REIT Housing Price Analysis
Estimating home prices and ranking the drivers behind them, using real King County sales records.
- Role
- Data Analyst
- Focus
- Regression & EDA
- Language
- Python
- Dataset
- King County, WA
Overview
A real-estate investment trust needs to know two things before it commits capital: what a property should sell for, and which characteristics actually move that price. This project answers both against real transaction data rather than a synthetic teaching set.
The work runs end to end — sourcing the data, cleaning it, exploring it, modelling it, and translating the model output back into language an investment committee can act on.
Data preparation
The raw extract carries the usual defects of transactional property data: missing values in a handful of numeric fields, extreme sale prices that distort any fitted line, and columns typed in ways that block modelling.
- Handled missing values across numeric fields rather than dropping rows wholesale, preserving sample size
- Identified and removed outliers whose sale prices sat far outside the credible range for their features
- Transformed variables into model-ready form — recasting types and deriving fields where the raw column was not directly usable
Exploratory analysis
Before fitting anything, the goal was to understand the shape of the market: how price distributes, which features correlate with it, and where relationships are non-linear enough to matter.
- Examined square footage of living space as the dominant continuous predictor
- Compared price behaviour across bedrooms, bathrooms, and floor count
- Assessed how property condition and the King County grade rating separate the market into tiers
- Used visualisations — distributions, correlation views, and scatter relationships — to surface patterns rather than assert them
Modelling
With a clean frame, regression models were built and evaluated in Python to estimate sale price from property characteristics, and to rank which of those characteristics carry the most weight.
- Built regression models in scikit-learn to predict sale price
- Evaluated model fit rather than accepting the first result
- Ranked feature importance to establish which attributes genuinely drive price
Communicating the result
A model that only the analyst understands is not a deliverable. The final step converted statistical output into business-focused interpretation — what the numbers imply for which properties are worth acquiring, and which attributes justify a premium.
What the analysis showed
- Interior living space is the strongest single predictor of sale price — more so than raw room counts.
- Grade and condition ratings separate the market into tiers that a bedroom count alone does not capture.
- Cleaning decisions — particularly outlier handling — materially change what the fitted model reports.
- Feature importance is the deliverable an investment committee acts on, not the prediction itself.