Miami Housing Price Prediction — Executive Summary

The Problem

When a homeowner receives an offer on their property, they face a difficult question: is this a fair price? Professional appraisals are costly and slow. This project builds a machine learning model that predicts Miami residential sale prices from observable property features, giving homeowners a fast, data-driven reference point.

The Data

The dataset covers 13932 residential properties sold in Miami in 2016. Each property is described by 15 features including location (latitude, longitude), distance from the ocean, highway, and city center, land and living area, property age, structural quality rating, and whether the property falls under an aircraft noise zone.

The Approach

Eight model types were trained and evaluated using 5-fold cross-validation with 3 repeats. Four different feature engineering strategies were tested for each model, producing a comprehensive comparison across model families and preprocessing approaches.

tribble(
  ~"Model Family",    ~"Best CV RMSE",  ~"Recipe",
  "XGBoost",          "0.1440",         "EDA Transformations",
  "Random Forest",    "0.1540",         "EDA Transformations",
  "SVM-RBF",          "0.1590",         "EDA Transformations",
  "MARS",             "0.1640",         "EDA Transformations",
  "KNN",              "0.2330",         "EDA Transformations",
  "Linear",           "0.2170",         "Interactions + PCA",
  "Elastic Net",      "0.2170",         "Interactions + PCA",
  "Lasso",            "0.2170",         "Interactions + PCA"
) %>%
  gt() %>%
  tab_header(title = "Cross-validation results by model") %>%
  tab_style(
    style = cell_fill(color = "#e8f5e9"),
    locations = cells_body(rows = 1)
  )

Cross-validation results by model
Model Family	Best CV RMSE	Recipe
XGBoost	0.1440	EDA Transformations
Random Forest	0.1540	EDA Transformations
SVM-RBF	0.1590	EDA Transformations
MARS	0.1640	EDA Transformations
KNN	0.2330	EDA Transformations
Linear	0.2170	Interactions + PCA
Elastic Net	0.2170	Interactions + PCA
Lasso	0.2170	Interactions + PCA

The Winning Model

XGBoost with EDA-driven feature engineering was the best-performing model. The feature engineering strategy was motivated by exploratory analysis, including log-transforming highway distance, applying a square root to special features value, squaring longitude to capture a U-shaped price relationship, and using natural splines for latitude and ocean distance where the relationships were too curved to linearize.

Test Set Results

test_preds <- predict(final_fit, new_data = miami_test) %>%
  bind_cols(miami_test) %>%
  mutate(
    sale_prc_actual    = exp(sale_prc_log),
    sale_prc_predicted = exp(.pred)
  )

tibble(
  Metric = c("RMSE (log scale)", "R-squared", "RMSE (dollars)", "MAE (dollars)"),
  Value  = c(
    round(rmse_vec(test_preds$sale_prc_log, test_preds$.pred), 4),
    round(rsq_vec(test_preds$sale_prc_log, test_preds$.pred), 4),
    scales::dollar(round(rmse_vec(test_preds$sale_prc_actual, test_preds$sale_prc_predicted))),
    scales::dollar(round(mae_vec(test_preds$sale_prc_actual, test_preds$sale_prc_predicted)))
  )
) %>%
  gt() %>%
  tab_header(title = "Final model performance on held-out test set")

Final model performance on held-out test set
Metric	Value
RMSE (log scale)	0.1415
R-squared	0.9363
RMSE (dollars)	$89,864
MAE (dollars)	$42,294

Key Findings

The final model explains 93.6% of the variance in Miami home sale prices. On the held-out test set the model’s predictions are off by about $42,294 at the median, meaning half of all predictions are within $42,294 of the true sale price. The average error of $89,864 reflects a small number of high-value properties that are difficult to predict accurately, a common limitation when luxury properties are underrepresented in training data.

Location dominates the prediction. Ocean distance, latitude, and longitude are consistently the strongest predictors, reflecting Miami’s geography where proximity to the water commands a significant price premium. Structural quality and living area are the next most influential factors, confirming that buyers weigh both location and physical characteristics heavily.

The EDA-driven feature engineering recipe outperformed automated alternatives across every nonlinear model family. This finding demonstrates that domain-informed transformations add measurable value even when powerful models like XGBoost are available, and that investing time in exploratory analysis before modeling pays off in final performance.

Limitations

The model is trained on 2016 sales data only. Miami real estate prices have changed significantly since then and predictions on current listings would require retraining on recent data. The model also does not account for interior condition, recent renovations, or neighborhood-level factors not captured in the available features. Predictions for luxury properties above $1 million should be interpreted with additional caution given the limited representation of this segment in the training data.