking or regional improvement with global deterioration. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 11: Multivariate mean

Begin with this mechanism: Use a vector of target means. Recreate the concrete case: Align coordinates and units. Write outcome units, predictive mean, standard deviation, variance, log variance, squared error, standardised error, covariance determinant and score with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Mixing scales distorts covariance. Complete the transfer step: Standardise thoughtfully. Change exactly one interval endpoint, alpha level, outcome, median, quantile grid, interval weight, normalisation, predictive mean, variance, covariance, threshold, weight function, ensemble value or case weight. Explain which mathematical object changes, which score term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving lucky coverage, excessive width, quantile crossing, unmatched grids, scale domination, overconfidence, singular covariance, identical moments with different tails, threshold cherry-picking or regional improvement with global deterioration. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 12: Covariance

Begin with this mechanism: Estimate positive-definite Sigma. Recreate the concrete case: Use log determinant and inverse. Write outcome units, predictive mean, standard deviation, variance, log variance, squared error, standardised error, covariance determinant and score with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Singularity makes the score unstable. Complete the transfer step: Check eigenvalues. Change exactly one interval endpoint, alpha level, outcome, median, quantile grid, interval weight, normalisation, predictive mean, variance, covariance, threshold, weight function, ensemble value or case weight. Explain which mathematical object changes, which score term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving lucky coverage, excessive width, quantile crossing, unmatched grids, scale domination, overconfidence, singular covariance, identical moments with different tails, threshold cherry-picking or regional improvement with global deterioration. State the extra evidence that 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: Correlation

Begin with this mechanism: Keep cross-target dependence. Recreate the concrete case: Off-diagonal terms affect Mahalanobis error. Write outcome units, predictive mean, standard deviation, variance, log variance, squared error, standardised error, covariance determinant and score with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Marginal scores miss dependence. Complete the transfer step: Report joint result. Change exactly one interval endpoint, alpha level, outcome, median, quantile grid, interval weight, normalisation, predictive mean, variance, covariance, threshold, weight function, ensemble value or case weight. Explain which mathematical object changes, which score term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving lucky coverage, excessive width, quantile crossing, unmatched grids, scale domination, overconfidence, singular covariance, identical moments with different tails, threshold cherry-picking or regional improvement with global deterioration. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 14: Regularisation

Begin with this mechanism: Stabilise covariance when dimension is high. Recreate the concrete case: Declare shrinkage or structure. Write outcome units, predictive mean, standard deviation, variance, log variance, squared error, standardised error, covariance determinant and score with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Regularisation changes the evaluated forecast. Complete the transfer step: Sensitivity check. Change exactly one interval endpoint, alpha level, outcome, median, quantile grid, interval weight, normalisation, predictive mean, variance, covariance, threshold, weight function, ensemble value or case weight. Explain which mathematical object changes, which score term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving lucky coverage, excessive width, quantile crossing, unmatched grids, scale domination, overconfidence, singular covariance, identical moments with different tails, threshold cherry-picking or regional improvement with global deterioration. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 15: Shape check

Begin with this mechanism: Inspect quantiles and tail events separately. Recreate the concrete case: Compare with a fuller distributional score. Write outcome units, predictive mean, standard deviation, variance, log variance, squared error, standardised error, covariance determinant and score with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: DSS cannot certify multimodality or skewness. Complete the transfer step: Use companion evidence. Change exactly one interval endpoint, alpha level, outcome, median, quantile grid, interval weight, normalisation, predictive mean, variance, covariance, threshold, weight function, ensemble value or case weight. Explain which mathematical object changes, which score term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving lucky coverage, excessive width, quantile crossing, unmatched grids, scale domination, overconfidence, singular covariance, identical moments with different tails, threshold cherry-picking or regional improvement with global deterioration. State the extra evidence that 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 moment-based scope and formula. Recreate the concrete case: Teaching scores are 3.6363,4 and 3.2189. Write outcome units, predictive mean, standard deviation, variance, log variance, squared error, standardised error, covariance determinant and score with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Do not claim full-distribution correctness. Complete the transfer step: Bound conclusion. Change exactly one interval endpoint, alpha level, outcome, median, quantile grid, interval weight, normalisation, predictive mean, variance, covariance, threshold, weight function, ensemble value or case weight. Explain which mathematical object changes, which score term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving lucky coverage, excessive width, quantile crossing, unmatched grids, scale domination, overconfidence, singular covariance, identical moments with different tails, threshold cherry-picking or regional improvement with global deterioration. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.