Lab

Real: the macro forecast of U.S. private nonresidential construction spending, its backtest, the O'Hare weather history, and the price and rate series, all public data pulled without API keys on 2026-10-04. Fictional: Prairie Ridge Builders (fictional). Its backlog, bookings, margins and cost shares are synthetic, written for this demo, and every one is labeled in the engine's config as sourced or as an assumption.

The target is PNRESCON, in millions of dollars per month, not seasonally adjusted. Six models were tested with a rolling-origin backtest: 71 monthly forecast origins from Sep 2020 to Jul 2026, each forecasting 1 to 12 months ahead. The rule for picking the published model was fixed before any backtest ran: lowest average MAPE over 1 to 12 months. It was run twice: lag 0 treats a month as known when it ends, lag 1 treats it as known a month later, which is about when Census publishes it. The published forecast uses lag 1.

Holt-Winters damped wins under both. With lag 1 its average error is 34.5% lower than seasonal naive (6.65% against 10.15%). The lagged drivers (employment, job openings, rates, metals, lumber and diesel) scored 7.83%: they did not beat the simpler model, which is why the contractor layer is presented as scenarios and not as a forecast.

The P10 to P90 band is read off the published model's backtest errors. It was also checked out of sample: rebuilt from earlier folds only, then tested on origins from Sep 2023 on. A calibrated band would hold about 80%. It holds far less beyond the first few months, because spending surged in 2022 and 2023 and then stalled, and errors from the boom said little about the slowdown. Read it as the range of past misses, not an 80% probability.

All public, no API keys: FRED's CSV download and NOAA NCEI's data service. Every series ID was found by searching the FRED website, and the engine re-runs those searches to confirm them.

Weather: NOAA station USW00094846, CHICAGO OHARE INTERNATIONAL AIRPORT, IL US, 25,109 daily records from Jan 1958 to Sep 2026, pulled 2026-10-04. Architecture Billings Index: not included, because AIA does not publish it on FRED and no file was supplied at construction/manual/abi.csv.

A flagged weekday is a partly productive day, not a lost day. Each rule has a light and a severe tier, and a day takes the lowest factor among the rules it trips. Climatology covers 1996 to 2026, because 2-minute wind records start in 1996. Work not done slips to later months, so a harsh winter moves revenue later rather than erasing it.

Tariffs, named by mechanism only. Each episode's size is the change in metal prices over its window, pushed through pass-through coefficients estimated from 2015-01 to 2026-08 monthly data (R squared 0.6493). Raised is 1 times the episode, reduced gives back half (an assumption), held is zero.

• 2025 to 2026 episode (default). Jan 2025 to Aug 2026. Log change WPU1017 +37.7%, WPU102501 +40.7%, WPU102502 +23.0%. Through the pass-through coefficients that is 4.56% on the construction input PPI. From the last month before the first 2025 metals action to the latest month in the snapshot. It reflects current tariff levels, so it is the default. Over a 19-month window the price change also includes global price and demand moves, so treat it as the larger estimate. Federal Register actions in the window: Proclamations 10895 and 10896, aluminum and steel, Federal Register 2025-02-18, Proclamation 10947, aluminum and steel, Federal Register 2025-06-09, Proclamation 10962, copper, Federal Register 2025-08-05, Proclamations 11021, 11032 and 11045, aluminum, steel and copper, Federal Register 2026-04-09 to 2026-07-23.
• 2018 episode. Feb 2018 to Aug 2018. Log change WPU1017 +15.0%, WPU102501 +6.3%. Through the pass-through coefficients that is 0.91% on the construction input PPI. Six months around the 2018 steel and aluminum tariffs. Copper is left out because it was not covered and fell over the window.

Rates. History cannot separate how rates affect construction from how the Fed responds to booms: every regression the engine ran says higher rates go with more construction growth later (+0.021 to +0.049 per 100 bp). So the scenario uses a labeled assumption: bookings fall 5% per +100 bp, after 6 months. The size comes from Bloesch and Weber, "Structural Changes in Investment and the Waning Power of Monetary Policy", Federal Reserve Bank of New York Staff Reports no. 1190, March 2026. The 5% is the peak response of nonresidential structures SPENDING to a 100 bp CUT. Applying it to bookings, and to a rise, assumes the effect is symmetric. That is an assumption. In that paper the spending response peaks about 10 quarters after the shock. Bookings lead spending, so the 6-month lag on bookings is an assumption. The estimate uses 1969 to 1996 data, and Chicago Fed Letter 482 (2023) reports a positive response of structures to higher rates, which it calls puzzling. The sign is not settled. The base path follows the Fed's September 2026 projections, from 3.75% in Sep 2026 to 4.1% at the end of 2026, 4.1% at the end of 2027, 3.9% at the end of 2028, 3.6% at the end of 2029.

Public funding. A delayed surface transportation bill holds back 60% of public bookings for 6 months. An assumption.

Fuel. Diesel moves to the 10th or 90th percentile of its historical 12-month change (-23.7% or +30.4% in log terms) over the first year. Escalation clauses here cover materials, not fuel.

The company. Materials are 40% of cost, escalation clauses pass 80% of a materials cost change to the owner, 50% of work is weather-exposed, and new bookings start 2 months after signing and run 12 months. Base bookings are trailing bookings scaled by the macro forecast's growth. The synthetic backlog: