1# RECORD 8a99113b012c7dc5288863734eba9456a85fd9062# TITLE: Investigating the impact of property characteristics, cost of living, and environmental factors on rental prices in Baidoa’s climate-affected real estate market: a hybrid approach using hedonic regression and neural networks3# YEAR: 2025.0 VENUE: Frontiers in Sustainable Cities DOI: 10.3389/frsc.2025.15117614# AUTHORS: Mohamed Ibrahim Nor; Buba Audu; A. Mohamed5# PAGES: 15 CANDIDATE TABLE PAGES (1-based): [7, 8, 9]6# ABSTRACT: Understanding the determinants of residential rental prices is crucial for policymakers, investors, and real estate practitioners. This study investigates the influence of property value, property characteristics, cost of living, political stability, essential services, and environmental factors on rental prices in Baidoa city. Additionally, the research compares different modeling approaches to enhance rental price forecasting.A dual-method approach was employed, integrating hedonic regression analysis and artificial neural network (ANN) models to analyze rental values. The dataset includes key variables such as the number of bedrooms, essential services, cost of living, and environmental conditions. The predictive performance and interpretability of both models were assessed to determine their effectiveness in rental price estimation.The findings reveal that rental prices are significantly influenced by the number of bedrooms, essential services (e.g., electricity), cost of living, and environmental conditions. However, political stability and displacement did not exhibit significant effects. While hedonic regression provided clear, interpretable insights into direct predictors, ANN models captured nonlinear interactions and demonstrated superior prediction accuracy. Nevertheless, the ANN model exhibited mixed performance, with 53% of cases underperforming and 47% exceeding predictions, highlighting the need for improved precision in forecasting.The study emphasizes the importance of a mixed-method approach in rental price forecasting. Policymakers should integrate econometric and machine learning models to refine housing policies and ensure fair market regulations. Investors and property owners can leverage these findings to optimize rental pricing strategies, while real estate practitioners can benefit from data-driven decision-making. This research contributes to the real estate valuation literature by bridging traditional econometric analysis with advanced machine learning techniques. The study validates the applicability of hedonic pricing and information asymmetry theories within an emerging market context, offering a more comprehensive understanding of rental price determinants.789===== PAGE 1 (score 7, numeric-paren 0, table-words 0, var-words 7) =====1011Frontiers in Sustainable Cities120113frontiersin.org14Investigating the impact of 15property characteristics, cost of 16living, and environmental factors 17on rental prices in Baidoa’s 18climate-affected real estate 19market: a hybrid approach using 20hedonic regression and neural 21networks22Mohamed Ibrahim Nor 1*, Buba Audu 2 and 23Abdullahi Dahir Mohamed 3241 Institute of Climate and Environment (ICE), SIMAD University, Mogadishu, Somalia, 2 Buruuj 25Construction and Real Estate, Mogadishu, Somalia, 3 Faculty of Economics and Management Sciences, 26Hormuud University, Mogadishu, Somalia27Introduction: Understanding the determinants of residential rental prices 28is crucial for policymakers, investors, and real estate practitioners. This study 29investigates the influence of property value, property characteristics, cost of 30living, political stability, essential services, and environmental factors on rental 31prices in Baidoa city. Additionally, the research compares different modeling 32approaches to enhance rental price forecasting.33Methods: A dual-method approach was employed, integrating hedonic 34regression analysis and artificial neural network (ANN) models to analyze rental 35values. The dataset includes key variables such as the number of bedrooms, 36essential services, cost of living, and environmental conditions. The predictive 37performance and interpretability of both models were assessed to determine 38their effectiveness in rental price estimation.39Results: The findings reveal that rental prices are significantly influenced by 40the number of bedrooms, essential services (e.g., electricity), cost of living, 41and environmental conditions. However, political stability and displacement 42did not exhibit significant effects. While hedonic regression provided clear, 43interpretable insights into direct predictors, ANN models captured nonlinear 44interactions and demonstrated superior prediction accuracy. Nevertheless, the 45ANN model exhibited mixed performance, with 53% of cases underperforming 46and 47% exceeding predictions, highlighting the need for improved precision in 47forecasting.48Discussion: The study emphasizes the importance of a mixed-method approach 49in rental price forecasting. Policymakers should integrate econometric and 50machine learning models to refine housing policies and ensure fair market 51regulations. Investors and property owners can leverage these findings to 52optimize rental pricing strategies, while real estate practitioners can benefit 53from data-driven decision-making. This research contributes to the real estate 54valuation literature by bridging traditional econometric analysis with advanced 55machine learning techniques. The study validates the applicability of hedonic 56OPEN ACCESS57EDITED BY58Jianquan Cheng, 59Manchester Metropolitan University, 60United Kingdom61REVIEWED BY62Olatunji Shobande, 63University of Aberdeen, United Kingdom64Debora Anelli, 65Politecnico di Bari, Italy66*CORRESPONDENCE67Mohamed Ibrahim Nor 68 m.ibrahim@simad.edu.so69RECEIVED 15 October 202470ACCEPTED 03 March 202571PUBLISHED 19 March 202572CITATION73Nor MI, Audu B and Mohamed AD (2025) 74Investigating the impact of property 75characteristics, cost of living, and 76environmental factors on rental prices in 77Baidoa’s climate-affected real estate market: 78a hybrid approach using hedonic regression 79and neural networks.80Front. Sustain. Cities 7:1511761.81doi: 10.3389/frsc.2025.151176182COPYRIGHT83© 2025 Nor, Audu and Mohamed. This is an 84open-access article distributed under the 85terms of the Creative Commons Attribution 86License (CC BY). The use, distribution or 87reproduction in other forums is permitted, 88provided the original author(s) and the 89copyright owner(s) are credited and that the 90original publication in this journal is cited, in 91accordance with accepted academic 92practice. No use, distribution or reproduction 93is permitted which does not comply with 94these terms.95TYPE Original Research96PUBLISHED 19 March 202597DOI 10.3389/frsc.2025.15117619899100===== PAGE 7 (score 44, numeric-paren 6, table-words 12, var-words 8) =====101102Nor et al.10310.3389/frsc.2025.1511761104Frontiers in Sustainable Cities10507106frontiersin.org107learning algorithm with a suitable loss function to minimize the 108prediction error. An ANN’s functional relationship can be expressed 109as follows:110 11111122 31134114(115,116,117,118p119y120f x x x121x122x123=124……………125where 1x , 212631274128,129,130x131x132x ………………. p133x are the input variables and y 134is a target variable.135The MLP algorithm relies heavily on hidden units, which compute 136the weighted sum of their inputs to obtain their scalar net activation. 137The net activation is the inner product of the inputs and the weights 138at the hidden unit, which is expressed as follows:139 140014111420143d144d145T146j147i148ji149i150ji151j152i153i154net155x w156w157x w158W X159=160=161=162+163=164≡165∑166∑167This illustrates that every hidden unit emits an output that 168varies according to how it is activated. Table 1 presents the results 169of the experiment, the inputs (predictors), and the sampling 170strategy. The training, test, and validation sets of the experimental 171data are used to obtain the best results. The sample size of each 172component of the data is shown in Table 1. Similarly, the 173architecture of the MLP algorithm is explained in Table 1 and 174Figure 1. An MLP with at least three hidden layers and a maximum 175of ten layers is used in this investigation. In addition, five networks 176are 177retained, 178while 17920 180networks 181are 182used 183for 184the 185training experiment.186Results of hedonic regression187The results of the hedonic regression analysis indicate that 188property features, access to essential services, economic factors, 189political conditions, and displacement have significant impacts on 190rental property rates. Our model is robust because it explains a 191significant proportion of the variance in rental values (R2 = 0.96, 192F = 105.46, p < 0.000). It has also been established that the model has 193TABLE 2 Hedonic model estimates.194Independent variables195Coefficient196t-statistic197VIF198Cohens f2199Constant200143.582014.97202Residential property price (RPP)2030.022040.612051.152069.07207Number of bedrooms (NBR)2080.092092.72***2101.42119.15212Dedicated parking space (DPS)213−0.01214−0.362151.642169.07217Residential property size (RPS)2180.012190.262201.242219.15222Access to electricity (AEL)223−0.18224−5.96***2251.182269.09227Environmental factors228−0.12229−3.71***2301.392319.14232Cost of living (COL)233−0.26234−4.85***2353.652369.04237Political stability (POS)238−0.02239−0.662401.412419.13242Displacement (DIS)2430.052441.262451.682469.11247Interaction1 (I1)2480.252494.49***2503.932518.86252Interaction2 (I2)2530.6725412.24***2553.822568.85257R-squared (R2)2580.96259Adjusted (Adj. R2)2600.92261F-statistics (prob)262105.462630.00264Durbin-Watson test2652.052660.965267Heteroscedasticity test268No heteroscedasticity concerns were found in the data.269Normality test: Kolmogorov–Smirnov2700.072710.596272Mean squared error (MSE)2731507.13274Mean absolute error (MAE)27529.38276Dependent variable: Residential property monthly rentals (RPMR).277**p < 0.05; ***p < 0.01.278TABLE 3 ANN test results (forecasts).279Model/item280Number of 281items282% of items283Item: less 284than 10%285% of items 286(within 10%)287Item: more 288than 10%289% of items 290(more than 29110%)292Over-performed293729423.33295229628.57297529871.43299Under-performed3002330176.67302330313.043042030586.96306Total30730308100309531010031125312100313314315===== PAGE 8 (score 19, numeric-paren 0, table-words 5, var-words 7) =====316317Nor et al.31810.3389/frsc.2025.1511761319Frontiers in Sustainable Cities32008321frontiersin.org322high explanatory power, indicating the significant influence of the 323independent variables on monthly rental values.324The underlying assumptions of the regression model were 325carefully tested to confirm the robustness of the findings. The residuals 326of the model were normally distributed, as verified by the 327Kolmogorov–Smirnov test (stat. = 0.07; p-value = 0596), confirming 328that the assumption of a normal distribution was satisfied. 329Furthermore, the Durbin-Watson test is employed to ensure the 330absence of significant autocorrelation among residuals, and as 331reported in Table 2, the Durbin-Watson test (DW = 2.45, p = 0.053) 332affirmed the independence of the residuals of the model. While 333checking for multicollinearity, we found that all variance inflation 334factor (VIF) values are below five (5), signifying that there is no 335considerable redundancy among the predictors. Moreover, as reported 336in Table 2, the Cohen’s f2 test indicates a strong effect of the predictors 337on the dependent variable, well above the established benchmarks. 338These tests validate the integrity of the analysis, highlighting the 339reliability of the model.340Remarkably, the study revealed that the number of bedrooms in a 341property has a direct and positive effect on property rental prices 342(b = 0.09, t = 2.72, p < 0.00), implying that larger homes demand 343higher rents. In addition, properties with greater area demand an 344increase in rental prices (b = 0.07, t = 2.26, p < 0.05), confirming that 345spatial dimensions play a critical role in rental valuation. In contrast, 346the study reveals significant negative effects from socioeconomic 347factors such as high cost of living (b = −0.26, t = −4.85, p < 0.01) and 348political instability (b = −0.08, t = −2.48, p < 0.05), which reduce 349property rental prices, exemplifying the burden of economic strain on 350rental affordability, political risks and uncertainties, which might 351prevent long-term residential commitments or weaken consumer 352confidence. These findings underscore the sensitivity of property 353rental markets to economic and political climates, showing the 354necessity for good governance and effective economic policies to 355sustain healthier rental markets.356Shockingly, despite Baidoa city’s vulnerability to ecological shocks, 357the analysis indicates that environmental factors such as significant 358climate change and sustainable practices do not significantly affect 359rental decisions. These results may reflect a gap in tenant awareness or 360prioritization regarding environmental sustainability and resilience in 361the residential property market. Moreover, the study revealed a 362positive relationship between the influx of internally displaced persons 363(IDPs) and property rental prices (b = 0.07, t = 2.18, p < 0.05), 364demonstrating that demographic pressures from displacement can 365aggravate housing demand, thus increasing property rental prices in 366affected regions. The interaction terms further illuminate the 367compound effects of combined variables on property rental prices. 368Remarkably, the interaction between property rental rates and the cost 369of living significantly increases rental rates (b = 0.21, t = 3.52, p < 0.01), 370FIGURE 1371ANN model.372TABLE 4 Comparison metrics.373Model374Mean squared error (MSE)375Mean absolute error (MAE)376Regression model3771507.1337829379ANN model3806920.8338165.75382383384===== PAGE 9 (score 20, numeric-paren 0, table-words 4, var-words 9) =====385386Nor et al.38710.3389/frsc.2025.1511761388Frontiers in Sustainable Cities38909390frontiersin.org391whereas the synergy between property rental rates and essential 392services, principally electricity, is even more noticeable (b = 0.74, 393t = 12.48, p < 0.01). These results suggest that while individual factors 394influence property rental prices, their interdependencies can intensify 395impacts, underscoring the complexity of rental market dynamics. This 396comprehensive analysis assists policymakers and stakeholders in 397understanding the multidimensional effects on property rental prices, 398providing a basis for informed decision-making intended to 399accomplish sustainable urban development and housing affordability.400The results of the hedonic regression analysis indicate that several 401factors significantly influence residential property prices. Our model 402demonstrates robust explanatory power, accounting for a substantial 403proportion of the variance in property values (R2 = 0.96, F = 105.46, 404p < 0.00), signaling the significant impact of the independent variables 405on monthly rentals of residual properties. The underlying assumptions 406of the regression model were thoroughly assessed to ensure the 407reliability of the results. The residuals of the model were found to 408be normally distributed, as confirmed by the Kolmogorov–Smirnov 409test (stat. = 0.07, p = 0.596), ensuring the normality assumption was 410met. Additionally, the Durbin-Watson test (DW = 2.05, p = 0.965) 411confirmed the absence of autocorrelation among the residuals, 412affirming the independence of the error terms. Regarding 413multicollinearity, all variance inflation factor (VIF) values were below 414five (<5), suggesting no significant redundancy among the predictors. 415The Cohen’s f2 values further indicate a substantial effect of the 416predictors on the dependent variable, underscoring the robustness of 417the model. Among the significant variables, the number of bedrooms 418(b = 0.09, t = 2.72, p < 0.01) shows a positive relationship with 419property prices, suggesting that properties with more bedrooms tend 420to have higher values. The cost of living (b = −0.26, t = −4.85, p < 0.01) 421and access to electricity (b = −0.18, t = −5.96, p < 0.01) were found to 422have a significant negative impact on property prices, indicating that 423higher costs of living and limited access to electricity contribute to 424lower property values. The environmental factors (b = −0.12, 425t = −3.71, p < 0.01) also exert a significant negative effect, suggesting 426that environmental challenges, such as climate change, negatively 427impact property prices. Furthermore, the interaction between 428monthly rentals of residual properties and the cost of living (b = 0.25, 429t = 4.49, p < 0.01) and the interaction between monthly rentals of 430residual properties and access to electricity (b = 0.67, t = 12.24, 431p < 0.01) were both found to significantly amplify monthly rentals of 432residual properties. These interaction terms emphasize how the 433combination of economic factors and essential services can intensify 434their effects on property prices. These findings highlight the key 435drivers of residential monthly rentals, with economic and service-436related factors playing a critical role. Policymakers and urban planners 437should consider these insights to develop effective housing strategies 438that address the complexities of property price dynamics.439Results of the artificial neural network440In this study, we employed an artificial neural network (ANN) 441technique to forecast property rental values for Baidoa city using a 442multilayer perceptron (MLP) model. In this study, data preprocessing 443was conducted to ensure the integrity and usability of the dataset 444before training the artificial neural network (ANN) model. The dataset 445was partitioned into three subsets: a training set comprising 70 446samples (60%), a testing set with 30 samples (25%), and a validation 447set containing 18 samples (15%). A total of 11 predictors were 448identified (from the regression results), with numerical features 449standardized to ensure uniformity and categorical variables encoded 450to facilitate machine learning processing. The ANN model was 451structured with an input layer consisting of 11 nodes, corresponding 452to the number of predictors, and a hidden layer architecture ranging 453between 3 and 10 layers, with the final optimized model incorporating 454five hidden layers. The ReLU activation function was applied to the 455hidden layers to introduce non-linearity, while a linear activation 456function was used in the output layer to predict the Monthly Property 457Rentals (MPR). The Adam optimizer was employed to adjust model 458weights efficiently, and the mean squared error (MSE) loss function 459was used to measure prediction accuracy. The model underwent 2000 460training iterations, ultimately achieving an MSE of 6920.83 and a 461mean absolute error (MAE) of 65.75, indicating an optimal predictive 462performance (as illustrated in Table 3).463A significant valuation difference is evident in Figure 1, indicating 464significant fluctuations in prediction accuracy, with both 465underestimations and overestimations evident. Thirty property cases 466were critically evaluated to determine the effectiveness of the ANN 467model for predicting Monthly Property Rentals (MPR). The evaluation 468yielded two primary outcomes regarding the model’s performance: 469underperformance and over-performance against actual rental values. 470Among the 30 cases, 16 properties (53%) exhibited some degree of 471underperformance relative to their actual rental values, while 14 cases 472(47%) demonstrated over-performance (see Table 4). The data indicate 473a higher occurrence of underperformance in the predictions, where 474the majority (71%, or 11 out of 16 cases) underestimated rental prices 475by less than 10%, whereas the remaining five (5) cases exceeded this 476threshold, suggesting that moderate underestimation existed at 31%. 477Specifically, 16 properties—accounting for over half of the evaluated 478samples—failed to meet their actual rental values, signaling potential 479revenue discrepancies for property owners and investors. The 480predominance of underperformance, mostly within a 10% margin, 481reflects a conservative estimation pattern. While this approach 482mitigates the risk of rental overvaluation that could lead to unrealistic 483investment expectations, it simultaneously risks undervaluing 484properties, potentially limiting revenue generation opportunities and 485affecting the overall financial viability of rental investments. The 486findings from the ANN model evaluation highlight critical insights for 487stakeholders in the rental market, indicating both opportunities and 488challenges associated with current predictive modeling approaches. 489These insights inform strategic decision-making in pricing, 490forecasting, investment, and market entry strategies.491Pricing strategies adjustment492The analysis reveals that a significant proportion (53%) of the 493forecasts were lower than actual market rental values, indicating a 494persistent bias toward conservative pricing strategies among property 495owners and managers. While this conservative approach helps 496mitigate the risk of high vacancy rates, it also results in lower 497achievable rental earnings. To better align with actual market 498conditions, property managers should consider refining their pricing 499strategies, ensuring they accurately reflect market demand while 500maintaining competitiveness. Overly cautious pricing may limit 501potential revenue streams, requiring a more dynamic and data-driven 502approach to rental valuation.503504505===== PAGE 10 (score 8, numeric-paren 0, table-words 2, var-words 4) =====506507Nor et al.50810.3389/frsc.2025.1511761509Frontiers in Sustainable Cities51010511frontiersin.org512Development of forecast models513The substantial deviations observed between forecasted and actual 514rental values suggest that further refinements to predictive modeling 515are necessary. In some cases, rental values were significantly 516overestimated, while in others, they were considerably underestimated. 517These inconsistencies emphasize the need for incorporating a broader 518set of explanatory variables, including economic indicators, 519demographic trends, and historical market fluctuations, to enhance 520forecast accuracy. Developers should integrate real-time economic 521data reflective of the market’s conditions to improve model precision 522and bridge the gap between predicted and actual values.523Opportunities for niche investments524The findings indicate significant investment opportunities within 525specific market segments. Properties that consistently outperform 526their forecasts, particularly those exceeding predicted values by more 527than 10%, are likely undervalued and may provide higher-than-528expected returns. Investors should consider these properties as high-529potential opportunities for maximizing rental income. Conversely, 530areas where rental forecasts frequently underperform may require 531additional due diligence, as they could signal market volatility or 532external factors affecting rental stability. Identifying these trends 533enables investors to allocate capital strategically and optimize returns.534Market entry strategies and portfolio 535management536The mixed accuracy of forecasts suggests that prospective investors 537should adopt a diversified approach to portfolio management. While 538some properties align closely with predicted rental values, others exhibit 539substantial variation, either exceeding or falling short of expectations. To 540mitigate risks, investors should consider a balanced portfolio, including 541stable properties with predictable returns and speculative investments 542with higher growth potential. This strategic diversification allows market 543participants to capitalize on emerging opportunities while safeguarding 544against unforeseen market fluctuations.545Advanced market analysis and continual 546monitoring547The variability in rental market forecasts underscores the 548importance of ongoing market analysis and continuous model 549adjustments. Given the dynamic nature of the housing market, real 550estate professionals must remain vigilant and responsive to economic 551and regulatory changes that may impact rental demand. Regular 552updates to predictive models, informed by market trends and real-553time data, will enhance forecasting accuracy and enable property 554owners to make proactive adjustments. By continually refining their 555strategies, stakeholders can maintain a competitive edge and ensure 556long-term financial sustainability in the rental market (Figure 2).557Comparison of regression and ANN model 558performance559The regression model outperforms the artificial neural network 560(ANN) model in terms of predictive accuracy, as indicated by its 561significantly lower error metrics. The regression model achieves a 562Mean Squared Error (MSE) of 1507.13 and a Mean Absolute Error 563(MAE) of 29, suggesting that its predictions are closer to the actual 564values. In contrast, the ANN model exhibits a much higher MSE of 5656920.83 and an MAE of 65.75, indicating greater deviations from the 566true values (see Table 5). This suggests that, in this particular case, the 567regression model is more effective in minimizing errors and providing 568more reliable predictions compared to the ANN model.569Discussion570The main aim of this study is to investigate the influence of 571property value, property characteristics, cost of living, political 572stability, essential services and environmental factors on residential 573property rental prices in Baidoa city. This study uses a dual-method 574TABLE 5 Summary of data preprocessing, ANN model configuration, and model performance results.575Stage576Details577Data preprocessing578Training set: 70 samples (60%)579Testing set: 30 samples (25%)580Validation set: 18 samples (15%)581Feature count: 11 predictors582Normalization: standardized numerical predictors583Categorical encoding: converted non-numeric predictors584ANN (MLP) model585Input layer: 11 nodes (one for each predictor)586Hidden layers: between 3 and 10 layers (based on tuning)587Output layer: 1 node (predicting MPR)588Activation functions:589- ReLU for hidden layers590- Linear for output591Optimizer: Adam592Loss function: mean squared error (MSE)593ANN model results594Mean squared error (MSE): 6920.83595Mean absolute error (MAE): 65.75596Training iterations: 2000597Hidden layers: 5598