WHY MATHEMATICS?

Why Mathematics? | Quantile Regression Averaging: Forecast Pools, Tau-Specific Weights, Prediction Intervals and Rolling Calibration

How separate quantile regressions combine several point forecasts into lower, median and upper predictions, and why time-respecting evaluation determines whether the interval is useful.

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

Build an auditable pool of forecasts available at each origin
Estimate separate combination weights for every target quantile
Keep lower and upper outputs ordered and supported
Evaluate rolling check loss, interval coverage, width and regime stability

Chapter map

1. Forecast origins
2. Component forecasts
3. Median weights
4. Median output
5. Lower weights
6. Lower output
7. Upper weights
8. Upper output
9. Interval
10. Realisation
11. Lower loss
12. Upper loss
13. Rolling window
14. Constraints
15. Calibration
16. Claim

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

1. Forecast origins

A small example makes the mechanism visible. Define when every input forecast was available.

A concrete situation

Use three model forecasts.

What the mathematics does not license

Revised data cause leakage. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Freeze vintages. 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. Component forecasts

This is where modelling choices begin to affect evidence. Collect 10,12,14.

A concrete situation

Keep model identities.

What the mathematics does not license

Highly correlated forecasts reduce coefficient stability. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Audit pool. 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. Median weights

The central mathematical idea is concrete. Use .2,.5,.3 for illustration.

A concrete situation

They sum to one.

What the mathematics does not license

Real QRA weights need not be convex. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

State constraints. 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. Median output

Here the geometry becomes operational. Compute 2+6+4.2.

A concrete situation

Get 12.2.

What the mathematics does not license

A weighted point is not automatically a calibrated median. Evidence becomes stronger when the limitation is tested instead of hidden.

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.

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. Lower weights

This step deserves a careful pause. Use .5,.4,.1 at tau=.10.

A concrete situation

Emphasise lower forecasts.

What the mathematics does not license

Separate tau models may cross. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Check order. 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. Lower output

The useful question is not only how to calculate, but what the calculation assumes. Compute 5+4.8+1.4.

A concrete situation

Get 11.2.

What the mathematics does not license

Rounding changes interval width. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Keep precision. 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. Upper weights

A small example makes the mechanism visible. Use .1,.3,.6 at tau=.90.

A concrete situation

Emphasise upper forecast.

What the mathematics does not license

Large weights can amplify one weak model. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Stress test. 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. Upper output

This is where modelling choices begin to affect evidence. Compute 1+3.6+8.4.

A concrete situation

Get 13.0.

What the mathematics does not license

Ordered result is illustrative. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Enforce coherence if needed. 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. Interval

The central mathematical idea is concrete. Combine lower and upper outputs.

A concrete situation

Get [11.2,13.0].

What the mathematics does not license

Nominal width is not coverage. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Evaluate outcomes. 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. Realisation

Here the geometry becomes operational. Set y=13.

A concrete situation

It lies on upper endpoint.

What the mathematics does not license

One hit says nothing about long-run calibration. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Use many origins. 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. Lower loss

This step deserves a careful pause. Residual is 1.8 at tau=.10.

A concrete situation

Check loss is .18.

What the mathematics does not license

Lower loss alone ignores upper tail. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Score full 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.

12. Upper loss

The useful question is not only how to calculate, but what the calculation assumes. Residual is zero at tau=.90.

A concrete situation

Check loss is zero.

What the mathematics does not license

Perfect one-case loss can be chance. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Aggregate honestly. 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. Rolling window

A small example makes the mechanism visible. Estimate weights only from past origins.

A concrete situation

Update through time.

What the mathematics does not license

Random folds leak the future. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Respect sequence. 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. Constraints

This is where modelling choices begin to affect evidence. Choose unconstrained, nonnegative or sum-to-one weights.

A concrete situation

Match interpretation.

What the mathematics does not license

Constraints trade flexibility for stability. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Compare versions. 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. Calibration

The central mathematical idea is concrete. Check unconditional and conditional coverage plus width.

A concrete situation

Compare benchmarks.

What the mathematics does not license

Correct average coverage can hide regimes. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Stratify. 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 pool, origins, window, tau and constraints.

A concrete situation

Teaching interval is 11.2 to 13.0.

What the mathematics does not license

Do not guarantee the next outcome. 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

Quantile regression averaging combines forecasts by learning a separate regression for each target quantile. Nowotarski and Weron proposed QRA for electricity spot-price prediction intervals using a pool of point forecasts from individual models. Instead of averaging model outputs once and attaching a generic error band, QRA lets the lower, middle and upper conditional outcomes use different intercepts and forecast weights.

Suppose three models predict 10,12,14. An illustrative median weight vector .2,.5,.3 gives 12.2. Lower-.10 weights .5,.4,.1 give 11.2, while upper-.90 weights .1,.3,.6 give 13.0, forming an ordered interval [11.2,13.0]. If the realised outcome is 13, the lower-quantile residual is 1.8 and its check loss is .10(1.8)=.18, while the upper-quantile residual and loss are zero.

Real QRA weights are estimated from rolling or otherwise time-respecting training data; they need not be nonnegative or sum to one unless constraints impose that. Correlated model forecasts can make coefficients unstable even when combined predictions are useful. A separate fit at every tau can cross, and regime change can make yesterday’s best combination unreliable.

Report the component forecasts, information available at each forecast origin, rolling window, tau grid, intercept, constraints, regularisation, crossing repair and held-out coverage. Compare QRA against individual models and simple combinations. Performance in one electricity dataset does not establish universal superiority, and a prediction interval is not a guarantee.

Frequently asked questions

What is QRA?

A quantile regression that uses a pool of model forecasts as predictors.

Why estimate weights by tau?

Lower, central and upper outcomes may reward different combinations.

Must weights sum to one?

Only if the chosen constrained formulation requires it.

Why use rolling evaluation?

Forecasts must be trained without information from future origins.

What makes an interval useful?

Good held-out coverage, appropriate width and stable conditional calibration.

What must be reported?

Forecast pool, vintages, rolling window, tau grid, constraints, repair and benchmarks.

Primary source and fact-check note

Jakub Nowotarski and Rafał Weron, “Computing Electricity Spot Price Prediction Intervals Using Quantile Regression and Forecast Averaging,” Computational Statistics 30, 791–803 (2015), DOI 10.1007/s00180-014-0523-0. 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 latent and observed covariates, validation or replicate data, event definition, time origin, censoring, forecast origins, predictor scaling and train-test boundaries.

Objective audit

Write joint estimating equations, survival risk-set contributions, kernel check-loss penalties or tau-specific forecast-combination objectives explicitly.

Tuning audit

Record tau grid, latent integration grid, Monte Carlo draws, kernel bandwidth, RKHS penalty, rolling window, constraints and stopping tolerance.

Computation audit

Check noisy-proxy loss, EM-type convergence, risk-set counts, kernel conditioning, forecast vintages, coefficient instability and crossing.

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.

Support audit

Inspect validation coverage, censoring support, training distances, tail sample size, component forecast diversity and regime change.

Uncertainty audit

Propagate measurement, censoring, kernel tuning, weight estimation, resampling and repair uncertainty through the complete workflow.

Validation audit

Use untouched units or future origins for check loss, survival quantile calibration, interval coverage, width and subgroup diagnostics.

Final claim

Translate the conditional result without turning a corrected proxy, censored estimate, kernel similarity or forecast interval into a causal guarantee.

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: Forecast origins

Begin with this mechanism: Define when every input forecast was available. Recreate the concrete case: Use three model forecasts. Write forecast origins, model identities, point forecasts, tau-specific weights, intercepts, quantile outputs, interval endpoints, realisations, check losses, coverage and width with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Revised data cause leakage. Complete the transfer step: Freeze vintages. Change exactly one proxy error, validation pair, replicate, tau level, latent grid, event code, censoring time, risk-set member, kernel bandwidth, regularisation value, scaled distance, component forecast, weight constraint, rolling window, forecast origin 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, support, 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 differential measurement error, dependent censoring, unsupported survival tails, kernel extrapolation, correlated forecasts, regime change or subgroup miscalibration. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 2: Component forecasts

Begin with this mechanism: Collect 10,12,14. Recreate the concrete case: Keep model identities. Write forecast origins, model identities, point forecasts, tau-specific weights, intercepts, quantile outputs, interval endpoints, realisations, check losses, coverage and width with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Highly correlated forecasts reduce coefficient stability. Complete the transfer step: Audit pool. Change exactly one proxy error, validation pair, replicate, tau level, latent grid, event code, censoring time, risk-set member, kernel bandwidth, regularisation value, scaled distance, component forecast, weight constraint, rolling window, forecast origin 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, support, 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 differential measurement error, dependent censoring, unsupported survival tails, kernel extrapolation, correlated forecasts, regime change or subgroup miscalibration. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 3: Median weights

Begin with this mechanism: Use .2,.5,.3 for illustration. Recreate the concrete case: They sum to one. Write forecast origins, model identities, point forecasts, tau-specific weights, intercepts, quantile outputs, interval endpoints, realisations, check losses, coverage and width with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Real QRA weights need not be convex. Complete the transfer step: State constraints. Change exactly one proxy error, validation pair, replicate, tau level, latent grid, event code, censoring time, risk-set member, kernel bandwidth, regularisation value, scaled distance, component forecast, weight constraint, rolling window, forecast origin 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, support, 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 differential measurement error, dependent censoring, unsupported survival tails, kernel extrapolation, correlated forecasts, regime change or subgroup miscalibration. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 4: Median output

Begin with this mechanism: Compute 2+6+4.2. Recreate the concrete case: Get 12.2. Write forecast origins, model identities, point forecasts, tau-specific weights, intercepts, quantile outputs, interval endpoints, realisations, check losses, coverage and width with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: A weighted point is not automatically a calibrated median. Complete the transfer step: Validate. Change exactly one proxy error, validation pair, replicate, tau level, latent grid, event code, censoring time, risk-set member, kernel bandwidth, regularisation value, scaled distance, component forecast, weight constraint, rolling window, forecast origin 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, support, 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 differential measurement error, dependent censoring, unsupported survival tails, kernel extrapolation, correlated forecasts, regime change or subgroup miscalibration. State what extra 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: Lower weights

Begin with this mechanism: Use .5,.4,.1 at tau=.10. Recreate the concrete case: Emphasise lower forecasts. Write forecast origins, model identities, point forecasts, tau-specific weights, intercepts, quantile outputs, interval endpoints, realisations, check losses, coverage and width with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Separate tau models may cross. Complete the transfer step: Check order. Change exactly one proxy error, validation pair, replicate, tau level, latent grid, event code, censoring time, risk-set member, kernel bandwidth, regularisation value, scaled distance, component forecast, weight constraint, rolling window, forecast origin 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, support, 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 differential measurement error, dependent censoring, unsupported survival tails, kernel extrapolation, correlated forecasts, regime change or subgroup miscalibration. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 6: Lower output

Begin with this mechanism: Compute 5+4.8+1.4. Recreate the concrete case: Get 11.2. Write forecast origins, model identities, point forecasts, tau-specific weights, intercepts, quantile outputs, interval endpoints, realisations, check losses, coverage and width with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Rounding changes interval width. Complete the transfer step: Keep precision. Change exactly one proxy error, validation pair, replicate, tau level, latent grid, event code, censoring time, risk-set member, kernel bandwidth, regularisation value, scaled distance, component forecast, weight constraint, rolling window, forecast origin 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, support, 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 differential measurement error, dependent censoring, unsupported survival tails, kernel extrapolation, correlated forecasts, regime change or subgroup miscalibration. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 7: Upper weights

Begin with this mechanism: Use .1,.3,.6 at tau=.90. Recreate the concrete case: Emphasise upper forecast. Write forecast origins, model identities, point forecasts, tau-specific weights, intercepts, quantile outputs, interval endpoints, realisations, check losses, coverage and width with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Large weights can amplify one weak model. Complete the transfer step: Stress test. Change exactly one proxy error, validation pair, replicate, tau level, latent grid, event code, censoring time, risk-set member, kernel bandwidth, regularisation value, scaled distance, component forecast, weight constraint, rolling window, forecast origin 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, support, 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 differential measurement error, dependent censoring, unsupported survival tails, kernel extrapolation, correlated forecasts, regime change or subgroup miscalibration. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 8: Upper output

Begin with this mechanism: Compute 1+3.6+8.4. Recreate the concrete case: Get 13.0. Write forecast origins, model identities, point forecasts, tau-specific weights, intercepts, quantile outputs, interval endpoints, realisations, check losses, coverage and width with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Ordered result is illustrative. Complete the transfer step: Enforce coherence if needed. Change exactly one proxy error, validation pair, replicate, tau level, latent grid, event code, censoring time, risk-set member, kernel bandwidth, regularisation value, scaled distance, component forecast, weight constraint, rolling window, forecast origin 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, support, 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 differential measurement error, dependent censoring, unsupported survival tails, kernel extrapolation, correlated forecasts, regime change or subgroup miscalibration. State what extra 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: Interval

Begin with this mechanism: Combine lower and upper outputs. Recreate the concrete case: Get [11.2,13.0]. Write forecast origins, model identities, point forecasts, tau-specific weights, intercepts, quantile outputs, interval endpoints, realisations, check losses, coverage and width with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Nominal width is not coverage. Complete the transfer step: Evaluate outcomes. Change exactly one proxy error, validation pair, replicate, tau level, latent grid, event code, censoring time, risk-set member, kernel bandwidth, regularisation value, scaled distance, component forecast, weight constraint, rolling window, forecast origin 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, support, 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 differential measurement error, dependent censoring, unsupported survival tails, kernel extrapolation, correlated forecasts, regime change or subgroup miscalibration. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 10: Realisation

Begin with this mechanism: Set y=13. Recreate the concrete case: It lies on upper endpoint. Write forecast origins, model identities, point forecasts, tau-specific weights, intercepts, quantile outputs, interval endpoints, realisations, check losses, coverage and width with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: One hit says nothing about long-run calibration. Complete the transfer step: Use many origins. Change exactly one proxy error, validation pair, replicate, tau level, latent grid, event code, censoring time, risk-set member, kernel bandwidth, regularisation value, scaled distance, component forecast, weight constraint, rolling window, forecast origin 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, support, 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 differential measurement error, dependent censoring, unsupported survival tails, kernel extrapolation, correlated forecasts, regime change or subgroup miscalibration. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 11: Lower loss

Begin with this mechanism: Residual is 1.8 at tau=.10. Recreate the concrete case: Check loss is .18. Write forecast origins, model identities, point forecasts, tau-specific weights, intercepts, quantile outputs, interval endpoints, realisations, check losses, coverage and width with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Lower loss alone ignores upper tail. Complete the transfer step: Score full grid. Change exactly one proxy error, validation pair, replicate, tau level, latent grid, event code, censoring time, risk-set member, kernel bandwidth, regularisation value, scaled distance, component forecast, weight constraint, rolling window, forecast origin 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, support, 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 differential measurement error, dependent censoring, unsupported survival tails, kernel extrapolation, correlated forecasts, regime change or subgroup miscalibration. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 12: Upper loss

Begin with this mechanism: Residual is zero at tau=.90. Recreate the concrete case: Check loss is zero. Write forecast origins, model identities, point forecasts, tau-specific weights, intercepts, quantile outputs, interval endpoints, realisations, check losses, coverage and width with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Perfect one-case loss can be chance. Complete the transfer step: Aggregate honestly. Change exactly one proxy error, validation pair, replicate, tau level, latent grid, event code, censoring time, risk-set member, kernel bandwidth, regularisation value, scaled distance, component forecast, weight constraint, rolling window, forecast origin 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, support, 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 differential measurement error, dependent censoring, unsupported survival tails, kernel extrapolation, correlated forecasts, regime change or subgroup miscalibration. State what extra 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: Rolling window

Begin with this mechanism: Estimate weights only from past origins. Recreate the concrete case: Update through time. Write forecast origins, model identities, point forecasts, tau-specific weights, intercepts, quantile outputs, interval endpoints, realisations, check losses, coverage and width with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Random folds leak the future. Complete the transfer step: Respect sequence. Change exactly one proxy error, validation pair, replicate, tau level, latent grid, event code, censoring time, risk-set member, kernel bandwidth, regularisation value, scaled distance, component forecast, weight constraint, rolling window, forecast origin 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, support, 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 differential measurement error, dependent censoring, unsupported survival tails, kernel extrapolation, correlated forecasts, regime change or subgroup miscalibration. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 14: Constraints

Begin with this mechanism: Choose unconstrained, nonnegative or sum-to-one weights. Recreate the concrete case: Match interpretation. Write forecast origins, model identities, point forecasts, tau-specific weights, intercepts, quantile outputs, interval endpoints, realisations, check losses, coverage and width with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Constraints trade flexibility for stability. Complete the transfer step: Compare versions. Change exactly one proxy error, validation pair, replicate, tau level, latent grid, event code, censoring time, risk-set member, kernel bandwidth, regularisation value, scaled distance, component forecast, weight constraint, rolling window, forecast origin 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, support, 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 differential measurement error, dependent censoring, unsupported survival tails, kernel extrapolation, correlated forecasts, regime change or subgroup miscalibration. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 15: Calibration

Begin with this mechanism: Check unconditional and conditional coverage plus width. Recreate the concrete case: Compare benchmarks. Write forecast origins, model identities, point forecasts, tau-specific weights, intercepts, quantile outputs, interval endpoints, realisations, check losses, coverage and width with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Correct average coverage can hide regimes. Complete the transfer step: Stratify. Change exactly one proxy error, validation pair, replicate, tau level, latent grid, event code, censoring time, risk-set member, kernel bandwidth, regularisation value, scaled distance, component forecast, weight constraint, rolling window, forecast origin 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, support, 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 differential measurement error, dependent censoring, unsupported survival tails, kernel extrapolation, correlated forecasts, regime change or subgroup miscalibration. State what extra 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 pool, origins, window, tau and constraints. Recreate the concrete case: Teaching interval is 11.2 to 13.0. Write forecast origins, model identities, point forecasts, tau-specific weights, intercepts, quantile outputs, interval endpoints, realisations, check losses, coverage and width with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Do not guarantee the next outcome. Complete the transfer step: Bound conclusion. Change exactly one proxy error, validation pair, replicate, tau level, latent grid, event code, censoring time, risk-set member, kernel bandwidth, regularisation value, scaled distance, component forecast, weight constraint, rolling window, forecast origin 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, support, 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 differential measurement error, dependent censoring, unsupported survival tails, kernel extrapolation, correlated forecasts, regime change or subgroup miscalibration. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.