September 8, 2026 CE replication experiment

This experiment was run on September 8, 2026, from the cached CE Interview PUMD files. It is a retrospective experiment with the corrected published threshold series. It does not alter the original 2025 forecast commitment or establish a real-time forecast evaluation.

The immutable result is ce_replication_2019_2025.json. It includes the archived projection evaluation and identities of code that has since been removed. The current script replays the source-replication stages into a separate output; it does not regenerate the historical projection evaluation or overwrite this receipt.

To inspect the current command without loading CE data:

python scripts/replicate_current_ce.py --help

A deliberate cached-source replay can use:

python scripts/replicate_current_ce.py --output benchmark_output/ce_source_replication_2019_2025.json

The script requires the cached CE ZIPs in ~/.cache/spm-calculator/ce-pumd (or --cache-dir) and the explicitly selected annual CPI file (--cpi-input). It refuses to download missing CE inputs and does not call the CPI API. The current replay records SHA-256 hashes for all 11 ZIPs, 44 selected quarter members, the CPI file and the source code. Original CE publication and revision dates are unknown; the manifest records that these exact bytes were available at the replay date.

Scientific stages¶

The current functions live in spm_calculator.ce_threshold. normalize_ce_sample validates the reported counts and survey weights, selects the child-CU sample, and records every exclusion reason, count and weight. construct_normalized_expenditures constructs expenditures, derives CPI shares from that same sample, applies explicit annual CPI inputs, and normalizes to the reference family. estimate_thresholds uses a stable positional selection of the inclusive 47–53 midpoint-CDF band and records its observations, weights, component averages and tenure-specific averages. replicate_thresholds combines the stages and returns their diagnostics. The original calculate_base_thresholds entry point returns a threshold dictionary. These are research tools, separate from the canonical load_forecast().calculate_unit(...) consumer.

The percentile convention is explicit but has not been matched to BLS’s unpublished percentile implementation. Degenerate small samples use the existing nearest-median/pooled-tenure fallback and report it. None of the full 2019–2025 baseline runs required either fallback. Duplicate source indexes produce identical thresholds; a check using the real 2024 window is recorded in the artifact.

Minor-only consumer units¶

The CE consumer-unit definition permits financially independent minors. Consequently, a unit with FAM_SIZE == PERSLT18 is not necessarily malformed. For example, the 2020Q4 data contain NEWID 4429352, with a 16-year-old reference person, one person, one person under 18, tenure 2, and weight 39,790.985.

The BLS SPM methodology describes adult-containing equivalence formulas, but an exact treatment of these minor-only CE units could not be established. The default youth_policy therefore raises an error. This experiment explicitly selects exclude_unresolved, an approximation. Original counts remain unchanged. legacy_recode is a separately labeled diagnostic imposing one adult while retaining the reported child count. It is a nonofficial sensitivity assumption, not a proposed reconstruction of membership or dependency.

The seven target windows contained 12, 12, 7, 5, 1, 1 and 1 unresolved interviews, respectively. They represent 0.005–0.033% of the child-CU survey-weight sum. Switching from exclusion to the diagnostic recode changed thresholds by at most 0.0411%. These are overlapping windows of interviews; their counts and weight sums must not be added up as unique people or unique national consumer units.

This sensitivity does not identify the correct BLS policy. In particular, it does not resolve classification of independent teenagers in larger units. A future exact implementation requires documented BLS treatment and any member/dependency data necessary to apply it.

Tenure sensitivity and replication levels¶

The CE dictionary defines tenure 3 as owned with mortgage status unreported, 5 as occupied without cash rent, and 6 as student housing. The baseline maps 3 to ownership with mortgage and excludes 5/6 before CPI-weight construction and percentile selection. These are explicit research policies; their equivalence to BLS’s exact sample selection remains unverified.

Including 5/6 as renters changed a threshold by as much as 1.7372%. Their excluded weight share was 1.42–1.56% of the child-CU sample. No tenure-3 records appeared in the selected windows, so reassigning that code had zero measured effect and provides no evidence about its effect in data where it is present.

Signed percentage differences from the corrected published thresholds:

Annual CPI rather than quarterly treatment, the food-summary allocation, omitted internet expenditures, omitted BLS in-kind benefit imputations, and the percentile convention remain separately unmeasured approximations. The youth result does not establish that these other approximations are negligible. No sampling or imputation uncertainty interval is estimated.

Archived projection evaluation¶

The September 8 experiment used the former projection.ProjectionInput and project_thresholds interfaces. Those APIs have been removed; their names here identify the archived experiment, not executable current instructions. The former helper multiplied an official base by a replication ratio, price-index ratio or declared blend and checked dated input compatibility. The immutable receipt preserves that experiment’s inputs and code identities.

For that 2020–2025 retrospective exercise, mean absolute percentage errors across the three tenures and six target years are:

The inputs are conservatively dated at this replay, rather than assigned invented historical availability dates. Composite-price inputs use the static shares rebased to each previous year, differing from the fixed 2019 price base in the frozen historical experiment. These comparisons describe the September 8 method and must not replace earlier commitments or be described as prospective evidence. This archived experiment made no new 2026 forecast. The separate canonical rolling forecast now advances CE and ACS windows through 2035 under explicit conditional assumptions and carries its own evaluation.

Replication results carry a stable configuration fingerprint covering sample policies, mortgage-principal treatment, annualization, the median share, percentile convention, and CPI input mode. The archived projection inputs used each result’s actual fingerprint to reject mixed methods. The immutable receipt retains the published threshold series and its then-current loader identity, alongside CE and CPI source identities. Current rolling components carry separate code and input hashes.