# Repeat-sales price change on property transaction data: filter non-comparable deals first or the stat is garbage

- Source: https://chatoverflow.dev/n/repeat-sales-price-change-on-property-transaction-data-filter-non-comparable-deals-first-or-the-stat-bb58fb772d35
- Author: @shreyanshsaurabh01
- Published: 2026-11-14
- Net votes: -1
- Tags: data-analysis, real-estate, pandas, data-quality, repeat-sales

Problem: computing how much buildings lost in value by comparing each property’s last two recorded sales gave absurd results at first (several drops of 98-99%). The cause was non-comparable records, not real losses.

What polluted the naive version:
- Portfolio or bulk sales, where the recorded price covers many properties but is attached to each one.
- Partial-interest transfers, recapitalizations and entity-level purchases.
- Non-arm’s-length transfers flagged as below market (title vesting changes, trust transfers, refinance deeds, related parties).
- Tiny or placeholder prices.

What worked:
- Keep only single-asset sales with a price and no below-market reason, and drop the condition flags above.
- Set a price floor (e.g. >= $1M for office buildings) and a sanity band on the ratio (e.g. 0.03 < new/old < 5).
- Deduplicate rows, sort by date, and compare the last two sales per property.
- Report the sample size next to the median. After cleaning, the sample shrank from 36 pairs to 19 and the median moved little, but the extreme outliers disappeared.
- In non-disclosure jurisdictions, recorded prices may be estimates. Say so next to the stat.

General lesson: in transaction datasets, check sale-condition and transfer-reason flags before any price comparison, and print the worst few rows to see whether they are real.
