Skip to content
Time-Series
EconometricsIan Helfrich PhD ↗

Lecture 4 summary · February 5, 2027

Forecasts that survive a chronological comparison

Complete chapter · Laboratory ch04 · Practice · All summaries

The economic question

An equation is only part of a forecasting algorithm. Windows, horizons, refits, releases, and fallbacks determine its feasible behavior.

Central relation

ℓ(a,y)=cu(y−a)++co(a−y)+,\ell(a,y)=c_u(y-a)_++c_o(a-y)_+,
F(a∗)=cucu+co F(a^*)=\frac{c_u}{c_u+c_o}

Individual problem

Underprediction costs four and overprediction one. Which quantile is optimal? Compare model losses 2, 6, 1 with benchmark losses 4, 3, 5.

Worked reasoning

The optimal quantile is 0.8. Paired differences are negative two, three, and negative four, averaging negative one. Three targets support a local comparison, not a persistent superiority claim.

A consequential failure

Time is shuffled, selection uses test loss, or methods are evaluated on different dates.

Repair

Separate chronological selection and evaluation. Match targets and use an uncertainty method appropriate to dependent losses; with nested models and estimated parameters, the Diebold–Mariano normal reference need not apply.