The Journey Through Real Estate Price Model Development
summary
Public registries of every Danish property and sale contracts feed a pipeline that targets median absolute percentage error (baseline ~5–15%). K‑nearest neighborhood features and simple official time indices produce a workable zero model; switching to neural networks +15% improves in‑sample fit but yields unpredictable out‑of‑sample bias. Replacing that with a Bayesian time index preserves historical values and fixes update drift; a unified “U‑model” adds ~3%. A quality score predicts per‑house reliability and a random‑forest explainer gives feature‑level reasons for each price.
AI-written summary · about
chapters (17)
- 0:00Start
- 4:08journey begin
- 4:30data
- 5:47blem of time dimension
- 7:10blem of property features
- 9:58N features
- 12:48good enough?
- 19:45at to do with hyper parameters?
- 20:48machine learning to decide
- 22:58edges
- 26:37veloping ML package
- 27:51geography index anymore
- 28:28find some edge cases
- 30:20proportions of neighbors
- 30:49we trust the prediction?
- 31:35ality Score (QS) model
- 36:24sons learned
related talks
28mInterpretable and realistic generative models in data science?
25mWhy Good Models Fail After Deployment
34mLearning Market Dynamics for Optimal Pricing
59mA developers guide to Machine Learning
26mProduction Data to the Model: “Are You Getting My Drift?”
01h 02mA developers guide to Machine Learning
21mPredicting Bond Prices: Regime Changes
19mOpen Source Property Assessment: Tidymodels to Allocate $16B in Property Taxes