WHY MATHEMATICS?

Why Mathematics? | Quantile Factor Models: Low-Rank Check Loss, Tail-Specific Latent Factors, Heavy-Tailed Panels and Rank Selection

How a large panel can contain common structure in its lower, middle and upper outcomes, why factor spaces change with tau, and what stability checks prevent over-interpretation.

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 the high-dimensional panel and target quantile explicitly
Estimate low-rank loadings and factor paths with check loss
Choose factor count and align spaces across quantiles
Validate reconstruction, forecasts, starts and subperiod stability

Chapter map

1. Panel
2. Target quantile
3. Rank
4. Loadings
5. Factor path
6. Reconstruction
7. Residuals
8. Perturbation
9. Median loss
10. Tail factors
11. Factor count
12. Heavy tails
13. Missing cells
14. Computation
15. Validation
16. Claim

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

1. Panel

A small example makes the mechanism visible. Define units i and times t.

A concrete situation

Use a 2 by 3 matrix.

What the mathematics does not license

Rows and columns must not be swapped. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Record dimensions. 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. Target quantile

This is where modelling choices begin to affect evidence. Choose tau=.5.

A concrete situation

Fit a median factor.

What the mathematics does not license

Another tau can need another factor space. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Repeat 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.

3. Rank

The central mathematical idea is concrete. Start with one latent factor.

A concrete situation

Set r=1.

What the mathematics does not license

Rank is a modelling choice. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Use criteria. 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. Loadings

Here the geometry becomes operational. Use lambda=(2,1).

A concrete situation

First series loads twice as much.

What the mathematics does not license

Scale is not identified alone. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Normalise. 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. Factor path

This step deserves a careful pause. Use f=(1,2,3).

A concrete situation

Factor rises through time.

What the mathematics does not license

Sign can flip with loadings. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Align 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.

6. Reconstruction

The useful question is not only how to calculate, but what the calculation assumes. Multiply loadings by factor path.

A concrete situation

Rows are (2,4,6) and (1,2,3).

What the mathematics does not license

Exact fit is constructed. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Label teaching case. 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. Residuals

A small example makes the mechanism visible. Subtract reconstruction from data.

A concrete situation

All residuals are zero.

What the mathematics does not license

Zero training loss can overfit. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Hold out cells. 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. Perturbation

This is where modelling choices begin to affect evidence. Change first row final value to 9.

A concrete situation

Residual becomes 3.

What the mathematics does not license

One cell may be noise or a regime. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Investigate source. 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. Median loss

The central mathematical idea is concrete. Apply .5 absolute residual.

A concrete situation

Added loss is 1.5.

What the mathematics does not license

Loss scale grows with panel size. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Normalise comparisons. 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. Tail factors

Here the geometry becomes operational. Refit at high tau.

A concrete situation

Allow different loadings and factors.

What the mathematics does not license

Spaces across tau need alignment. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Track continuity. 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. Factor count

This step deserves a careful pause. Use information or rank-minimisation criteria.

A concrete situation

Compare r=1 and r=2.

What the mathematics does not license

Selection uncertainty remains. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Report paths. 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. Heavy tails

The useful question is not only how to calculate, but what the calculation assumes. Use quantile loss instead of squares.

A concrete situation

Limit influence of extreme residual magnitude.

What the mathematics does not license

Leverage can still dominate factors. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Check 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.

13. Missing cells

A small example makes the mechanism visible. Define estimation under missingness.

A concrete situation

Mask unavailable entries.

What the mathematics does not license

Informative missingness biases factors. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Model or qualify. 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. Computation

This is where modelling choices begin to affect evidence. Alternate factor and loading updates.

A concrete situation

Use multiple starts.

What the mathematics does not license

Nonconvex joint problem has local optima. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Compare solutions. 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. Validation

The central mathematical idea is concrete. Forecast future periods and reconstruct held-out cells.

A concrete situation

Use check loss by tau.

What the mathematics does not license

Random cells can leak time structure. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Use blocked tests. 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 tau, rank, normalisation and stability.

A concrete situation

Teaching perturbation loss is 1.5.

What the mathematics does not license

Do not name a latent factor as cause. 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 factor models compress a large panel into latent factors that can influence different parts of the data distribution. Classical approximate factor models target conditional means, while Chen, Dolado and Gonzalo use quantile regression to estimate quantile-dependent factors and loadings. Their quantile factor analysis remains meaningful under heavy-tailed idiosyncratic errors and includes model-selection criteria for the number of factors at each quantile.

Consider a two-series by three-time matrix with rows (2,4,6) and (1,2,3). Loadings (2,1) multiplied by factor path (1,2,3) reconstruct every cell, so the rank-one median check loss is zero. If the final value in the first row becomes 9, the same reconstruction leaves residual 3 and adds median check loss 1.5. A tail-specific factor or loading may explain that change, but a single altered cell does not prove another population factor.

Factor signs, scales and rotations are not unique without normalisation. Choosing too many factors fits noise; too few forces different mechanisms into one latent series. Quantile-specific factors estimated at nearby tau values may not line up automatically, and a factor that helps upper-tail prediction is not necessarily an economic or scientific cause. Cross-sectional and time dependence affect uncertainty.

Report panel dimensions, missing-data treatment, scaling, tau levels, factor-number criterion, normalisations, starts, convergence, smoothing, factor alignment and out-of-sample validation. Reconstruct held-out cells and future periods, compare mean and quantile factors, and test stability across subperiods. Keep latent statistical structure separate from named real-world mechanisms until independent evidence supports the interpretation.

Frequently asked questions

How does QFA differ from mean factor analysis?

It estimates factors and loadings for chosen conditional quantiles.

Why can tail factors differ?

Common shocks may shift upper or lower outcomes without moving the centre equally.

How is factor count selected?

With declared information or rank-minimisation criteria evaluated for each quantile.

Are factors uniquely named?

No. Scale, sign and rotation normalisations are required.

Why are heavy tails relevant?

Check loss does not square large residuals, though leverage and dependence still matter.

What must be reported?

Panel, tau, rank, normalisation, starts, smoothing, alignment and validation.

Primary source and fact-check note

Liang Chen, Juan J. Dolado and Jesús Gonzalo, “Quantile Factor Models,” Econometrica 89(2), 875–910 (2021), DOI 10.3982/ECTA15746. 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 time-series vintages, sampling frequency, margins, pseudo-observations, panel dimensions, missing-data rules, running variable, cutoff and treatment assignment before estimation.

Objective audit

Write the cointegrating check loss, copula conditional-CDF plug-in, low-rank quantile-factor objective or local distribution-regression discontinuity estimand explicitly.

Tuning audit

Record tau grids, lag augmentation, long-run covariance bandwidth, copula family, vine order, factor count, starts, RD bandwidth, kernel, polynomial order and CDF grid.

Computation audit

Check residual integration, correction terms, response weights, copula likelihood, alternating-factor convergence, local CDF monotonicity, first-stage jumps and numerical tolerance.

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 time breaks, marginal and joint tails, concentrated copula weights, sparse panel cells, quantile-specific factor support and observations immediately around the cutoff.

Uncertainty audit

Propagate long-run dependence, marginal estimation, copula fitting, factor selection, nonconvex optimisation, local bandwidth and CDF inversion uncertainty through the complete workflow.

Validation audit

Use future periods, held-out units, blocked panels or bandwidth-stable cutoff windows for quantile hits, weighted-CDF calibration, reconstruction loss and distributional effect sensitivity.

Final claim

Translate the result without turning cointegration, copula dependence, a latent factor or a local discontinuity effect into a universal 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: Panel

Begin with this mechanism: Define units i and times t. Recreate the concrete case: Use a 2 by 3 matrix. Write panel units, time periods, response units, quantile levels, factor counts, factor values, loading values, reconstructed cells, residuals, check losses and forecast origins with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Rows and columns must not be swapped. Complete the transfer step: Record dimensions. Change exactly one observation, series transformation, long-run slope, residual, tau level, lag, correction bandwidth, marginal rank, copula family, query weight, response value, factor count, loading, factor path, missing panel cell, running-variable value, cutoff, bandwidth, kernel weight, treatment jump, CDF threshold or validation block. 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 spurious trends, structural breaks, misspecified margins, tail-dependence failure, concentrated weights, factor rotation, local optima, cutoff manipulation, weak first stages or simultaneous policy changes. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 2: Target quantile

Begin with this mechanism: Choose tau=.5. Recreate the concrete case: Fit a median factor. Write panel units, time periods, response units, quantile levels, factor counts, factor values, loading values, reconstructed cells, residuals, check losses and forecast origins with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Another tau can need another factor space. Complete the transfer step: Repeat grid. Change exactly one observation, series transformation, long-run slope, residual, tau level, lag, correction bandwidth, marginal rank, copula family, query weight, response value, factor count, loading, factor path, missing panel cell, running-variable value, cutoff, bandwidth, kernel weight, treatment jump, CDF threshold or validation block. 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 spurious trends, structural breaks, misspecified margins, tail-dependence failure, concentrated weights, factor rotation, local optima, cutoff manipulation, weak first stages or simultaneous policy changes. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 3: Rank

Begin with this mechanism: Start with one latent factor. Recreate the concrete case: Set r=1. Write panel units, time periods, response units, quantile levels, factor counts, factor values, loading values, reconstructed cells, residuals, check losses and forecast origins with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Rank is a modelling choice. Complete the transfer step: Use criteria. Change exactly one observation, series transformation, long-run slope, residual, tau level, lag, correction bandwidth, marginal rank, copula family, query weight, response value, factor count, loading, factor path, missing panel cell, running-variable value, cutoff, bandwidth, kernel weight, treatment jump, CDF threshold or validation block. 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 spurious trends, structural breaks, misspecified margins, tail-dependence failure, concentrated weights, factor rotation, local optima, cutoff manipulation, weak first stages or simultaneous policy changes. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 4: Loadings

Begin with this mechanism: Use lambda=(2,1). Recreate the concrete case: First series loads twice as much. Write panel units, time periods, response units, quantile levels, factor counts, factor values, loading values, reconstructed cells, residuals, check losses and forecast origins with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Scale is not identified alone. Complete the transfer step: Normalise. Change exactly one observation, series transformation, long-run slope, residual, tau level, lag, correction bandwidth, marginal rank, copula family, query weight, response value, factor count, loading, factor path, missing panel cell, running-variable value, cutoff, bandwidth, kernel weight, treatment jump, CDF threshold or validation block. 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 spurious trends, structural breaks, misspecified margins, tail-dependence failure, concentrated weights, factor rotation, local optima, cutoff manipulation, weak first stages or simultaneous policy changes. 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: Factor path

Begin with this mechanism: Use f=(1,2,3). Recreate the concrete case: Factor rises through time. Write panel units, time periods, response units, quantile levels, factor counts, factor values, loading values, reconstructed cells, residuals, check losses and forecast origins with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Sign can flip with loadings. Complete the transfer step: Align orientation. Change exactly one observation, series transformation, long-run slope, residual, tau level, lag, correction bandwidth, marginal rank, copula family, query weight, response value, factor count, loading, factor path, missing panel cell, running-variable value, cutoff, bandwidth, kernel weight, treatment jump, CDF threshold or validation block. 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 spurious trends, structural breaks, misspecified margins, tail-dependence failure, concentrated weights, factor rotation, local optima, cutoff manipulation, weak first stages or simultaneous policy changes. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 6: Reconstruction

Begin with this mechanism: Multiply loadings by factor path. Recreate the concrete case: Rows are (2,4,6) and (1,2,3). Write panel units, time periods, response units, quantile levels, factor counts, factor values, loading values, reconstructed cells, residuals, check losses and forecast origins with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Exact fit is constructed. Complete the transfer step: Label teaching case. Change exactly one observation, series transformation, long-run slope, residual, tau level, lag, correction bandwidth, marginal rank, copula family, query weight, response value, factor count, loading, factor path, missing panel cell, running-variable value, cutoff, bandwidth, kernel weight, treatment jump, CDF threshold or validation block. 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 spurious trends, structural breaks, misspecified margins, tail-dependence failure, concentrated weights, factor rotation, local optima, cutoff manipulation, weak first stages or simultaneous policy changes. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 7: Residuals

Begin with this mechanism: Subtract reconstruction from data. Recreate the concrete case: All residuals are zero. Write panel units, time periods, response units, quantile levels, factor counts, factor values, loading values, reconstructed cells, residuals, check losses and forecast origins with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Zero training loss can overfit. Complete the transfer step: Hold out cells. Change exactly one observation, series transformation, long-run slope, residual, tau level, lag, correction bandwidth, marginal rank, copula family, query weight, response value, factor count, loading, factor path, missing panel cell, running-variable value, cutoff, bandwidth, kernel weight, treatment jump, CDF threshold or validation block. 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 spurious trends, structural breaks, misspecified margins, tail-dependence failure, concentrated weights, factor rotation, local optima, cutoff manipulation, weak first stages or simultaneous policy changes. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 8: Perturbation

Begin with this mechanism: Change first row final value to 9. Recreate the concrete case: Residual becomes 3. Write panel units, time periods, response units, quantile levels, factor counts, factor values, loading values, reconstructed cells, residuals, check losses and forecast origins with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: One cell may be noise or a regime. Complete the transfer step: Investigate source. Change exactly one observation, series transformation, long-run slope, residual, tau level, lag, correction bandwidth, marginal rank, copula family, query weight, response value, factor count, loading, factor path, missing panel cell, running-variable value, cutoff, bandwidth, kernel weight, treatment jump, CDF threshold or validation block. 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 spurious trends, structural breaks, misspecified margins, tail-dependence failure, concentrated weights, factor rotation, local optima, cutoff manipulation, weak first stages or simultaneous policy changes. 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: Median loss

Begin with this mechanism: Apply .5 absolute residual. Recreate the concrete case: Added loss is 1.5. Write panel units, time periods, response units, quantile levels, factor counts, factor values, loading values, reconstructed cells, residuals, check losses and forecast origins with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Loss scale grows with panel size. Complete the transfer step: Normalise comparisons. Change exactly one observation, series transformation, long-run slope, residual, tau level, lag, correction bandwidth, marginal rank, copula family, query weight, response value, factor count, loading, factor path, missing panel cell, running-variable value, cutoff, bandwidth, kernel weight, treatment jump, CDF threshold or validation block. 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 spurious trends, structural breaks, misspecified margins, tail-dependence failure, concentrated weights, factor rotation, local optima, cutoff manipulation, weak first stages or simultaneous policy changes. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 10: Tail factors

Begin with this mechanism: Refit at high tau. Recreate the concrete case: Allow different loadings and factors. Write panel units, time periods, response units, quantile levels, factor counts, factor values, loading values, reconstructed cells, residuals, check losses and forecast origins with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Spaces across tau need alignment. Complete the transfer step: Track continuity. Change exactly one observation, series transformation, long-run slope, residual, tau level, lag, correction bandwidth, marginal rank, copula family, query weight, response value, factor count, loading, factor path, missing panel cell, running-variable value, cutoff, bandwidth, kernel weight, treatment jump, CDF threshold or validation block. 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 spurious trends, structural breaks, misspecified margins, tail-dependence failure, concentrated weights, factor rotation, local optima, cutoff manipulation, weak first stages or simultaneous policy changes. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 11: Factor count

Begin with this mechanism: Use information or rank-minimisation criteria. Recreate the concrete case: Compare r=1 and r=2. Write panel units, time periods, response units, quantile levels, factor counts, factor values, loading values, reconstructed cells, residuals, check losses and forecast origins with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Selection uncertainty remains. Complete the transfer step: Report paths. Change exactly one observation, series transformation, long-run slope, residual, tau level, lag, correction bandwidth, marginal rank, copula family, query weight, response value, factor count, loading, factor path, missing panel cell, running-variable value, cutoff, bandwidth, kernel weight, treatment jump, CDF threshold or validation block. 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 spurious trends, structural breaks, misspecified margins, tail-dependence failure, concentrated weights, factor rotation, local optima, cutoff manipulation, weak first stages or simultaneous policy changes. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 12: Heavy tails

Begin with this mechanism: Use quantile loss instead of squares. Recreate the concrete case: Limit influence of extreme residual magnitude. Write panel units, time periods, response units, quantile levels, factor counts, factor values, loading values, reconstructed cells, residuals, check losses and forecast origins with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Leverage can still dominate factors. Complete the transfer step: Check units. Change exactly one observation, series transformation, long-run slope, residual, tau level, lag, correction bandwidth, marginal rank, copula family, query weight, response value, factor count, loading, factor path, missing panel cell, running-variable value, cutoff, bandwidth, kernel weight, treatment jump, CDF threshold or validation block. 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 spurious trends, structural breaks, misspecified margins, tail-dependence failure, concentrated weights, factor rotation, local optima, cutoff manipulation, weak first stages or simultaneous policy changes. 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: Missing cells

Begin with this mechanism: Define estimation under missingness. Recreate the concrete case: Mask unavailable entries. Write panel units, time periods, response units, quantile levels, factor counts, factor values, loading values, reconstructed cells, residuals, check losses and forecast origins with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Informative missingness biases factors. Complete the transfer step: Model or qualify. Change exactly one observation, series transformation, long-run slope, residual, tau level, lag, correction bandwidth, marginal rank, copula family, query weight, response value, factor count, loading, factor path, missing panel cell, running-variable value, cutoff, bandwidth, kernel weight, treatment jump, CDF threshold or validation block. 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 spurious trends, structural breaks, misspecified margins, tail-dependence failure, concentrated weights, factor rotation, local optima, cutoff manipulation, weak first stages or simultaneous policy changes. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 14: Computation

Begin with this mechanism: Alternate factor and loading updates. Recreate the concrete case: Use multiple starts. Write panel units, time periods, response units, quantile levels, factor counts, factor values, loading values, reconstructed cells, residuals, check losses and forecast origins with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Nonconvex joint problem has local optima. Complete the transfer step: Compare solutions. Change exactly one observation, series transformation, long-run slope, residual, tau level, lag, correction bandwidth, marginal rank, copula family, query weight, response value, factor count, loading, factor path, missing panel cell, running-variable value, cutoff, bandwidth, kernel weight, treatment jump, CDF threshold or validation block. 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 spurious trends, structural breaks, misspecified margins, tail-dependence failure, concentrated weights, factor rotation, local optima, cutoff manipulation, weak first stages or simultaneous policy changes. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 15: Validation

Begin with this mechanism: Forecast future periods and reconstruct held-out cells. Recreate the concrete case: Use check loss by tau. Write panel units, time periods, response units, quantile levels, factor counts, factor values, loading values, reconstructed cells, residuals, check losses and forecast origins with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Random cells can leak time structure. Complete the transfer step: Use blocked tests. Change exactly one observation, series transformation, long-run slope, residual, tau level, lag, correction bandwidth, marginal rank, copula family, query weight, response value, factor count, loading, factor path, missing panel cell, running-variable value, cutoff, bandwidth, kernel weight, treatment jump, CDF threshold or validation block. 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 spurious trends, structural breaks, misspecified margins, tail-dependence failure, concentrated weights, factor rotation, local optima, cutoff manipulation, weak first stages or simultaneous policy changes. 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 tau, rank, normalisation and stability. Recreate the concrete case: Teaching perturbation loss is 1.5. Write panel units, time periods, response units, quantile levels, factor counts, factor values, loading values, reconstructed cells, residuals, check losses and forecast origins with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Do not name a latent factor as cause. Complete the transfer step: Bound conclusion. Change exactly one observation, series transformation, long-run slope, residual, tau level, lag, correction bandwidth, marginal rank, copula family, query weight, response value, factor count, loading, factor path, missing panel cell, running-variable value, cutoff, bandwidth, kernel weight, treatment jump, CDF threshold or validation block. 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 spurious trends, structural breaks, misspecified margins, tail-dependence failure, concentrated weights, factor rotation, local optima, cutoff manipulation, weak first stages or simultaneous policy changes. State what extra measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.