Predicting Time to Sell Real Estate: Why You Can't Just Fit-Predict

Vasily Baranov24 May 2025

Problem

Avito Real Estate needs to know how fast a listing will sell at a given price. Lead scoring depends on it: only liquid objects should go down the funnel. So does a seller hint that trades a high price and a long wait against a low price and a fast sale. The valuation (IMV, Instant Market Value) gives the expected price but not the speed, and the old liquid/illiquid classifier reached 0.69 ROC AUC. A plain regression on time to sale scored better but could not ship: its prediction does not fall steadily as the price falls and breaks at extreme prices. Close dates are also a noisy proxy for sale dates.

Example from the article: for one object valued at about 8.5 million, the baseline predicted around 45 days at any price up to 7.5 million, then rose steeply up to 10–12 million, then dropped slightly.

Solution in brief

A model predicts a price-independent base exposure. It is multiplied by (extprice/extIMV)k, which is monotonic in price and equals the base exposure at the valuation. One exponent k is tuned once on a holdout.

Questions this case answers
  • How to predict time to event for an item at a given price so that the prediction is monotonic in price and stays sensible outside the range where price is observed
  • How to train a regression on noisy event dates when the exact event date is unknown and only a proxy date is recorded
  • How to predict time to sale of a listing at a given price for lead scoring
  • How to make a boosting model's prediction monotonic in price without monotonic constraints

Tools and methods

  • CatBoost Regressor — base exposure from features without price or valuation
  • Price correction — base exposure times (extprice/extIMV)k
  • Cleaning rules — five filters that make noisy close dates usable as a target
  • Aqueduct — AvitoTech open-source library for serving the model in steps
Continue reading — it's free

Google or a one-time email code. No password.