Payday Loan Marketing in Social Media Networks.
Traditional ways to collect opinion data are slow and expensive. The recently rising social networks provide researches real-time "Big Data" to analyze consumer attitudes and behaviors. There are several advantages of using social media for PDL monitoring: (1) People may behave and express their feelings in a different way on social media networks than they do when they respond to traditional surveys. Therefore, social media posts may contain "back-stage" information that might not be available from a traditional survey. (2) Social media networks provide researchers with "Big Data" to analyze consumer attitudes and behaviors. An average of 500 million tweets (short messages that are fewer than 140 characters) are posted on Twitter each day and 23% of online adults/19% of the entire adult population currently use Twitter (Duggan et al. 2015). Over one billion people use Facebook actively each month (Statista 2015) and 71% of adult Internet users currently use Facebook (Duggan et al. 2015). These networks yield a large sample size and rich information for research, including sentiment assessment, prevalence by geographic locations, and topics of social media posts. (3) Most social media information is real-time and can be analyzed continuously at a relatively low cost. Past studies (Dai and Hao 2017; Ginsberg et al. 2009; Salathe and Khandelwal 2011; Wong et al. 2015) have shown that using social media platforms as a real-time monitoring tool can provide valuable insights into the general public's opinions and attitudes.
The PDL activities from social media networks, including sentiment, opinion, and knowledge along with user count, favorite count, followers, impact factors, and so on, have not been studied in literature. This study seeks to fill gaps in the literature by analyzing the social media data related to PDLs, examining PDL online marketing on Twitter, and investigating the impact of state regulations on PDL marketing and sentiment among Twitter users. To our knowledge, this study is the first of its kind to analyze PDL activities on social networks.
PDL Lending Trend
Nonbank lenders offer PDLs through storefronts and over the Internet. The PDL lending industry has grown dramatically since its inception in the 1990s. The Center for Responsible Lending reports that there are over 22,000 PDL locations in the United States and the industry generates an estimated $27 billion in annual loan volume. A PDL typically carries a 14-day loan term for no more than $1,000 with a cost of $10-$20 per $100 borrowed, with annual percentage rates (APRs) often in excess of 400% (Center for Responsible Lending 2015).
Many Americans with good credit or stable income might never use a PDL. However, many other Americans are vulnerable to these high-cost loans, often because they have limited access to mainstream banking services. An estimated 27.7% of all American households representing 67.6 million adults are either "unbanked" (7.7% of all households) or "underbanked" (20% of households) (FDIC 2014). The highest unbanked rate is found among non-Asian minorities, lower-income households, younger households, and unemployed households. In 2013, e.g., 53.6% of African American and 46.4% of Hispanic households were unbanked or underbanked compared with 19.5% of White non-Black non-Hispanic households. With regard to age, 46.5% of 15-24 year olds and 37.2% of 25-36 year olds were unbanked or underbanked compared, with 23.3% for 55-64 year olds and 15.1% for 65 years or more. FDIC (2009) estimated that 16.2% of underbanked households have used PDLs and 6.6% of unbanked household have used PDLs, compared with only 3.5% of all households.
PDL lenders are expected to evolve away from a predominately storefront model given the regulatory pressure (National People's Action's 2012). Stephens Inc. (2014) reported that the total number of PDL stores peaked at 24,043 in 2007 and have since decreased to 17,862 as of 2013. Online PDLs still make up only a small percentage of overall PDL volume. FDIC (2014) reported that 15.5% of households that used PDLs in the last 12 months had taken out of online PDLs. However, online PDL lending will likely continue to increase and may eventually overtake storefront loan volume (CFPB 2013).
State Laws and Enforcement
Due to high interest rates and recurring payments for PDLs, the industry has come under increased scrutiny, resulting in regulations in many states. For example, 15 states and the District of Columbia prohibit extremely high-cost PDL lending and have no PDL storefronts due to the regulation of interest rate. A total of 27 states take a permissive stance toward PDL lending, allowing lenders to have a single-repayment loan with APRs of 391% or higher. Another nine states are considered as hybrid states, where PDL storefronts exist but there are more exacting requirements, such as lower limits on fees or loan usage, or longer repayment periods (Pew Charitable Trusts 2014). The Consumer Financial Protection Bureau (CFPB) announced in March 26, 2015 that it is considering proposing rules that would regulate PDL with new requirements (CFPB 2015).
CFA (2011) started to survey and report on Internet PDL lending in 2004. Its initial results found a proliferation of PDL marketing sites, anonymous domain registrations, and difficulty locating or identifying lenders. In 2011, the CFA surveyed a sample of 20 Internet PDL websites and reviewed all web pages and other information available. Their survey found that a growing number of online lenders had associations with Native American tribes and claimed to be exempt from state law enforcement due to tribal sovereign immunity. The typical cost of a $500 two-week loan from these lenders was $125, or 625% APR. Montezemolo (2013) also found that some PDL lenders claimed to be exempt from compliance with state consumer standards because of associations with Native American tribes. Other Internet lenders claimed that "choice of law" allowed them to comply only with the laws in the states in which they were headquartered (generally those with minimal or no PDL regulations; Montezemolo 2013).
DATA AND METHODS
Twitter, with over 400 million active users across the world, provides public access to a random sample of approximately 1 % of all tweets in real time through an advanced programming interface (API) (Murthy 2013). The metadata include text, time when the tweet was sent, user's language, number of favorites and followers, expanded URL if included in the tweet, author's location, along with the geocode of latitude and longitude if users choose to enable this feature. We searched for keyword of "PDL" to collect tweets related to this study. To assess the quality of our data, two human judges randomly reviewed a sample of 1,000 processed tweets. The results confirmed that more than 99% of tweets were related to PDLs and more than 99% of words were in English.
In this study, we used a combination of human judgment and machine learning algorithm to classify tweets as either commercial or noncommercial. Commercial tweets included tweets from lenders who used Twitter as a platform to promote and advertise their PDL products. These tweets could contain branded promotional information, and/or URLs linked to PDL websites. Noncommercial tweets were tweets from PDL users or the general public or consumer advocacy groups that reflected opinions or experience with PDLs and/or contained URLs to nonpromotional web pages. In the training stage, two human judges first classified a random sample of 1,000 tweets as commercial or noncommercial based on the tweet contents and web links. Examples of tweet classification were provided in the online appendix (Table A1). These judges had a high agreement in classification (92%) and jointly reclassified the undecided tweets. Then a Naive Bayes machine learning algorithm (Murphy 2012) was developed on the training tweets to create a predictive model for classification and the details of the model is provided in the Appendix. The Naive Bayes model achieved an accuracy of 97% on the training data (Table A2 in the online appendix). Finally, we applied the Naive Bayes model to a full set of collected tweets. A random sample of 1,000 tweets from full data set was appended with predicted classifier to serve as the test/validation data. A human judge fully reviewed these 1,000 tweets and compared the human classifier with predicted classifier. The accuracy on the test data was 89%.
Sentiment analysis, also called opinion mining, extracts and identifies from the relevant emotions or opinions within a document that concern a specific event or interest (Liu 2012). Sentiment score can be calculated by analyzing each word included in tweets. The lower the sentiment score is, the more negatively people express their feeling,
Sentiment score = [n.summation over (i=1)](positive-negative) x 100 (1)
where n represents number of words in the tweet.
We started with the opinion lexicon compiled by Liu (2012), which includes a list of 2,914 positive and 4,914 negative opinion words or sentiment words in English. The common opinion lexicon is insufficient to capture the sentiment in financial products. For instance, the words "grow," "growing," or "rising" are typically considered as positive when people express opinions. However, people often use these words to express their concern or frustration when talking about financial burden. In addition, "default" and "shark" are considered as neutral words in the general lexicon, but we had to reclassify them as negative in our study. To better capture the sentiment toward PDL lending, our two human judges further adjusted the opinion lexicon after reviewing the 1,000 randomly selected tweets. The word cloud of positive and negative words from PDL-related tweets using our customized lexicon were presented in the appendix (Figure A1a and b), where the font size is proportional to the frequency of words within each category.
We collected a total of 23,276 tweets from June 28 to October 2, 2015. These tweets were posted by 9,555 unique users and the average tweets per user over the study period was 2.44 (SD = 20.20). Table 1 presents the summary statistics of PDL-related tweets in total and separated by commercial and noncommercial tweets. Among all collected tweets, 16,583 (71 %) were classified as commercial tweets. These tweets had significantly more tweets per user than did noncommercial tweets (2.99 tweets per user for commercial, SD =20.60; 1.67 tweets per user for noncommercial, SD = 5.33; p < .0001). For the tweets related to commercial interests, the top ten users accounted for 25.48% of total tweets, compared with 11.54% for the noncommercial tweets.
In addition to the text information, Twitter API also provides other user-level metadata. Table 2 presents the statistics of other metadata in total and separated by commercial and noncommercial tweets. The average followers per user for commercial tweets were 3,521 versus 5,663 for noncommercial users. The potential reaches in Twitter were 19,515,960 for commercial tweets and 22,721,649 for noncommercial tweets, suggesting a significant presence of PDL information on Twitter given we collected only less than 1% of tweets in the study period. Of all commercial tweets, 86% included URLs, suggesting that the payday lenders do use Twitter as a platform to market and promote their products. On the other hand, 66% of noncommercial tweets also included URLs, indicating that consumer advocacy groups also use social media networks to disseminate information. The average number of tweets reposted (retweets) was 16.99 for commercial tweets and 2.63 for noncommercial tweets (p < .001). The average number of "favorites" (indications that the user liked the content of the tweet) per tweet was 1,370 for commercial tweets and 2,916 for noncommercial tweets, implying that noncommercial PDL-related tweets were much more likely to be considered favorites than commercial tweets.
Table 3 presents the summary of sentiment score. We found a significant difference in sentiment scores between commercial and noncommercial tweets. The commercial tweets had a high sentiment score of 34.9 (SD = 90.6), indicating that PDL lenders expressed a positive sentiment when marketing and promoting their PDL products or services on Twitter. In contrast, the noncommercial tweets had a much lower sentiment score of -50.7 (SD = 108.9, p < .0001 with commercial tweets), suggesting the concerns about PDLs from noncommercial users. The average sentiment score was 10.3 (SD = 103.7) but this finding is less meaningful given we have two distinct populations mixed in the data.
In addition, the key words contained in these two groups were very different. The words "cash" and "online" appeared in 16.51% and 12.52% of commercial tweets but in only 0.30% and 3.82% of noncommercial tweets. Commercial tweets often used the words "fast/instant/speedy/quick" (11.05%) or "cheap/free/easy" (8.68% of tweets) to allure Twitter users to apply for a PDL (compared with noncommercial tweets, only 0.25% and 0.37% tweets included such words). On the other hand, consumers or consumer advocates expressed their strong opinions regarding PDL lenders, using words such as "shark/loan shark" (12.63% of noncommercial tweets). They also expressed concerns about their own financial health with the words "debt/poor" (9.74%) or "scams/predatory/trap/fraud" (8.73%). About 4.29% of noncommercial tweets also included "CFPB," the acronym for the CFPB, indicating a potential push by consumer advocates for more regulations on PDL lending. The presence of such words in commercial tweets was very limited (0.22%).
PDLs are traditionally subject to state regulations that vary by state. One of the limitations of previous social media studies is that a majority of tweets were not geocoded and only less than 1 % of tweets had longitude and latitude data (Ram et al. 2015; Young, Rivers, and Lewis 2014). We imputed the geographic locations based on the city or state information provided in the tweet metadata (Broniatowski, Paul, and Dredze 2013; Daniulaityte et al. 2015; Ram et al. 2015), an approach that significantly increases the sample size when analyzing the social media data at the geographic level. We were able to identify 17.5% of tweets (4,083/23,276) that had a valid state or city location in the United States. Table A3 in the online appendix presents the number of PDL-related tweets in the United States with geo locations, separated by state legal status on PDL lending (P: permissible; R: restricted; H: hybrid) and type of tweets (commercial versus noncommercial).
Table 4 summarizes the key characteristics of metadata at the state group level (permissive, hybrid, and restrictive), separated by commercial and noncommercial purpose. The characteristics of commercial versus noncommercial tweets were different depending on the state's legal stance on PDL lending. The states with no restrictions on PDLs had the highest percentage of noncommercial tweets (42% vs. 37% for restricted states and 33% for hybrid states) and the lowest average commercial tweets per user (2.17 versus 3.87 for restricted states and 4.37 for hybrid states). On the other hand, the states that restricted PDLs had the highest number of followers per user (9,182 versus 4,481 for hybrid states and 3,821 for permissible states). These commercial tweets appeared 1,333,714 times on feeds despite laws that restrict physical stores from marketing PDL products.
The relationship between sentiment score and state regulations was further analyzed. The noncommercial tweets from states that restricted PDL lending had a significantly higher sentiment score of -30.2 compared to -54.5 for those from hybrid states (p = .0045) and -56.6 for those from permissible states (p = .0001), indicating the heterogeneity in the perception of PDLs by state regulations. On the other hand, the sentiment scores from commercial tweets were slightly lower in states that permitted PDL lending (22.0 vs. 26.6 for hybrid states and 24.8 for restricted states), but the differences were not significantly different, implying that commercial tweets across different states convey similar messages.
SUMMARY AND DISCUSSION
The PDL industry has seen explosive growth over the last two decades. As high-cost PDL products are often marketed heavily to financially vulnerable consumers who sometimes repeatedly use these loans, consumers can become trapped in debt (Center for Responsible Lending 2015). As a result, the CFPB has proposed stronger consumer protections on PDLs and other deposit advance products to end PDL debt traps (CFPB 2015).
Our study is the first to examine the prevalence of tweets related to PDLs on Twitter. We found that 71% of PDL tweets on Twitter were commercial tweets, which often included either promotional messages or URLs linked to Web sites that promote PDL use. These commercial tweets had a positive sentiment score of 34.9 and a significant number of them mentioned the words "cash," "online," "quick," "fast," "cheap," "free," and "easy." These findings suggest that PDL lenders proactively market their products on social media networks. On the other hand, 29% of tweets (6,693) were classified as noncommercial tweets, suggesting that consumers or consumer advocates do use social media sites to express their experience and opinions about PDL products. The noncommercial tweets had a sentiment score of -50.7 and over 35% of tweets included the words "shark," "loanshark," "debt," "trap," "scam," "fraud," "predatory," as well as the acronym for the "CFPB." The use of these terms suggests that many consumers or consumer advocates are concerned by PDL lending practices and are demanding changes in the industry.
Our descriptive study also sheds light on the relationship between state regulations, PDL commercial marketing and consumer sentiment toward PDLs. The noncommercial tweets from states that restricted PDL lending had a significantly higher sentiment score (-30.34) compared to those from states with hybrid regulation (-54.49) and those from states that permitted PDL (-56.62). In addition, state regulations affected how PDL lenders marketed their products on Twitter. PDL marketing was more prevalent in states that restricted PDLs than in other states with limited restrictions or no restrictions.
Our research confirms that social media data and "Big Data" analytics could provide valuable insights into how PDL products are marketed on social media platforms and how consumers or consumer advocates feel about these products. Used in this way, social media can serve as a tool to educate and prevent people from falling into the PDL debt trap.
Bertrand, M. and A. Morse. 2011. Information Disclosure, Cognitive Biases, and Payday Borrowing. Journal of Finance, 66: 1865-1893.
Bhutta, N. 2014. Payday Loans and Consumer Financial Health. Journal of Banking and Finance, 47: 230-242.
Bhutta, N., P. M. Skiba, and J. Tobacman. 2015. Payday Loan Choices and Consequences. Journal of Money, Credit, and Banking, 47: 223-259.
Broniatowski, D. A., M. J. Paul, and M. Dredze. 2013. National and Local Influenza Surveillance through Twitter: An Analysis of the 2012-2013 Influenza Epidemic. PLoS One, 8 (12): e83672. DOI: https://doi.org/10.1371/journal.pone.0083672.
Center for Responsible Lending. 2015. Fast Facts - Payday Loans, http://www.responsiblelending.org/payday-lending/tools-resources/fast-facts.html.
CFA. 2011. CFA Survey of Online Payday Loan Websites, http://www.consumerfed.org/pdfs/CFAsurveyInternetPaydayLoanWebsites.pdf.
CFPB. 2013. Payday Loans and Deposit Advance Products, a White Paper of Initial Data Findings. http://files.consumerfinance.gOv/f/201304_cfpb_payday-dap-whitepaper.pdf.
__. 2015. Factsheet: the CFPB Considers Proposal to End Payday Debt Traps. http://www.consumerfinance.gov/newsroom/cfpb- considers- proposal-to-end- payday- debt- traps/.
Dai, H. and J. Hao. 2017. Mining Social Media Data for Opinion Polarities about Electronic Cigarettes. Tobacco Control, 26 (2): 175-180. DOI: https://doi.Org/10.1136/tobaccocontrol-2015-052818.
Daniulaityte, R., R. W. Nahhas, S. Wijeratne, R. G. Carlson, F. R. Lamy, S. S. Martins, E. W. Boyer, G. A. Smith, and A. Sheth. 2015. "Time for Dabs": Analyzing Twitter data on Marijuana Concentrates across the U.S. Drug and Alcohol Dependence, 155: 307-311. DOI: https://doi.org/10.1016/j.drugalcdep.2015.07.1199.
Duggan, M., N. B. Ellison, C. Lampe, A. Lenhart, and M. Madden. 2015. Social Media Update 2014. Pew Research Center. http://www.pewinternet.org/files/2015/01/PI_SocialMediaUpdate20144.pdf.
FDIC. 2009. National Survey of Unbanked and Underbanked Households, https://www.fdic.gov/HOUSEHOLDSURVEY/2009/full_report.pdf.
__. 2014. 2013 FDIC National Survey of Unbanked and Underbanked Households. https://www.fdic.gov/householdsurvey/2013report.pdf.
Ginsberg, J., M. H. Mohebbi, R. S. Patel, L. Brammer, M. S. Smolinski, and L. Brilliant. 2009. Detecting Influenza Epidemics Using Search Engine Query Data. Nature, 457 (7232): 1012-1014. DOI: https://doi.org/10.1038/nature07634.
Li, M., K. J. Mumford, and J. L. Tobias. 2012. A Bayesian Analysis of Payday Loans and Their Regulation. Journal of Econometries, 17: 205-216.
Liu, B. 2012. Sentiment Analysis and Opinion Mining. Synthesis Lectures on Human Language Technologies, 5(1): 1-167.
Montezemolo, S. 2013. Payday Lending Abuses and Predatory Practices. The State of Lending in America & Its Impact on U.S. Households, Center for Responsible Lending, https://ssrn.com/abstract=2391403.
Murphy, Kevin P. 2012. Machine Learning: A Probabilistic Perspective. Cambridge, MA: MIT Press. DOI: https://doi.org/0262018020.
Murthy, D. 2013. Twitter: Social Communication in the Twitter Age, Digital Media and Society. Cambridge: Polity.
National People's Action. 2012. Profiting from Poverty: How Payday Lenders Strip Wealth from the Working - Poor for Record Profits, http://npa-us.org/files/images/profiting_from_poverty_npa_payday_loan_report _jan_2012_0.pdf.
Pew Charitable Trusts. 2014. State Payday Loan Regulation and Usage Rates, http://www.pewtru.sts.org/en/multimedia/data- visualizations/2014/state-payday-loan-regulation-and-usage-rates.
Ram, S., W. Zhang, M. Williams, and Y. Pengetnze. 2015. Predicting Asthma-Related Emergency Department Visits Using Big Data. IEEE Journal of Biomedical and Health Informatics, 19 (4): 1216-1223. DOI: https://doi.org/10.1109/JBHI.2015.2404829.
Salathe, M. and S. Khandelwal. 2011. Assessing Vaccination Sentiments with Online Social Media: Implications for Infectious Disease Dynamics and Control. PLoS Computational Biology. 7 (10): el002199. DOI: https://doi.org/10.1371/journal.pcbi.1002199.
Statista. 2015. Number of Monthly Active Facebook Users Worldwide as of 1st Quarter 2015. hup://www.statista.com/statistics/264810/number-of-monthly-active-facebook-users-worldwide/.
Stephens Inc. 2014. Alternative Financial Services: Innovate to Meet Customer Needs in an Evolving Regulatory Framework. http://cfsaa.eom/Portals/0/cfsa2014_conference/Presentations/CFSA2014_THURSDAY_GeneralSession_JohnHecht_Stephens.pdf.
Wong, C. A., M. Sap, A. Schwartz, R. Town, T. Baker, L. Ungar, and R. M. Merchant. 2015. Twitter Sentiment Predicts Affordable Care Act Marketplace Enrollment. Journal of Medical Internet Research, 17 (2): e51. DOI: https://doi.org/10.2196/jmir.3812.
Young, S. D., C. Rivers, and B. Lewis. 2014. Methods of Using Real-Time Social Media Technologies for Detection and Remote Monitoring of HIV Outcomes. Preventive Medicine, 63: 112-115. DOI: https://doi.org/10.1016/j.ypmed.2014.01.024.
JIANQIANG HAO AND HONGYING DAI [iD]
Jianqiang Hao (email@example.com) is a Vice President of Customer Data Management at First National Bank of Omaha and Hongying Dai (firstname.lastname@example.org) is an Associate Professor of Health Services and Outcomes Research at Children's Mercy Hospital and Department of Biomedical & Health Informatics and Department of Pediatrics at University of Missouri-Kansas City. There are no competing interests to this work. We thank editors for their constructive comments, which helped us improve the manuscript. We also thank the Medical Writing Center at Children's Mercy Hospital for proofreading the manuscript.
The Journal of Consumer Affairs, Summer 2018: 441-451
Copyright 2017 by The American Council on Consumer Interests
TABLE 1 Summary Statistics of PDL-Related Tweets Tweets Type Number of Number of Average Number of % of Tweets from Tweets Users Tweets Per User Top Ten Users Commercial 16,583 5,543 2.99 (SD 22.60) 25.48% Noncommercial 6,693 4,012 1.67 (SD 5.33) 11.64% Total 23,276 9,555 2.44 (SD 20.20) 21.50% p value of p < .0001 difference between commercial and noncommercial tweets TABLE 2 Metadata Characteristics of Commercial and Noncommercial Users Tweets Type Average Potential Includes Average Number of Number of Reach in URL Number of Favorites Followers Twitter (a) Retweets Per Tweet Per User Commercial 3,521 19,515,960 86% 16.99 1,177 Noncommercial 5,663 22,721,649 66% 2.63 3,395 Total 4,420 42,237,609 81% 12.86 1,815 (a)Potential reaches in Twitter were calculated by summing the total number of followers for each tweet. Total times posted represent the number of times tweets appeared on feeds. TABLE 3 Summary of Sentiment Score Tweets Type Sentiment Score p-Value of Sentiment Scores Between Commercial and Noncommercial Tweets Commercial 34.9 (SD 90.6) p < .0001 Noncommercial -50.7 (SD 108.9) Total 10.3 (SD 103.7) TABLE 4 Characteristics of Metadata at the State Group Level State Regulation Number of Number of Number of Average on PDL Tweets Tweets Users Tweets Per User Hybrid 922 100% 320 2.88 Commercial 621 67% 142 4.37 Noncommercial 301 33% 178 1.69 Permissible 1,741 100% 944 1.84 Commercial 1,011 58% 465 2.17 Noncommercial 730 42% 479 1.52 Restricted 1,420 100% 521 2.73 Commercial 897 63% 232 3.87 Noncommercial 523 37% 289 1.81 State Regulation Potential Average Number of on PDL Reach in Number of Favorites Twitter Followers Per Tweet Per User Hybrid 1,433,967 4,481 1,055 Commercial 411,459 2,898 667 Noncommercial 1,022,508 5,744 1,854 Permissible 3,607,326 3,821 2,012 Commercial 1,403,789 3,019 1,017 Noncommercial 2,203,537 4,600 3,390 Restricted 4,783,612 9,182 906 Commercial 1,332,714 5,744 503 Noncommercial 3,450,898 11,941 1,596
|Printer friendly Cite/link Email Feedback|
|Title Annotation:||TRENDS AND APPLICATIONS|
|Author:||Hao, Jianqiang; Dai, Hongying|
|Publication:||Journal of Consumer Affairs|
|Date:||Jun 22, 2018|
|Previous Article:||A Systematic Literature Review to Identify Successful Elements for Financial Education and Counseling in Groups.|
|Next Article:||A Whole or Two Halves: Serving Size Framing Effects and Consumer Healthfulness Perceptions.|