Investment candidate sourcing
Use a calibrated linear probability model to rank distressed-sale candidates, then compare observed asking price with an OLS fair-value estimate.
Data signals differ by target
Select a feature to compare its direction and strength for price versus distressed-sale status.
Interpretation boundary: correlation is descriptive and feature importance is predictive; neither establishes a causal effect on price or distress.
Statistical model selection
Nested out-of-fold testing compares probability quality and candidate capture. Trees remain useful challengers, but the selected model is simpler and more productive.
Correlated predictors are regularized into one stable linear score, then calibrated into probabilities.
The strongest standalone tree remains competitive on ranking, but captures fewer cases in the review budget.
Price is mostly additive and linear here; OLS improves 26.8% on the best tree MSE.
Resampling, calibration and early stopping occur inside training folds only.
The selected screen captures 51.1% of known distressed cases with 40.0% precision.
Housing price benchmark
OLS has the lowest repeated-CV error. The best tree, XGBoost, remains useful as a nonlinear challenger but does not justify replacing the linear valuation anchor.
The additive price structure rewards a transparent linear fair-value estimate.
Shallow boosted trees are the best nonlinear benchmark, but error is 36.7% higher than OLS.
The dataset contains sale price, not a live asking price; obtain the ask before underwriting.
Distressed-sale probabilities
At a fixed 20% review budget, the selected model surfaces 24 of 47 distressed homes. That is four more than random forest and six more than calibrated XGBoost.
Review the top 60 ranked homes to find 24 of the 47 observed distressed cases.
Probability quality is statistically tied with LDA, but the queue finds three fewer cases.
Retained as the strongest pure-tree comparator for nonlinear structure.
Out-of-fold analytics
The most decision-useful plots from the source analysis are updated to include the calibrated LDA model and the fixed 20% investment-review budget.
Investment screening decision
Use distress probability to decide what to inspect, then use the price model and current market evidence to decide whether the opportunity deserves underwriting.
Sort homes by calibrated distress probability.
Inspect the top 20% queue for genuine urgency and condition.
Compare asking price with the OLS estimate and local comps.
Underwrite repairs, title, liquidity and transaction costs.
Decision boundary: distress does not imply a bargain. A live asking price, repair estimate, title review and resale/rental comps are required before investment.
Cross-model diagnostics