New York mortgage market benchmarks

How to read this. Raw, unadjusted HMDA denial rates (conditional on a terminal decision — denied ÷ (denied + originated)). NOT controlled for loan purpose, amount, geography, or credit — these show disparities, not proven discrimination.

Which denominator. These rates use hmda.denial_rate.orig_basis — denied ÷ (denied + originated), 25.374% nationally in 2024. Two other conventions exist in this product and are NOT interchangeable with it: hmda.denial_rate.action3 (denied ÷ actions 1–5, 19.526%), which is what the market tiles show, and hmda.denial_rate.decisioned (denied ÷ actions 1–3 over closed-end consumer loans only, 21.286%), which the fair-lending screen uses because it is the one that isolates the credit decision. A gap between two “denial rates” on two pages is usually this, not an error.

Unadjusted, and it matters. A raw gap here pools purchase with cash-out, first lien with subordinate, site-built with manufactured — and product mix alone can manufacture most of a gap or hide one entirely. One 2024 lender's raw gap for Black applicants is +47.9pp and +1.3pp once the comparison is like-for-like; another looks 12.8pp better than market raw and is 3.1pp adverse adjusted. The mix-adjusted screen does that comparison per lender.

Subject: White conditional denial rate (denied of terminal decisions). Benchmark: all applicants in New York, same denominator. Raw and unadjusted — a gap reflects many factors (application mix, credit, geography), not discrimination by itself.

Originated only. Every cell counts loans MADE (HMDA action taken = 1). Applications, approvals the borrower did not accept, withdrawals and denials are all excluded, so this is a mix of lending, not of demand. Purpose classes partition the universe: refinance total = rate/term + cash-out, which is stated rather than stored as a column — the same discipline that ended the `refinance_count` collision between two surfaces.