P1 · Build journal
How the instructor build actually went
This page records the process behind the instructor build — decisions, dead ends included. Students rarely get to see an expert’s intermediate states; this is ours, kept honest.
Entry 1 — data archaeology
Planned inputs included EIA’s weekly gasoline product supplied — a real-economy volume series the BSE course also used. It is not on FRED (the id we guessed, WGFUPUS2, does not exist; EIA’s own API needs a key). Rather than add a second data provider on day one, we swapped in two weekly credit aggregates from the Fed’s H.8 release (TOTBKCR, CCLACBW027SBOG). Decision: keep the v1 build single-provider and keyless so anyone can run it from a clean checkout; revisit EIA for v2. Lesson: the data layer, not the model, is where practicum time actually goes — budget accordingly.
Entry 2 — the standardization trap
First complete draft: 52-week log differences, standardize over the full sample, first principal component, regression scaling. Everything ran; the tracker correlated 0.89 with the official WEI; the figures looked plausible at a glance. And the Great Recession was almost invisible — the tracker bottomed near +0.3% in early 2009 while the WEI read −3.5%.
The diagnosis took one table. Mid-recession z-scores, full-sample standardization, week of 2009-03-28:
| series | z-score |
|---|---|
| initial claims | −1.16 |
| continued claims | −1.38 |
| bank credit | −0.63 |
| consumer loans | +0.64 |
The worst labor market in a generation registered barely one standard deviation — because the standard deviation had been computed on a sample containing April 2020, when the same transformed series sat near eight standard deviations from normal. One episode had redefined the yardstick for every other episode. The regression scaling then compounded it: fit through COVID’s enormous swings, the slope flattened everything before 2020.
Fix: calibrate everything — means, variances, factor weights, GDP scaling — on the pre-2020 sample, then apply those fixed weights to the full history. Crises get measured on a normal-times yardstick instead of being allowed to compress it. The 2009 trough moved from +0.3% to −2.3%, against the WEI’s −3.3%; correlation rose to 0.90.
Lesson (and it generalizes): every standardization has a reference population, and choosing it is a modeling decision, not housekeeping. The official WEI’s own maintainers hit a version of this in 2020 and had to adapt their procedure. Session 6’s discussion of outliers and regime dependence starts exactly here.
Entry 3 — what the four series can and cannot see
With calibrated weights, the loadings tell a clean story: the factor is essentially the claims pair (0.69 each), with consumer loans a distant third (0.20) and bank credit near zero (−0.08) — banks’ balance sheets expand in panics as firms draw credit lines, so the series is nearly acyclical in yoy terms and the PCA correctly learns to ignore it. Our largest WEI divergence is the COVID trough (−11.4pp deeper than the WEI): when the factor is nearly all claims, a claims catastrophe is a tracker catastrophe. The official index’s retail, staffing, fuel, and electricity series pulled its trough up. Moral: breadth is not decoration — it is variance insurance for the weeks that matter most.
Open items for v2
- EIA fuel volumes (API key) and any public rail/electricity weekly series.
- Replace PCA with the Session 1 one-factor Kalman filter; compare factors.
- ALFRED claims vintages: recompute the tracker as of specific 2020 Thursdays — the real-time discipline this course preaches, applied to itself.