True Rent Index — Methodology

How the exit-validated market-clearing rent is computed, and what to distrust about it. v1.0.

Every rent index you have seen is built on ASKING prices: what landlords

hope to get. The True Rent Index measures what units actually DISAPPEAR

at: the exit-validated market-clearing rent. On 2026-07-10 the downtown

Toronto corpus showed an asking median of $2,225 against an exit-validated

median of $1,790. The market cleared 19.6% under ask.

1. Corpus. An autonomous pipeline scans 7 listing sources twice daily

(08:00 / 18:00 ET) within 3 km of Bay & Queen. Every listing-day is

stored; nothing is overwritten.

1. Lemon screening. Before any listing can inform the index, a

classifier removes non-comparable products: shared rooms and roommate

listings (lexicon over title AND full page text), micro-units posing

as apartments (under 250 sqft claiming a bedroom), and implausible

rents (under 45% of cohort median with no stated incentive). Rejections

are kept as data with their reasons, not discarded.

1. Cohorts. Discounts are judged against the median of the listing's

own bed-count cohort (studio / 1 / 2 / 3+), computed from this corpus

over the trailing 14 scan-days. A cohort under 8 observations is not

trusted; unknown bed counts are recovered from listing page text where

possible and flagged where not.

1. Exit validation. A listing "exits" when it stops appearing in

scans. The exit-validated median is the median rent of qualified,

lemon-screened listings that exited within the trailing 30 days. Fast

exits are the market saying yes.

1. Capture and provenance. Every scan writes per-cohort rows

(asking median, exit median, days-to-exit, lemon rate) stamped with

the git revision of the engine that computed them. Improved engines

never silently rewrite history: old rows keep their provenance.

· Corpus size. As of v1.0 the corpus spans 17 scan-days and ~2,400

listing-rows; per-cohort exit counts are small (single digits). Treat

early numbers as directional. The corpus grows every 12 hours.

· Exit is not signing. A delisting usually means leased, but can mean

withdrawn. Reposts (same unit under a new id) are detected by a

conservative multi-signal matcher (photo fingerprint + bed cohort +

a 7-day gap cap + rent band or title overlap) and excluded from exit

validation; on first corpus-wide run it screened 79 repost pairs,

including id-recycling clusters on one source. v1.1 of this section.

· Survivorship of scan cadence. A unit listed and leased entirely

between two scans is invisible. Twice-daily scanning bounds this window

at 12 hours.

· Asking medians move with mix. A flood of studios shifts the "all"

cohort; per-cohort rows are the honest comparison.

· One geography. Downtown Toronto, 3 km radius. Other markets require

their own corpus.

Survival analysis is the accepted method for listing liquidity (Liu 2024,

JTLU; Opendoor research). The index publishes a Kaplan-Meier median

time-on-market per cohort over the screened candidate set, treating

still-live listings as right-censored observations. Two corrections no

prior implementation makes: repost chains never count as exits (Tucker

2013 documents the relisting reset problem; MLS answers with CDOM rules,

rental platforms answer with nothing), and exits are progressively

page-confirmed rather than inferred from feed disappearance (first audit:

7 of 8 inferred exits still had live pages).

The headline clearing gap now carries a split-conformal interval built

from one-step residuals of the index's own history (distribution-free

under exchangeability; approximate under trend, and labeled as such).

Early corpus verdict: 19.6% with a wide band. The band narrows only as

scan-days accumulate; publishing it wide is the honest price of being

new. Exit events are additionally split by cause label: page-confirmed

gone versus unverified feed disappearance, with counts published, per

the competing-risks framing the exit trichotomy demands.

A backlog audit page-checked 148 inferred exits and 93 came back

still-live: most feed disappearance is platform noise, not leasing. The

index now censors these false exits everywhere they could bias a

statistic — the Kaplan-Meier event indicator, the competing-risks cause

labels, and the exit-rent clearing anchor itself (a rent that never

cleared cannot anchor a clearing median). The correction moved the KM

median from 2 days to not-reached-inside-the-window, which is the honest

reading: the earlier figure measured platform churn, not absorption.

With the confirmed-label gate open (30+ page-confirmed exits) the index

publishes the Aalen-Johansen 25% quantile for confirmed absorption; the

AJ median stays null while the terminal cumulative incidence sits below

0.5, and that null is reported rather than substituted.

The methodology version bumps when the computation changes (screening

rules, cohort definition, exit window). Data rows carry the engine

revision, so any published number can be traced to the exact code that

produced it. Version history lives in the public repository's commit log.