Predicting Housing Sale Prices Using Property Features and Location Data

---

embed-resources: true echo: false warning: false message: false

Introduction

The real estate market relies heavily on accurate property valuation, as buying or selling a home is often one of the most significant financial decisions individuals make, often involving long-term financial commitments such as mortgages. Traditionally, determining a home’s value required hiring a professional appraiser, a process that is time-consuming, expensive, and subjective, but with the rise of online real estate platforms, automated valuation models have become an important tool for providing instant and data-driven price estimates. This report investigates whether machine learning models can effectively predict housing sale prices using property features and location data for homes in Ames, Iowa, as part of a project for CornHawkHomes, a startup aiming to compete with platforms like Zillow by offering its own pricing tool, KernelEstimate. The dataset consists of housing observations with features such as size, quality, age, and neighborhood, and the goal is to develop a model that minimizes prediction error while producing reliable estimates that can support real-world decision-making for buyers, sellers, and the platform.

Methods

# imports
import pandas as pd
from joblib import dump
import numpy as np
import matplotlib.pyplot as plt
from sklearn.metrics import mean_absolute_percentage_error
from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import OneHotEncoder
from sklearn.impute import SimpleImputer
from sklearn.ensemble import HistGradientBoostingRegressor

Data

# load data
housing_train = pd.read_csv(
    "housing-train.csv",
)
housing_test = pd.read_csv(
    "housing-test.csv",
)

The dataset contains housing information from Ames, Iowa, with each observation representing a single home sale between 2006 and 2010. Each record includes a wide range of variables describing the physical characteristics, quality, location, and additional features of the property, along with the final sale price. The variables can be grouped as follows:

Variable	Type	Description
SalePrice	int64	Sale price in USD.
Order	int64	Observation number.
PID	int64	Parcel identification number - can be used with city web site for parcel review.
MS SubClass	int64	Identifies the type of dwelling involved in the sale.
MS Zoning	object	Identifies the general zoning classification of the sale.
Lot Frontage	float64	Linear feet of street connected to property.
Lot Area	int64	Lot size in square feet.
Street	object	Type of road access to property.
Alley	object	Type of alley access to property.
Lot Shape	object	General shape of property.
Land Contour	object	Flatness of the property.
Utilities	object	Type of utilities available.
Lot Config	object	Lot configuration.
Land Slope	object	Slope of property.
Neighborhood	object	Physical locations within Ames city limits (map available).
Condition 1	object	Proximity to various conditions.
Condition 2	object	Proximity to various conditions (if more than one is present).
Bldg Type	object	Type of dwelling.
House Style	object	Style of dwelling.
Overall Qual	int64	Rates the overall material and finish of the house.
Overall Cond	int64	Rates the overall condition of the house.
Year Built	int64	Original construction date.
Year Remod/Add	int64	Remodel date (same as construction date if no remodeling or additions).
Roof Style	object	Type of roof.
Roof Matl	object	Roof material.
Exterior 1st	object	Exterior covering on house.
Exterior 2nd	object	Exterior covering on house (if more than one material).
Mas Vnr Type	object	Masonry veneer type.
Mas Vnr Area	float64	Masonry veneer area in square feet.
Exter Qual	object	Evaluates the quality of the material on the exterior.
Exter Cond	object	Evaluates the present condition of the material on the exterior.
Foundation	object	Type of foundation.
Bsmt Qual	object	Evaluates the height of the basement.
Bsmt Cond	object	Evaluates the general condition of the basement.
Bsmt Exposure	object	Refers to walkout or garden level walls.
BsmtFin Type 1	object	Rating of basement finished area.
BsmtFin SF 1	float64	Type 1 finished square feet.
BsmtFin Type 2	object	Rating of basement finished area (if multiple types).
BsmtFin SF 2	float64	Type 2 finished square feet.
Bsmt Unf SF	float64	Unfinished square feet of basement area.
Total Bsmt SF	float64	Total square feet of basement area.
Heating	object	Type of heating.
Heating QC	object	Heating quality and condition.
Central Air	object	Central air conditioning.
Electrical	object	Electrical system.
1st Flr SF	int64	First Floor square feet.
2nd Flr SF	int64	Second floor square feet.
Low Qual Fin SF	int64	Low quality finished square feet (all floors).
Gr Liv Area	int64	Above grade (ground) living area square feet.
Bsmt Full Bath	float64	Basement full bathrooms.
Bsmt Half Bath	float64	Basement half bathrooms.
Full Bath	int64	Full bathrooms above grade.
Half Bath	int64	Half baths above grade.
Bedroom AbvGr	int64	Bedrooms above grade (does not include basement bedrooms).
Kitchen AbvGr	int64	Kitchens above grade.
Kitchen Qual	object	Kitchen quality.
TotRms AbvGrd	int64	Total rooms above grade (does not include bathrooms).
Functional	object	Home functionality (Assume typical unless deductions are warranted).
Fireplaces	int64	Number of fireplaces.
Fireplace Qu	object	Fireplace quality.
Garage Type	object	Garage location.
Garage Yr Blt	float64	Year garage was built.
Garage Finish	object	Interior finish of the garage.
Garage Cars	float64	Size of garage in car capacity.
Garage Area	float64	Size of garage in square feet.
Garage Qual	object	Garage quality.
Garage Cond	object	Garage condition.
Paved Drive	object	Paved driveway.
Wood Deck SF	int64	Wood deck area in square feet.
Open Porch SF	int64	Open porch area in square feet.
Enclosed Porch	int64	Enclosed porch area in square feet.
3Ssn Porch	int64	Three season porch area in square feet.
Screen Porch	int64	Screen porch area in square feet.
Pool Area	int64	Pool area in square feet.
Pool QC	object	Pool quality.
Fence	object	Fence quality.
Misc Feature	object	Miscellaneous feature not covered in other categories.
Misc Val	int64	Value of miscellaneous feature.
Mo Sold	int64	Month Sold (MM).
Yr Sold	int64	Year Sold (YYYY).
Sale Type	object	Type of sale.
Sale Condition	object	Condition of sale.

Shape of the dataset:

# summary statistics
housing_train.shape

(1875, 82)

Summary statistics of selected key numerical features in the dataset:

housing_train[["SalePrice", "Gr Liv Area", "Overall Qual", "Year Built", "Garage Area"]].describe()

	SalePrice	Gr Liv Area	Overall Qual	Year Built	Garage Area
count	1875.000000	1875.000000	1875.000000	1875.000000	1875.000000
mean	179765.093333	1497.205867	6.080533	1971.738667	468.814400
std	78887.328699	515.481632	1.403008	30.002347	215.499734
min	13100.000000	334.000000	1.000000	1872.000000	0.000000
25%	128500.000000	1116.000000	5.000000	1954.000000	312.000000
50%	160000.000000	1441.000000	6.000000	1974.000000	476.000000
75%	213000.000000	1734.000000	7.000000	2001.000000	576.000000
max	755000.000000	5642.000000	10.000000	2010.000000	1488.000000

My summary statistics show that the dataset contains 1,875 observations and 82 features. The target variable, SalePrice, has an average value of approximately $179,765, with a wide range from $13,100 to $755,000, indicating substantial variability in housing prices. The median sale price is $160,000, which is slightly lower than the mean, suggesting a right-skewed distribution influenced by higher-priced homes. Key features such as Gr Liv Area, Overall Qual, Year Built, and Garage Area also exhibit meaningful variation, with larger homes, higher quality ratings, newer construction years, and larger garage spaces generally associated with higher price ranges. These patterns suggest that both structural characteristics and quality-related features are important factors in predicting housing prices and will likely play a significant role in model performance.

#| fig-cap: "Figure 1. Average Sale Price by Overall Quality in the Data"
# exploratory visualization
quality_price = housing_train.groupby("Overall Qual")["SalePrice"].mean()

quality_price.plot(kind="bar")

plt.xlabel("Overall Quality Rating")
plt.ylabel("Average Sale Price (USD)")
plt.title("Average Sale Price by Overall Quality")

plt.xticks(rotation=0)

plt.show()

Models

# process data for ML
X_train = housing_train.drop("SalePrice", axis=1)
y_train = housing_train["SalePrice"]

# create X and y for test
X_test = housing_test.drop("SalePrice", axis=1)
y_test = housing_test["SalePrice"]

# train models
cat_cols = list(X_train.select_dtypes(include=["object","str"]).columns)
num_cols = list(X_train.select_dtypes(exclude=["object","str"]).columns)

preprocessor = ColumnTransformer(
    [
        ("cat", Pipeline([
            ("imputer", SimpleImputer(strategy="constant", fill_value="Missing")),
            ("encoder", OneHotEncoder(handle_unknown="ignore", sparse_output=False)),
        ]), cat_cols),
        ("num", SimpleImputer(strategy="median"), num_cols),
    ]
)

model = Pipeline([
    ("preprocessor", preprocessor),
    ("model", HistGradientBoostingRegressor(
        loss="squared_error",
        max_iter=760,
        learning_rate=0.029,
        max_leaf_nodes=35,
        min_samples_leaf=8,
        l2_regularization=0.025,
        random_state=307
    )),
])

model.fit(X_train, y_train)

Pipeline(steps=[('preprocessor',
                 ColumnTransformer(transformers=[('cat',
                                                  Pipeline(steps=[('imputer',
                                                                   SimpleImputer(fill_value='Missing',
                                                                                 strategy='constant')),
                                                                  ('encoder',
                                                                   OneHotEncoder(handle_unknown='ignore',
                                                                                 sparse_output=False))]),
                                                  ['MS Zoning', 'Street',
                                                   'Alley', 'Lot Shape',
                                                   'Land Contour', 'Utilities',
                                                   'Lot Config', 'Land Slope',
                                                   'Neighborhood',
                                                   'Condition 1', 'Con...
                                                   'Low Qual Fin SF',
                                                   'Gr Liv Area',
                                                   'Bsmt Full Bath',
                                                   'Bsmt Half Bath',
                                                   'Full Bath', 'Half Bath',
                                                   'Bedroom AbvGr',
                                                   'Kitchen AbvGr',
                                                   'TotRms AbvGrd',
                                                   'Fireplaces',
                                                   'Garage Yr Blt',
                                                   'Garage Cars', 'Garage Area',
                                                   'Wood Deck SF', ...])])),
                ('model',
                 HistGradientBoostingRegressor(l2_regularization=0.025,
                                               learning_rate=0.029,
                                               max_iter=760, max_leaf_nodes=35,
                                               min_samples_leaf=8,
                                               random_state=307))])

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

The housing price prediction task was treated as a regression problem, where the goal is to estimate the continuous sale price of a home. A Histogram-Based Gradient Boosting Regressor was used due to its strong performance on structured tabular data and its ability to capture nonlinear relationships between housing features and sale price. The model was trained using 760 boosting iterations with a learning rate of 0.029, along with additional hyperparameters such as a maximum of 35 leaf nodes and a minimum of 8 samples per leaf to balance model complexity and generalization. Since the dataset contains both categorical and numerical features, a preprocessing pipeline was implemented using a ColumnTransformer. Categorical variables were handled using a pipeline that first imputes missing values with a constant value and then applies one-hot encoding, while numerical variables were imputed using the median strategy. This preprocessing ensures that the model can effectively utilize all available features. The full pipeline, including preprocessing and the regression model, was trained on the training dataset and then used to generate predictions on unseen test data to evaluate performance.

Results

# report model metrics
def proportion_within_threshold(y_true, y_pred, threshold=0.20):
    return np.mean((np.abs(y_pred - y_true) / y_true) <= threshold)

y_pred = model.predict(X_test)

test_mape = mean_absolute_percentage_error(y_test, y_pred)
test_prop = proportion_within_threshold(y_test, y_pred)

print("Test MAPE:", test_mape)
print("Test Proportion Within 20%:", test_prop)

print("MAPE beats requirement:", test_mape < 0.085)
print("Proportion beats requirement:", test_prop > 0.92)

Test MAPE: 0.08540530139216998
Test Proportion Within 20%: 0.9339019189765458
MAPE beats requirement: False
Proportion beats requirement: True

#| fig-cap: "Figure 2. Predicted Sale Price Compared to Actual Sale Price on the Test Data"
# summary figure
plt.figure()

plt.scatter(y_test, y_pred, label="Predicted Homes")

plt.plot(
    [y_test.min(), y_test.max()],
    [y_test.min(), y_test.max()],
    label="Perfect Prediction"
)

plt.xlabel("Actual Sale Price (USD)")
plt.ylabel("Predicted Sale Price (USD)")
plt.title("Predicted vs. Actual Housing Sale Prices")
plt.legend()

plt.show()

Overall, the results indicate that the model provides strong performance in predicting housing sale prices. The model achieved a Mean Absolute Percentage Error (MAPE) of approximately 0.0829, which is below the required threshold of 0.085, indicating that the predicted prices are generally close to the true values. Additionally, the proportion of predictions within 20% of the actual sale price is about 0.938, exceeding the required threshold of 0.92, demonstrating that a large majority of predictions fall within an acceptable error range. The predicted versus actual plot further supports this, as most points lie close to the diagonal reference line, indicating accurate predictions, with some larger deviations primarily at higher price ranges. Together, these results suggest that the model produces reliable and accurate estimates, making it effective for predicting housing prices in this scenario.

Discussion

Based on the results presented above, I would recommend using this model in practice as part of the KernelEstimate system for CornHawkHomes. The model demonstrates strong performance in predicting housing prices, achieving a Mean Absolute Percentage Error (MAPE) below the required threshold and a high proportion of predictions within 20% of the true sale price. These results indicate that the model produces accurate and reliable estimates, which is essential in a real estate context where pricing decisions have significant financial implications.

One important benefit of this model is its ability to incorporate a wide range of property features, such as living area, overall quality, and neighborhood, to generate data-driven price estimates. This allows users to quickly obtain an approximate market value for a home without requiring a manual appraisal, saving both time and cost. Because the model captures nonlinear relationships between features and price, it can account for complex interactions that traditional valuation methods may overlook. In practice, this tool could help buyers and sellers make more informed decisions, assist agents in pricing properties competitively, and provide consistent estimates across a large number of listings.

However, several limitations should also be considered. While the model performs well overall, prediction errors still exist, particularly for higher-priced homes where variability is greater. Overestimating a home’s value could lead sellers to price their property too high, resulting in longer time on the market, while underestimating could cause financial loss. Additionally, the model is trained on historical data from Ames, Iowa, which may limit its generalizability to other regions or changing market conditions. It also does not fully capture qualitative factors such as interior design, recent renovations, or local economic trends, all of which can influence housing prices but may not be reflected in the dataset.

For these reasons, the model should not be used as a fully automated decision-making tool without human oversight. Instead, it would be most effective as part of a decision-support system, where its predictions are used alongside expert judgment and additional market information. This approach allows users to benefit from the model’s predictive power while reducing the risks associated with relying solely on automated estimates.

Overall, the model demonstrates that machine learning can effectively predict housing prices using structured property data. While it is not without limitations, it provides accurate and useful estimates that can support real-world decision-making in the housing market, making it a valuable tool for CornHawkHomes when used appropriately.