Skip to content
Time-Series
EconometricsIan Helfrich PhD ↗

Lecture 14 summary · April 23, 2027

Freeze a revision before its next test

Complete chapter · Laboratory ch14 · Practice · All summaries

The economic question

A revision is a new empirical argument. Its selection record, update rule, fallback, and untouched next comparison must be preserved together.

Central relation

m^=arg⁡min⁡m∈MLvalidation(m),\widehat m=\arg\min_{m\in\mathcal M}L_{\mathrm{validation}}(m),
Lcontinuation(m^) L_{\mathrm{continuation}}(\widehat m)

Individual problem

Validation losses are 1.3, 1.1, 1.2 and continuation losses 1.0, 1.4, 0.9. Which procedure was selected and what is its continuation result?

Worked reasoning

The second procedure was selected and its continuation loss is 1.4. Choosing the third after seeing continuation creates a new selection step, requiring a later test.

A consequential failure

Only coefficients are frozen while features, cleaning, window length, or fallback change after continuation is observed.

Repair

Freeze the whole algorithm and data hash. Record allowed refits, and label post-continuation alterations as a new version.