Replication package: When a Deed Is Not a Market Sale: Foreclosure Transfers, Statutory Consideration, and Repeat-Sales House-Price Measurement

Abstract

Replication package and working paper for When a Deed Is Not a Market Sale: Foreclosure Transfers, Statutory Consideration, and Repeat-Sales House-Price Measurement, by Pablo Loschi (independent researcher, Berlin). When a lender takes title through a referee's foreclosure deed, a deed in lieu or a transfer between lenders, New York State and City transfer-tax rules can base the recorded consideration on a foreclosure judgment, a bid or outstanding debt. The amount then follows a legal rule, and the state and city calculations differ. The package identifies 16,655 such deed-family documents among 1.47 million recorded in Manhattan, the Bronx, Brooklyn and Queens from 2003 to 2025. Standard screens keep many of them, and 94–95% of screened legal-rule deeds appear in the Department of Finance sales file at the recorded amount. In a Bailey–Muth–Nourse repeat-sales index of 1–3 family houses, dropping pairs with a lender take-title endpoint changes the quarterly index by up to 3.7 log points (95% CI 2.5–4.9). The point estimates of the cumulative peak-to-trough decline differ by only 0.6 log points, with an interval that includes zero: the correction changes the quarterly path of the index, and the evidence does not establish a change in the total decline. The largest quarterly difference is 10.9 log points for Bronx houses; for condominiums it is negligible. Version 1.0.6 adds Appendix C (stage 09). It repeats the drop correction on the 138,380 pairs whose two endpoints are both 1–3 family deeds. The 2010Q3 difference is +3.61 log points for the published definition and +2.62 for the narrowest, against +3.69 and +2.63 in the main tables. A 95% band that holds jointly across six definitions and all quarters runs from +1.63 to +5.59 and from +0.64 to +4.60 at that quarter. For the published definition it excludes zero in five of the eight 2011–12 quarters; for the narrowest, in none, although that definition's 2011–12 average (−1.08, interval −1.30 to −0.88) is negative. A deed placebo matched on borough and quarter never reaches the observed values; a second one that also matches each deed's number of pairs exceeds the narrowest definition's value in 1 of 999 draws. The placebo margin therefore depends on the design, and neither share is a p-value. The version also tightens four phrases in the abstract and conclusion. No earlier estimate changes. Version 1.0.5 added three checks on the 1–3 family result (Section 4.6, stage 08). A stricter placebo removes random ordinary pairs, exactly as many as the correction drops and with the same timing (borough, first-sale quarter and second-sale quarter). It is a much harder test: its largest draw is 3.41 log points, against 1.94 in the deed placebo. The published 2010Q3 difference of 3.69 still exceeds all 200 draws, narrowly, and the narrowest definition clears it by more (2.63 against 1.89). A simultaneous 95% band for the whole path runs from 1.6 to 5.8 at 2010Q3 and excludes zero in 28 quarters. Over 2008–10 the raw index lies 1.07 log points above the corrected one (95% CI 0.55–1.59), and over 2011–12 it lies 2.28 below (1.72–2.76). Assigning every group the document review leaves unresolved against the classifier keeps the 2011–12 shortfall in every scenario; the 2009–10 excess becomes uncertain only when all four assignments are combined. No earlier estimate changes. Version 1.0.4 added a test with narrower definitions of the dropped set (Section 4.5, stage 07). Dropping only referee's deeds to lenders with non-zero cents and no securitization-type grantee, 3,309 deeds or about a third of the flagged set, still moves the 2010Q3 index by 2.63 log points (pointwise 95% CI 1.83–3.67). The largest of 200 matched placebo draws is 1.19. Under every definition the result exceeds all of its own placebo draws. These sets come from the same classifier and are not an independent validation. Version 1.0.4 also describes the document review more precisely. No earlier estimate changes. Version 1.0.3 added a document review of the stratified 400-deed sample. The original protocol called for two human readers; the review uses AI instead. ChatGPT read every deed image, and Claude then checked each reading against ACRIS open data and the protocol's rules. Claude did not read the images, so the rates measure agreement with those readings, not accuracy against a verified record. Among resolved deeds the classifier's precision is 0.993 (95% interval 0.978–1.000). Its recall is 0.966 (0.934–0.998) for foreclosure-type transfers, the class it targets, and 0.496 (0.168–0.824) against every non-price transfer, because it does not look for sales between relatives. 84 deeds remain unresolved, 48 of them third-party auction purchases whose images do not show whether the bidding was competitive. Files in this record: ACRIS_CONSIDERATION_v1.0.6.zip: code, classifier, manuscript (source, PDF and Word), tables, figures, documentation, the 400-deed review sample, the AI readings and revision log (audit/cross_ai_review/) and the analyst key audit/audit_key.csv, whose SHA-256 matches the value published with version 1.0.1. ACRIS_CONSIDERATION_v1.0_DATA.zip: the frozen NYC Open Data pulls and the derived data. ACRIS_CONSIDERATION_v1.0.6_MANUSCRIPT.pdf: the manuscript on its own. revision_notes.md: changes between versions. The README explains how to reproduce the results from the frozen snapshot; code/score_cross_ai_review.py --check rescores the review and code/07_narrower_definitions.py --check confirms Table 5, each against the manuscript. Source data are public NYC Open Data records; the included LICENSE gives the source-data terms. For details of changes between versions, see revision_notes.md.