DC Escalation Scoreboard did the basis you carried hold?
Every contingency basis graded at every anchor, on vintage-true and final-revision legs. /escalation offers 5 bases to carry as a contingency factor. This grades the 3 that are rules — Long-run, Trailing 3yr, Current momentum — against what data-center construction escalation actually did. For every month we can reconstruct what the DC Build index actually read at the time, we compute what each basis would have told a reader to carry, and check it against what escalation actually did next. The metric is the one a capital program is judged on — did you carry enough — not the one a forecaster reaches for. The two hand-picked historical regimes /escalation also offers (the GFC downturn, the COVID peak) are shown further down, unscored: they were selected with hindsight, which makes them ungradeable by construction.
Cite
MacroGauge DC escalation grades (strict + extended legs), 2026-10-02, 2018-01=100, 290 vintage anchors — https://macrogauge.vercel.app/dc-scoreboardExport data
Paired grading: 3 rules × 4 horizons
Two legs, always shown together. The strict leg is vintage-true but its anchors begin 2018-01 and contain the 2021-22 spike with no downturn; the extended leg reaches back to 2010-12 on final-revision data, at a measured 0.672pp maximum distortion. Quoting either leg alone overstates how much the answer is known.
- Long-run under-provisioned 62.2% of 12-month windows on the vintage-true sample and 46.9% of 12-month windows on the deeper sample that includes a downturn.
- Trailing 3yr under-provisioned 41.1% of 12-month windows on the vintage-true sample and 41.2% of 12-month windows on the deeper sample that includes a downturn.
- Current momentum under-provisioned 50.0% of 12-month windows on the vintage-true sample and 54.2% of 12-month windows on the deeper sample that includes a downturn.
| Basis | Horizon | Strict — vintage-true (101 anchors) no downturn in sample | Extended — final-revision (189 anchors) includes a downturn | ||||
|---|---|---|---|---|---|---|---|
| Shortfall | Bias / MAE | Draws | Shortfall | Bias / MAE | Draws | ||
| Long-run | 12mo | 62.2% mean 6.22pp · worst 18.90pp | -3.57pp MAE 4.17pp | 7.5 | 46.9% mean 4.98pp · worst 18.88pp | -1.35pp MAE 3.33pp | 14.8 |
| Long-run | 24mo | 75.6% mean 5.13pp · worst 13.26pp | -3.75pp MAE 4.01pp | 3.3 | 53.9% mean 4.16pp · worst 13.35pp | -1.38pp MAE 3.10pp | 6.9 |
| Long-run | 36mo | Withheld — vintage-true sample too thin at this horizon | 64.7% mean 3.55pp · worst 9.05pp | -1.54pp MAE 3.06pp | 4.3 | ||
| Long-run | 48mo | Withheld — vintage-true sample too thin at this horizon | 65.2% mean 3.76pp · worst 7.01pp | -1.74pp MAE 3.17pp | 2.9 | ||
| Trailing 3yr | 12mo | 41.1% mean 7.54pp · worst 18.11pp | -0.64pp MAE 5.57pp | 7.5 | 41.2% mean 5.22pp · worst 18.00pp | -0.38pp MAE 3.93pp | 14.8 |
| Trailing 3yr | 24mo | 46.2% mean 6.12pp · worst 12.59pp | -0.47pp MAE 5.18pp | 3.3 | 43.6% mean 4.54pp · worst 12.66pp | -0.39pp MAE 3.58pp | 6.9 |
| Trailing 3yr | 36mo | Withheld — vintage-true sample too thin at this horizon | 55.6% mean 3.80pp · worst 8.35pp | -0.89pp MAE 3.33pp | 4.3 | ||
| Trailing 3yr | 48mo | Withheld — vintage-true sample too thin at this horizon | 66.0% mean 3.61pp · worst 6.45pp | -1.61pp MAE 3.14pp | 2.9 | ||
| Current momentum | 12mo | 50.0% mean 5.87pp · worst 19.22pp | -0.30pp MAE 5.57pp | 7.5 | 54.2% mean 3.92pp · worst 18.91pp | -0.23pp MAE 4.01pp | 14.8 |
| Current momentum | 24mo | 51.3% mean 6.37pp · worst 14.62pp | -0.01pp MAE 6.53pp | 3.3 | 53.9% mean 4.23pp · worst 14.89pp | -0.20pp MAE 4.36pp | 6.9 |
| Current momentum | 36mo | Withheld — vintage-true sample too thin at this horizon | 55.6% mean 3.83pp · worst 10.98pp | -0.24pp MAE 4.02pp | 4.3 | ||
| Current momentum | 48mo | Withheld — vintage-true sample too thin at this horizon | 70.2% mean 3.24pp · worst 8.94pp | -0.53pp MAE 4.03pp | 2.9 | ||
Independent draws fall as the horizon lengthens — consecutive monthly anchors overlap, so a longer horizon compresses more history into fewer genuinely separate windows. On the extended sample, the 48-month row is the thinnest, averaging 2.9 independent draws across the three bases: read its shortfall rate as a wide range of precedent, not a precise probability.
Expected vs realized — every vintage anchor
Each dot is one month the harness stood at, carried the basis forward, and then watched what the index did. Above the dashed line the basis ran short (red); below it the basis over-provisioned (green). This is the picture behind the shortfall rates in the table above.
90 anchors · shortfall in 62.2% of windows (mean 6.22pp, worst 18.90pp) · bias −3.57pp · MAE 4.17pp — recomputed from the dots, equal to the published grade for this cell.
The inversion
Of the three rolling bases, Long-run has the lowest mean absolute error on both samples — 4.09pp on the strict, vintage-true sample and 3.21pp on the extended sample. On the strict sample it is also the basis most likely to leave a reader short — a symmetric error metric rewards centering the error, not skewing it toward safety: its mean shortfall rate there is 68.9%, the highest of the three rolling bases. On the extended sample — deeper, and the one that actually contains a downturn — its mean shortfall rate is 50.4%, no longer the highest of the three (a different basis now is): the inversion attenuates once the sample includes a period escalation actually cooled.
Every mean in this section covers the 12- and 24-month horizons — the horizons both legs publish, and the only ones on which the two samples can be compared like for like (the strict leg withholds the longer ones as too thin). They are plain averages of the per-horizon figures in the table above, taken over that same set for each leg, so a reader can re-derive them from those cells by hand. The extended leg's longer horizons are graded in that table and deliberately left out of these means: averaging them in on one side only would flatter whichever leg reaches further.
Regimes carried on /escalation — ungradeable by design
/escalation also lets a reader carry either of these hand-picked historical regimes instead of a rolling rule. Both windows were chosen with hindsight, after the fact, from realized history — so neither is graded above or anywhere else on this page. They publish a rate and a window only: no shortfall rate, no MAE, no independent-draw count. Computing one would score hindsight against itself.
Window hand-picked in 2026 from realized history. Ungradeable: selecting it required hindsight, and restricting to anchors where the window had closed leaves too few independent draws to measure.
Window hand-picked in 2026 from realized history. Ungradeable: selecting it required hindsight, and restricting to anchors where the window had closed leaves too few independent draws to measure.
Lead-lag: do input-price moves forecast the index?
Verdict: A stable lead was found for 2 of 4 mappings by the pre-registered gate -- see caveats before treating this as forecasting evidence.
Gate: A lead counts only if the best lag agrees in sign and within 3 months across both sample halves. Stated before the numbers were computed.
- The recovered best lag is contemporaneous, not a lead (Electrical equipment -> Power & distribution transformers at 0 month(s)). This study exists to decide whether a forward model is buildable at 12-48 month horizons; a same-month correlation cannot support that at any gate strictness, and is equally consistent with both series reacting to a shared shock as with backlog-to-price transmission.
- A positive result under this gate is not guaranteed stable to where the sample happens to be split in half. Measured 2026-07-26: on a shallower 222-month sample the split midpoint fell near 2017-03 and this study's one currently-stable pairing FAILED the gate there -- its two halves genuinely disagreed (2008-2017 peaked at lag 24 months, r=0.327; 2017-2026 at lag 0, r=0.784). Deepening the sample to 1992 (402 months) moved the midpoint to roughly 2009-09, folding that entire disagreement inside one half where the gate no longer exercises it. The instability was moved out of view, not resolved -- and the gate was deliberately NOT strengthened after seeing this, because tuning a pre-registered test after it produces a positive is the exact failure mode stating it up front exists to prevent.
Conclusion: No forward model is warranted on this evidence.
45% of Build weight has a mapped input-price driver; 21% of Build weight cleared the pre-registered gate above (47% of the mapped set) — read the caveats before treating that as a usable lead.
| Driver | Component | Weight | Sample | Best lag | Correlation | Split-half lag (1st → 2nd) | Gate |
|---|---|---|---|---|---|---|---|
| Electrical equipment | Switchgear & switchboard | 14% | 404 mo 1993-01 – 2026-08 | 3mo | 0.543 | 8mo → 2mo | Not stable |
| Electrical equipment | Power & distribution transformers | 12% | 404 mo 1993-01 – 2026-08 | 0mo | 0.677 | 2mo → 0mo | Cleared 0-month lag — contemporaneous, not a lead (see caveats above) |
| Ventilation, heating & AC | AC & refrigeration equipment | 10% | 404 mo 1993-01 – 2026-08 | 8mo | 0.560 | 0mo → 9mo | Not stable |
| Turbines & generators | Generator sets & turbines | 9% | 404 mo 1993-01 – 2026-08 | 24mo | 0.303 | 24mo → 24mo | Cleared see caveats above before treating this as forecasting evidence |
Power nowcast: a fast read vs. the slow retail print
FAIL — a like-month year-ratio nowcast, backtested over 11 months of realized retail prints (as of 2026-07-01): best nowcast MAE 8.479pp (λ=0.25) vs. carry-forward MAE 5.463pp.
A like-month year-ratio nowcast, backtested against realized retail prints before letting it touch the index. It failed the pre-registered backtest gate -- the selected pass-through candidate must beat both naive baselines (simple carry-forward and zero pass-through) on MAE with every month's error inside the bound, and it did not -- so the ops index stays on official retail data and the machinery ships config-gated.
Methodology
Both legs price the DC Build index off ALFRED point-in-time vintages, whose raw release history for these twelve components reaches back to 2015-03. The strict leg is vintage-true (ALFRED as-of): each component takes its latest release known at the anchor date, but its anchors cannot start before 2018-01 regardless — a second, additional floor on top of that raw history, not a sign the underlying data runs out there: the index is based to that month, and an index based at its own base month cannot be reconstructed at a vintage that predates the base observation itself, however far back the raw releases go. So the strict leg's start is a conceptual constraint, not a data accident. Grading at a different base month would also grade a materially different index: this is a Laspeyres sum of separately rebased components, so its effective per-component weight is weight ÷ index-at-base, and that base constant does not cancel out of a weighted sum the way it would for a single series.
The extended leg is final-revision throughout: deeper sample, at a measured 0.672pp maximum distortion across every anchor month both legs share, reaching back to 2010-12. Substituting final-revision data for a real-time read understates how much a reader actually knew at the time — measured on this publish across every anchor month the two legs share, at most 0.672pp of distortion in a carried annualized rate, a figure re-derivable from the anchor rows in the raw artifact linked below. The deeper sample therefore publishes alongside the strict one rather than replacing it, with the distortion disclosed here rather than hidden.
Anchors dedupe by last-observation month: several ALFRED vintages can share one, when a release revises an old observation without extending the series. Grading every vintage would inflate both the anchor count and the independent-draw estimate without adding information, so each leg carries exactly one anchor per distinct last-observation month — the earliest vintage to reach it, since that is the first date a reader could actually have stood there.
The index graded here is reconstructed from official releases only. Every component is read from its published PPI/CES series and nothing else. The DC Build index on /datacenter and /escalation additionally splices a live futures tail onto Copper wire & cable and Aluminum mill shapes (8.5% of Build weight) past their last official print, so the two indexes agree in every month where that splice is inactive and differ where it is not — and the latest anchor, the month every basis above is read at, is such a month. Measured at 2026-08, the widest gap is Current momentum, which grades here at 9.24%/yr against the 9.32%/yr /escalation shows for the same rule. The statistics above are barely touched — only the handful of anchor-horizon pairs whose anchor falls in a splice month can differ at all — but the two numbers are not identical, and this page says so rather than leaving a reader to find it.
Receipts. Every figure on this page is re-derivable from the published artifact: /data/dc_grades.json carries all 290 anchor rows — for each anchor month and leg, what every basis said to carry and what escalation actually did over each horizon next. The array is deliberately not rendered here (it is a re-derivation dataset, not a reading experience) and deliberately not serialized into this page either; it is linked so the underlying rows stay one click away instead of shipping unread in every page load.