Predictive Analysis of Lagos Housing Prices using Supervised Learning Algorithms
- 1Department of Statistics, Faculty of Physical Sciences, Nnamdi Azikiwe University, P.O. Box 5025 Awka, Nigeria
- 2Department of Computer Science, Faculty of Physical Sciences Nnamdi Azikiwe University, Awka, Nigeria
- 3Department of Statistics, School of Chemical Engineering and Physical Sciences, Lovely Professional University, Punjab, India
Abstract
Accurate housing price prediction remains a formidable challenge in rapidly urbanizing areas because of heterogeneous property characteristics and volatile market conditions. This study re-examines supervised machine learning approaches for predicting property prices in Lagos State, Nigeria, using a publicly available dataset comprising 9,784 property-sale listings from Kaggle. Three predictors reported in previous work, bathrooms, toilets, and parking spaces were found to be absent from the source data and were therefore excluded. Bedroom count was retained, together with an explicit indicator for listings for which bedroom count is not applicable (e.g., land parcels), while neighborhood and house type were retained as categorical predictors. Seven regression model Multiple Linear Regression, Ridge Regression, Random Forest, Gradient Boosting, XGBoost, CatBoost, and Support Vector Regression were evaluated using a leakage-safe pipeline, in which price-outlier capping and categorical encoding were fitted exclusively on the training data. Model performance was assessed using a fixed 80/20 hold-out split and 5-fold cross-validation repeated 10 times with paired statistical comparisons. Random Forest achieved the best overall performance, with a repeated-cross-validation R² of 0.239 ± 0.039. It significantly outperformed Gradient Boosting in paired RMSE comparisons (p = 0.0043) but was statistically indistinguishable from CatBoost (p = 0.606). Permutation-importance analysis identified neighborhood as the overwhelmingly dominant predictor of price across all evaluated models, whereas house type and bedroom count made substantially smaller, secondary contributions. Nevertheless, all models explained only a limited proportion of price variation (R² ≈ 0.24), indicating substantial predictive uncertainty. Additional analyses showed that log-transforming the target and employing a quantile (median) loss reduced mean absolute error but did not improve R², highlighting an inherent trade-off between mean- and median-oriented predictive objectives. Systematic under-prediction of the most expensive properties persisted across all evaluated approaches, reflecting the absence of important property-level predictors such as floor area, structural condition, title documentation, and amenities. Overall, the findings demonstrate that location provides the strongest predictive signal in the available dataset, while the limited explanatory power underscores the need for richer property-level and market-context variables for reliable housing valuation in Lagos State.
Related
- Automatic Monitoring of Rock-Slope Failures Using Distributed Acoustic Sensing and Semi-Supervised Learning
- DeepWaveRL: Self-Supervised Full Waveform Inversion via Reinforcement Learning
- LAND COVER AND CROP IDENTIFICATION USING MACHINE LEARNING ALGORITHMS
- Estimating aman and aus rice area in Bangladesh using sentinel-1 imagery and machine learning algorithms
- Estimation of hydrological drought using regional risk ranking and machine learning algorithms
BibTeX
@article{obulezi2026predictive,
title = {{Predictive Analysis of Lagos Housing Prices using Supervised Learning Algorithms}},
author = {Okechukwu J. Obulezi and Rose Uzoamak Paul and Chinyere P. Okechukwu},
journal = {مجلة العلوم التجارية والبيئية},
year = {2026},
month = oct,
doi = {10.21608/jcese.2026.497049.1157},
url = {https://doi.org/10.21608/jcese.2026.497049.1157},
note = {CC BY},
}
RIS
TY - JOUR
TI - Predictive Analysis of Lagos Housing Prices using Supervised Learning Algorithms
AU - Okechukwu J. Obulezi
AU - Rose Uzoamak Paul
AU - Chinyere P. Okechukwu
PY - 2026
DA - 2026/10/25
JO - مجلة العلوم التجارية والبيئية
DO - 10.21608/jcese.2026.497049.1157
UR - https://doi.org/10.21608/jcese.2026.497049.1157
AB - Accurate housing price prediction remains a formidable challenge in rapidly urbanizing areas because of heterogeneous property characteristics and volatile market conditions. This study re-examines supervised machine learning approaches for predicting property prices in Lagos State, Nigeria, using a publicly available dataset comprising 9,784 property-sale listings from Kaggle. Three predictors reported in previous work, bathrooms, toilets, and parking spaces were found to be absent from the source data and were therefore excluded. Bedroom count was retained, together with an explicit indicator for listings for which bedroom count is not applicable (e.g., land parcels), while neighborhood and house type were retained as categorical predictors. Seven regression model Multiple Linear Regression, Ridge Regression, Random Forest, Gradient Boosting, XGBoost, CatBoost, and Support Vector Regression were evaluated using a leakage-safe pipeline, in which price-outlier capping and categorical encoding were fitted exclusively on the training data. Model performance was assessed using a fixed 80/20 hold-out split and 5-fold cross-validation repeated 10 times with paired statistical comparisons. Random Forest achieved the best overall performance, with a repeated-cross-validation R² of 0.239 ± 0.039. It significantly outperformed Gradient Boosting in paired RMSE comparisons (p = 0.0043) but was statistically indistinguishable from CatBoost (p = 0.606). Permutation-importance analysis identified neighborhood as the overwhelmingly dominant predictor of price across all evaluated models, whereas house type and bedroom count made substantially smaller, secondary contributions. Nevertheless, all models explained only a limited proportion of price variation (R² ≈ 0.24), indicating substantial predictive uncertainty. Additional analyses showed that log-transforming the target and employing a quantile (median) loss reduced mean absolute error but did not improve R², highlighting an inherent trade-off between mean- and median-oriented predictive objectives. Systematic under-prediction of the most expensive properties persisted across all evaluated approaches, reflecting the absence of important property-level predictors such as floor area, structural condition, title documentation, and amenities. Overall, the findings demonstrate that location provides the strongest predictive signal in the available dataset, while the limited explanatory power underscores the need for richer property-level and market-context variables for reliable housing valuation in Lagos State.
SN - 2974-3125
SN - 2974-3117
ER -