An automated valuation model (AVM) estimates what a property is worth without a human surveyor ever visiting it, by learning the statistical relationship between a property's features and the prices that similar, nearby properties recently sold for. This simulation implements the core of that pipeline honestly: a real multiple linear regression, fitted by ordinary least squares to a set of comparable sales, predicting price from square footage, a location desirability score, bedroom count, bathroom count and a condition rating.
The headline feature is the confidence interval around the estimated value, and how it behaves as data accumulates — add comparables one at a time or stream them with auto-add, and watch the band tighten as the underlying matrix term shrinks.
What is an Automated Valuation Model (AVM) and how do real ones work?
An AVM is a statistical or machine-learning system that estimates a property's market value from its measurable features — square footage, location, bedroom and bathroom count, condition, age — and from recent sales of comparable properties nearby, without a human surveyor visiting the property. This simulation implements the regression half of that pipeline with real ordinary least squares mathematics, not a lookup table.
Why does the confidence interval narrow as more comparable sales are added?
The half-width of the confidence interval is proportional to the residual standard error multiplied by the square root of x₀ᵀ(XᵀX)⁻¹x₀. As more comparable sales are added to X, that term shrinks toward zero, so the band around the price estimate tightens.
What is multiple linear regression and why is it used for property valuation?
It models price as a weighted sum of features, with weights chosen by ordinary least squares to minimise squared prediction error across comparable sales. It is popular because the coefficients are directly interpretable in pounds per feature unit.
R² is the fraction of price variance explained by the model, 1 minus the ratio of residual sum of squares to total sum of squares. It is a direct, tunable measure of how much noise you have injected via the market noise slider.
Each coefficient reflects that feature's price effect after accounting for all others, so features with larger typical price impact (like square footage) naturally get larger coefficients than features with smaller typical impact (like one point of condition).
The confidence interval shown bounds uncertainty in the average price for the subject's exact features and shrinks toward zero with more data. A prediction interval for one actual sale would also include irreducible transaction noise and would not shrink below roughly the residual standard deviation.
AVMs struggle with unusual properties outside the comparable range, stale data in fast-moving markets, and hard-to-quantify features like kerb appeal — which is why regulators generally require a human valuation for high-stakes lending decisions.