Rolling expenditure and rent forecasts

The calculator projects missing observations and advances the expenditure and rent windows used to estimate SPM thresholds. Price growth and real spending growth are separate assumptions. An inflation-only projection holds the real spending of new observations constant; it does not imply that actual real spending will remain constant.

Version 1.0 loads one canonical schema 2 artifact for 2022–2035, with published national inputs through 2025 and conditional research forecasts thereafter. Archived source snapshots and forecast commitments retain their original bytes. The artifact records an information date of September 9, 2026, source receipts and a content digest. Its Python consumer works without PolicyEngine, Microcosm, Axiom or a network connection.

Reference years and moving windows¶

The revised BLS methodology uses five years of CE data lagged by one year. A threshold for reference year T uses collection quarters (T−5)Q2 through TQ1. The current observed files end in 2025Q1. The Census SPM geography method uses the preceding five calendar years of ACS rents.

National thresholds and housing shares for 2022–2025 use published BLS values directly. Published Census indices anchor represented areas in 2022–2024; the post-2024 geographic projection uses the 2024 rent-index anchor. Residual areas without a published anchor remain explicitly modeled. A 2025 local amount therefore combines published national inputs with modeled geography. The dated source check records that 2025 local geography was not yet published at the September 9 information cutoff.

Prices and real spending¶

For a missing CE quarter, the donor is the latest observed quarter in the same season. For example, projected 2026Q1 uses observed 2025Q1, while projected 2025Q2 uses observed 2024Q2. Demographics, tenure, weights and missing-value patterns stay with that donor. Monetary expenditures scale by:

projected expenditure = donor expenditure × price factor × real spending factor.

The price factor uses the CE composite index with weights from the origin’s five-year sample. Historical annual CPI observations are pinned. The future path uses the February 2026 CBO calendar-year CPI-U forecast: 2.9248% in 2026, 2.5476% in 2027, 2.3614% in 2028, 2.2880% in 2029 and about 2.26% annually in 2030–2035. The pinned CBO source and horizon inputs record the exact annual levels and growth rates. Growth ratios apply to the observed BLS 2025 anchor. Applying aggregate growth uniformly to future CE price components and baseline rents is a package assumption; CBO does not supply those component or local-rent forecasts here. The committed BLS API receipt verifies all-items CPI-U annual averages of 313.689 for 2024 and 321.943 for 2025. The latter covers eleven months: October 2025 data were not collected.

Two spending scenarios use the same historical records and price path:

• No real spending growth (zero_real) gives new observations a real spending factor of one.

• CE real-spending trend (ce_trend) fits the log of near-median FCSU expenditures in five nonoverlapping annual blocks ending in Q1. All blocks use the same reference-family normalization and origin-year prices. The forecast applies half the fitted annual log slope: a declared 0.5 shrinkage toward zero. For a donor advanced by k years, the real factor is exp(0.5 × fitted log slope × k).

Each annual block selects its own 47–53 percentile expenditure band. The trend may be negative. Ten-block and pandemic-excluded fits show sensitivity to the estimation period; the latter excludes blocks ending in 2021Q1 and 2022Q1. Ordinary least-squares slope standard errors describe the fit, not survey-design uncertainty or a forecast interval. The default ce_trend specification was selected independently of the retrospective test results. It is not labeled the best-performing model.

The current-origin fits in the pinned CE component give the following annual rates after the declared 0.5 shrinkage:

The difference between these fits is sensitivity to the estimation period, not a confidence interval.

Aggregate real consumption growth is not interchangeable with real FCSU spending among families near the SPM estimation range. The common real-growth scenario also holds future sample composition and relative category spending fixed. It does not model changing household composition or behavioral responses. Existing CE replication approximations remain, including annual price treatment and omitted internet and in-kind imputations.

Thresholds and housing shares¶

Every projected year reruns the CE estimator on its own twenty-quarter window. Let C(T,h) be its raw national threshold for year T and tenure h. Future national thresholds use:

threshold(T,h) = published threshold(2025,h) × C(T,h) / C(2025,h).

The estimator’s housing fraction is 0.82 × tenure SU mean / C(T,h). The projected housing share equals the published BLS 2025 shelter-plus- utilities share times the ratio of that estimated fraction in T to its value in 2025. Anchoring recovers the published base exactly. Shares outside (0,1) cause an error.

The sample and composite-price weights are recomputed inside each target window. Donor scaling uses the frozen origin weights to bridge price years; the estimator then applies only the remaining target-year price adjustment. This is an explicit approximation, with no second application of the same inflation interval.

A positive real-growth assumption need not raise every tenure’s threshold at every horizon. In 2026, the national renter threshold is $43,829.55 under `ce_trend`, compared with$43,888.11 under zero_real. The reselected band’s mean FCSUti spending rises by $76.46, while its overall SU mean rises by$127.94 and its renter SU mean falls by $33.25. The formula subtracts the overall SU mean and adds the renter SU mean, so this housing adjustment offsets the increase in common spending. By 2030, renter thresholds are$50,011.91 and $48,954.72 respectively; by 2035, they are$58,520.42 and $54,735.04. These are conditional research results from the canonical artifact, not published BLS forecasts.

Geographic rent indices¶

ACS inputs are cash-rented, occupied two-bedroom housing units with complete kitchen and plumbing, positive gross rent and positive survey weight. GRNTP × ADJHSG / 1,000,000 applies the source product’s dollar adjustment once. The calculation pools weighted individual records; it does not average annual medians. Census explains why multiyear estimates describe a pooled period.

Public PUMS identifies PUMAs rather than exact Census SPM areas. Pinned fractional PUMA-to-county-to-area mappings use population allocation and Census’s 2013 metropolitan definitions. Historical 2022 estimation uses 2010 PUMAs; 2023 onward uses 2020 PUMAs. This approximates area membership and does not identify actual household counties.

The menu has 349 areas in each year: 342 published areas and 7 unanchored modeled residuals in 2022, then 341 published-menu areas and 8 unanchored modeled residuals. The later years’ published-menu identifiers carry modeled rent indices. Source population outside the published menu remains in the full-US denominator. Every calculated area must have positive coverage. County FIPS is a lookup input assigning a unit to its year-specific area, not a separate county estimation level.

The 2019–2023 and 2020–2024 PUMS products have different weights and dollar vintages. Let R be the modeled ratio of local to national pooled median rent. The forecast bridges the common 2020–2023 period:

index(T) = official index(2024) × R(old overlap)/R(old full) × R(new T)/R(new overlap).

Each ratio uses records from one product. The overlap comparison removes the measured common-window vintage shift; it does not eliminate every revision effect. The four-year overlap uses the five-year product’s supplied weights and is not an official four-year estimate. Weighted microdata medians also approximate Census’s grouped-data interpolation.

Missing ACS years copy the observed 2024 cohort. Baseline nominal gross rents grow at the all-items CPI path and are deflated by that same path. New cohorts therefore have zero real rent growth. Historical cohorts still roll out, so local and national pooled medians can change differently. A separate sensitivity adds two percentage points to each missing year’s nominal rent growth while leaving the deflator unchanged. The CE spending scenario does not change this ACS assumption.

For housing share s and rent index I, the location factor is 1 + s × (I−1). Both s and I can change over time. Under this artifact’s fixed 2024 donor distribution, uniform nominal growth and matching deflator, all 349 relative rent indices reach their final constant level in SPM year 2029, through 2035. The first unchanged transition is 2029→2030. The 2029 window contains the observed 2024 cohort plus four projected cohorts; 2030 is the first all-projected window. The estimator continues advancing windows; it does not manually freeze rent values.

Housing shares still move slightly from 2029 to 2030, so geographic factors change by up to about 0.0000001024 under ce_trend. After 2030, remaining factor changes are at floating-point precision. National dollar thresholds continue changing. These assumptions do not generate persistent local rent-growth differences. The artifact records this limit in assumptions.acs.relative_index_stabilization.

Historical series breaks¶

The artifact preserves two 2022→2023 breaks rather than interpreting them as annual rent growth:

The selected area’s series_breaks metadata labels both endpoints. Sumter’s notice applies to that county assignment; it is not a statement that all counties in the modeled residual changed assignment. Values and source identities remain available for audit.

Support and topcoding¶

Repeated projected copies do not create new original observations. Diagnostics combine their allocation-adjusted weights by original record before computing the Kish count, (sum weights)² / sum(weights²). An expected whole-record count sums each original’s geographic allocation fraction once. An area is flagged when the minimum of unique records, expected count and Kish count is below 30. Projected windows also flag thin support in their original donor cohort. This is a declared research heuristic, not a Census publication standard or a survey-design effective sample size.

The artifact reports cohort-specific cash-rent and utility topcoding flags, affected weight shares and a conservative median warning. A median below the cash-rent topcode does not prove that utility topcoding is irrelevant. Thin areas retain their values and warnings; the consumer never substitutes a different geography.

Retrospective evaluation¶

CE forecasts use origins 2019–2024 and targets through 2025, giving 6, 5, 4, 3, 2 and 1 folds at horizons one through six. Each fold uses its own published revised-method origin threshold and only CE records through that origin’s Q1. Both candidates and the inflation-only comparator use realized annual prices through the target. These are current-vintage, price-conditional retrospective tests, not a reconstruction of real-time forecast accuracy.

The primary error is mean absolute percentage error (MAPE), equally weighting the 63 fold-tenure results. Short horizons have more folds. A second metric equally weights the per-horizon MAPEs for displayed horizons one through five. The app warns when a scenario fails to beat the CPI-only baseline overall, on that horizon-balanced measure, or at the selected forecast horizon.

The ACS holdout uses the corrected 2018–2022 product, anchors to the official 2023 index, drops 2018 and adds a projected copy of the 2022 cohort as 2023. It compares predicted 2024 indices with official 2024 values and unchanged 2023 indices across all 341 areas. Its primary metric equally weights area absolute percentage errors. This tests the donor mechanism, not the later cross-vintage bridge. The app flags failure to improve on unchanged indices.

Results from the pinned CE and ACS components:

Both rolling methods outperform CPI extrapolation at every evaluated horizon. Zero real growth performs slightly better at one year; the trend performs better on the overall and horizon-balanced comparisons. Only two folds inform the five-year result and one informs the supplemental six-year result. The ACS donor method’s MAPE is 1.2537%, compared with 1.7284% for unchanged indices. Its median absolute percentage error is 0.8745%, compared with 1.5048%. These comparisons have the source-vintage and price-conditioning limits described above.

Missing folds, areas or invalid results block generation. Numerical underperformance is reported rather than treated as a data failure. The artifact also freezes a hashed 2025 geographic prediction for comparison with the later official workbook; that forward test is pending official data. Neither these tests nor the fitted trend provide forecast uncertainty bounds.

Standalone calculation¶

Run this example after installing the package:

from spm_calculator import SPMUnit, load_forecast

forecast = load_forecast()
assignment = forecast.resolve_county(2026, "06075", county_vintage="2020")
area = forecast.areas_for_year(2026)[assignment["area_id"]]
result = forecast.calculate_unit(
    SPMUnit(
        unit_id="household-1",
        num_adults=2,
        num_children=2,
        tenure="renter",
        year=2026,
        geography_kind=assignment["kind"],
        geography_id=assignment["area_id"],
        resources=50000,
    ),
    scenario="ce_trend",
)
print(area["name"], area["area_type"], area["status"])
print(round(result["threshold"], 2), result["is_in_poverty"])

Use geography_kind="national" without an area ID for an explicitly national calculation. entry(year, scenario=...) exposes each year’s thresholds, shares, rent indices and windows. Results record the selected scenario, artifact identity and per-area diagnostics. An area’s official_published_area describes its source-menu status; status describes the current geographic component.

Unknown years, areas, counties, scenarios or county vintages raise errors. An as_of date before the information cutoff also raises. No alternate runtime release path or geographic substitution supplies missing inputs. Pass expected_sha256 when loading a retained artifact to verify its content identity; retrieve the current identity from forecast.content_sha256. The provider, Frame and real Axiom examples use this same artifact. Their adapters ship inside this package; the PolicyEngine country and wrapper integration ships in policyengine-us 2.0 and the policyengine wrapper 6.0, both in progress. See the 1.0 migration guide.

Rebuild and verify¶

Raw CE ZIPs and ACS PUMS products stay in configurable external caches. The component manifests pin their URLs and hashes; compact crosswalks and source receipts ship with the package. With matching raw source bytes cached:

These commands regenerate outputs; use them only when intentionally rebuilding the scientific artifact, not to run a threshold calculation.

python scripts/build_ce_forecast.py --output spm_calculator/data/current/ce_rolling_forecast.json
python scripts/build_acs_forecast.py --output spm_calculator/data/current/acs_rolling_forecast.json
python scripts/build_rolling_forecast.py

Both scientific builders support --cache-dir and an offline --check that repeats the calculations from cached raw inputs. The assembly check below uses committed component results, verifies their scientific code and package input hashes, checks the official CPI receipt and required evaluation coverage, and reproduces the portable artifact without raw microdata:

python scripts/build_rolling_forecast.py --check

The committed browser export identifies the verified PyPI package 1.0.0.post1. Reproduce and verify the export with:

python scripts/export_web_release.py --published-package-version 1.0.0.post1
python scripts/export_web_release.py --published-package-version 1.0.0.post1 --check

For an unpublished development preview, omit --published-package-version when generating and checking that preview’s export. The default remains local_preview; an explicit published version must match pyproject.toml and be verified before use.

Changing scientific logic or pinned scientific inputs requires rebuilding the affected component before assembly. A change confined to code-identity serialization can use a separately recorded equivalent-code adaptation when the original source and receipts are retained, normalization logic and cached products are unchanged, and exact comparison confirms that every scientific field survives reassembly. Such an adaptation records both source identities; it does not claim a fresh raw-data parse. See the current artifact identity for the retained evidence and current consumer pins. Hashes check integrity; they are not signatures or independent evidence of source authenticity.