San Jose: Alum Rock: market area analysis

Single-family detached houses. What drives value in San Jose: Alum Rock, as fitted by three regression programs on the town's sales, with the adjustments each model implies. The switch at the top shows every table and graph from the log model (effects as percentages of value) or the $ model (effects in dollars).

At a glance

The models

The best model from each of the three programs: the GAM (mgcv-rs), MARS (earth-rs) and the elastic net (glmnet-rs). The one that values held-out sales best (highest OOF R²) is marked as the best model; the glance tiles, the solid line on each graph and the county table use the best model.

What drives value

Each graph changes one feature of the typical sale and shows the value change. Move the pointer over a line for the value at that point and the adjustment per step.

Show the other models on the graphs

Neighborhoods and remarks features

Surfaces (3D)

Two features at once: the location surface across the town, pairs of features, and how the CQA score scales the value.

Equations

The fitted models as written equations. MARS and the elastic net are sums of straight-line pieces (hinges); the GAM is a sum of smooth curves, drawn in the graphs above.

Comparison of the models

Definitions

Adjustment: The dollar amount by which a difference in one feature changes value, read here at the typical sale.
Best model: Of the three programs' models, the one with the highest OOF R²: the model that values sales held out of its fit most accurately. The log model and the $ model each have their own best model.
CQA score: Condition, quality and appeal: the percentile of a house's market appeal divided by 10, from 0.00 to 10.00, read from the residual and refined by the appraiser's judgment.
Contour map: A flat view of a surface, with lines joining points of equal value, like the height lines of a topographic map.
CV R²: The average, over the ten groups of sales, of the share of price variation explained for each group when the model is fitted without that group (cross-validated R²).
Effective date: The date the values refer to; every graph shows the market at this date.
Elastic net: A regression that keeps or drops terms by a penalty on the size of their coefficients (shrinkage), here fitted on the MARS hinge terms (glmnet-rs).
Fitting sales (training sample): The sales the models are fitted on.
Foreclosure sale: A sale forced by the lender, at the foreclosure auction (the trustee's sale) or afterward by the bank that took the house back (REO); the seller has no personal interest in the house.
GAM: A regression that gives each feature its own smooth curve and adds the curves together (generalized additive model, mgcv-rs).
Hinge: A straight line flat on one side of a value (the knot) and sloping on the other: max(0, x − knot) or max(0, knot − x).
Log model (log-transformed model), $ model (untransformed model): The log model is fitted to the logarithm of price, so each feature changes value by a percentage; the $ model is fitted to the price in dollars, so each feature changes value by a fixed amount.
MARS: A regression built from straight-line pieces joined at knots (multivariate adaptive regression splines, earth-rs). Degree 2 may also multiply the pieces of two features.
Median absolute percentage error* (MdAPE): The middle value of the held-out sales' percentage differences between price and the model's value, ignoring sign. Lower is better.
OOF R²: The share of price variation explained when every sale is valued by a model fitted without that sale's group (pooled out-of-fold R²).
R²: The share of the variation in sale prices that the model explains, on the sales it was fitted on (coefficient of determination). Higher is better. A model fits its own sales better than it values new ones, so R² runs above the held-out measures (CV R², OOF R²).
Residual: The sale price less the model's value; mostly the value of what the model does not measure: condition, quality and appeal.
Root mean square error* (RMSE): The square root of the average squared difference between price and the model's value on held-out sales, in dollars.
Sale type: Who sold the house, from the listing remarks: a standard sale, a trust/probate sale, a short sale or a foreclosure sale. In a town where a sale type has fewer than 3 sales, those sales count as standard sales.
Short sale: A sale by the owner for less than the mortgage owed, with the lender's consent; the seller is an interested party.
Smearing factor: The correction that turns a value from the log model back into dollars without a downward bias.
Trust/probate sale: A sale by a trustee, or by an executor under court supervision; the seller has no personal interest in the house.
Typical sale: The middle (median) of each feature over the fitting sales, each factor at its most common level, valued at the effective date. A typical sale, not a typical house of the town: houses that sell are often more updated than the houses that do not, so the town's typical house is likely less updated than the typical sale.
95% range (band): The range within which the curve would fall in 95 of 100 samples of sales like these (the GAM's standard error* × 1.96 on each side).

© 2026 William Bert Craytor · Valuation Engineer Research Labs. These pages show market area models: how features affect the value of single-family houses in each town, for adjustments and market analysis. They are not an appraisal, give no opinion of value for any property and are provided without warranty of any kind. Elevation from USGS 3DEP data (public domain).