![Sample our Built Environment journals, sign in here to start your access, Latest two full volumes FREE to you for 14 days]({"type":"image" , "src" : "/sda/5926/TOC-Subject-Sample-2021-Built-Environment.jpg"})
Abstract
This paper sheds some light on the effects of social capital variables (social network data, physical appearance, etc.) on loan-to-value (LTV), a crucial variable to evaluate systemic risk. Using a unique database created by merging several sources of data, we show that the introduction of social capital variables are shown to be statistically significantly related to LTV. In particular, Facebook likes in a month and creditworthiness are a negative determinant of LTV while beauty and certain personality traits play a role in borrowers obtaining a higher LTV. We distinguish these effects depending on the LTV variable used: loan-to-appraisal (entirely under the control of lender) and loan-to-transaction (in which the transaction price can also be influenced). As policy implications we found that social capital variables capture information that would otherwise be unobservable using only the traditional variables in the sense that they are related to information lenders may have at lending that the researchers do not observe.
Acknowledgements
The author thanks Claudia Vargas, Elena Fernández and Helena Bonnin for excellent research assistance, especially with respect to data compilation
Data
The data that support the findings of this study are available from a Real Estate Company (REC). Restrictions apply to the availability of these data, which were used under license for this study. Data are available from the authors with the permission of the REC.
Disclosure Statement
No potential conflict of interest was reported by the authors.
Notes
1 Fair Isaac Corp. The exact formula used is unknown but the following are, approximately, the weights of each component: history of previous payments (35%), use of credit (30%), length of credit history (15%), types of credit used (10%), amount of credit obtained recently (10%).
2 Zest Finance has found that loan applicants who only use uppercase or lowercase letters are less likely to repay loans.
3 The impact of other nontraditional variables used to improve models of credit default can be shown in (Jagtiani & Lemieux, Citation2018; Khandani et al., Citation2010; Kruppa et al., Citation2013; Netzer et al., Citation2019). These cases include variables from machine learning and big data.
4 This includes the uses of a French repayment method, which is a well-known method characterized by payment of constant instalments in each period wherein Interest decreases as loan periods pass and amortized capital increases in every new period.
5 Two examples of typical mortgage product during boom years. First is from the Banco de Santander during 2007. Reference index: euribor. Spread: 0.25 (4.97%). Term: 30 years Annual interest rate review. 0.5% of early amortization commission. Financing up to 80% of the housing value. Fixed payment amortization system. Offer conditional on payroll direct debit and hiring home insurance. The second is from Barclays during 2005. Reference index: euribor. Spread: 0.45 (2.90%). Term: 30 years Annual interest rate review. 0.5% of early amortization commission. Offer conditional on payroll direct debit and hiring home insurance. Fixed payment amortization system.
6 See Cecchetti et al. (Citation2011), in particular .
7 This can be done in a number of ways. First, when interest rates are as low as they are now, even homeowners with relatively recent mortgages may benefit from refinancing. Second, to take advantage of better personal financial circumstances (if you have more savings or a higher income, you may qualify for better credit conditions). Finally you can also change mortgage conditions, refinancing into fixed rate mortgages or negotiating a longer mortgage term to reduce monthly repayments (to alleviate worse financial circumstances).
8 Raya (Citation2018) examines the determinants of foreclosures in Spain.
9 For confidentiality reasons, we cannot identify the company.
10 This number excludes social housing and residential units that had some type of public subsidy.
11 ‘Registry’ refers to the Spanish Official Property Registry (‘Registro de la Propiedad’), which archives the property titles of all real estate assets.
12 Notice that prices declared to the real estate registration office do not have to coincide with market prices since there is an extended practice of using money that is undeclared to the tax authority as part of the payment in real estate transactions.
13 To be sure that the matching was properly performed, we compared the common variables available in our constructed dataset and the information from the Official Property Registry (size of the loan, appraisal price).
14 A decrease in the accuracy of credit scores based on borrowers' credit history for predicting loan delinquency has been proved in a Fitch study in the US. Indeed, some banks have abandoned credit scores for other risk analyses based on, among other factors, borrowers' employment. These more reliable scores use as inputs variables such as the ones we use here regarding borrowers' labour status.
15 We also checked other social networks, such as Instagram, but fewer than 2% of the individuals had an active Instagram account.
16 In 2017, Facebook had 1.97 billion monthly active users. The seventh-ranked photosharing app, Instagram, had over 600 million monthly active accounts. LinkedIn had 106 million monthly active accounts.
17 In particular, Facebook information is observed from 2006 to 2012 while LinkedIn information is observed from 2010 to 2019.
18 Other variables computed, such as frequency of profile updates or the number of friends, depend crucially on the amount of information the individual sets to be publicly accessible. We also compiled information on Twitter and Instagram (followers, number of posts, etc.), but, as we have pointed out previously, the sample was very small.
19 The Cohen’s kappas and t tests suggest a fairly high level of agreement across raters, although for every item, raters 1 and 5 were relatively generous, while rater 6 was relatively ungenerous.
20 A variance comparison test also could not reject the null hypothesis of equal variance.
21 In fact, the major source of drop outs were people with a private social network profile. Thus, the selection bias, if any, is not due to differences in activity.
22 Note that loan-to-transaction equations are models that also approximate the probability of future default, because the correlation between the loan-to-value ratio and the future probability of default is well known (Wong et al., Citation2011). In this sense, sociodemographic and social capital variables at time “t” are implicitly explaining default at time “t + n”; i.e., we are using social capital data to predict loan performance. We have no information about the effective future default of these loans. However, we know for a very small subsample whether the dwelling ends in foreclosure. We present some results in this respect in the Appendix 2 ( and )
23 The variable is not statistically significant. In fact, the housing price trend is captured by the time dummies. We have excluded this variable in the models presented in the Appendix.
24 In the Appendix 1 () we present, for a subsample, the same estimation adding text analysis variables as predictors of LTT, LTA and ATT.
25 Models without social capital variables are presented in of the Appendix.
26 We have also computed the root of mean squared error (RMSE). In all cases, the difference between the models with and without social capital variables is similar to that found in terms of R-squared. That is, the LTT is reduced from 19.05 to 16.86 when we introduce social capital variables. For LTA, the reduction is from 19.14 to 17.3, and for ATT, RMSE is reduced from 19.06 to 16.7.
27 In this respect, in , we present results of the estimation where we do not yet include social networks indicators. Again, variables like education, age, income, permanent occupation are all statistically insignificant.
28 In fact, the only explanatory variable that is significant in both equations is the characteristic of being from Spain, which reduces LTA and LTT. Diaz-Serrano and Raya (Citation2014) explore discrimination in the Spanish mortgage market.
29 Similarly, Dorfleitner et al. (Citation2016) examine information derived from the description texts to the probability of successful funding and to the default probability in peer-to-peer lending for two leading European platforms.
30 Obviously, with 121 observations, we have a problem of representativeness. However, as will be seen below, we obtain additional evidence from a recent literature result.
31 For 1 observation, we do not know the selling price.
32 For the whole sample of 3,307 observations, the number of homes is 318 (65 of which were foreclosed upon).