The Early Days of Neobanks in Europe: Identification, Performance, and Riskiness

The literature offers a variety of definitions of ‘neobanks’.³ For instance:

-“Neobanks make extensive use of technology in order to offer retail banking services predominantly through a smartphone app and internet-based platform.” (...) “They leverage scalable infrastructure through cloud providers or API-based systems to better interact through online, mobile and social media-based platforms.” (...) “[They] may adopt big data technologies and advanced data analytics” (BCBS 2018; p. 16);

-“[Digital-only banks] are recently established (...) [and] innovative banks that use primarily digital channels (e.g., online, mobile apps, etc.) to serve their existing and new clients. These banks do not have branches nor maintain a network of private bankers, while relying on new technologies for managing interactions with their customers.” (p.42) (...) “They tend to specialise in certain primary business lines, such as payment systems, trading and asset management”. (ECB 2020: p.45);

-“Digital-only banks are branchless banks, meaning that their customers can only transact with them using digital banking channels such as online banking and mobile banking” (Nel and Boshoff 2021: pp. 429–430);

-“Neobanks (...) are offering customer-friendly interfaces and employing more efficient IT processes.” (Boot et al. 2021: p.14).

In general, such citations suggest that, while there is no unique definition in the literature, ‘neobanks’ may be seen as a ‘new’ type of banking business model, that uses innovative digital technologies – including blockchain, smart contracts, robo-advisors, advanced analytics and big data (Arner et al. 2017; Frost 2020; Carbó-Valverde et al. 2021; Oehler et al. 2021) –, to build competitive advantage over traditional peers in specific business lines (e.g., lending, payments, trading and asset management), via the provision of a superior customer experience and/or lower costs.

More precisely, in this paper we define ‘neobanks’ as licensed banking institutions, born during the post internet-era (i.e., after 2000), focused on retail banking activities (as shown by their involvement in lending and deposit-taking activities), and with a strong orientation towards digital distribution channels, rather than physical branches (Ehrentraud et al. 2020; Boot et al. 2021). Our focus on banks with a strong retail orientation allows us to set a level-playing field, in terms of regulatory and supervisory treatment, between the digital banks and traditional peers in our sample, as opposed to using a broader definition of digital banks, closer to ‘fintech banks’ (Gelis 2016). These would include, for instance, institutions that are mostly focused on offering services to other banks, despite have a banking license, e.g., ‘Banking as a Service’ (BaaS), offered by Solarisbank (Germany), or Railsbank (UK).

Interestingly, the management literature has often placed the onset of academic interest for the term ‘business model’ in the context of the dot-com boom of the late 1990’s (Zott et al. 2011). Thus, in a sense, describing ‘neobanks’ as a ‘business model’ may be seen as a full-circle for business model literature. Moreover, a set of advances has been made regarding the identification of banking business models, particularly regarding the methods used. For instance, Ayadi et al. (2011) apply hard clustering techniques to the financial data (e.g., customer deposits, trading assets, loans to banks) of 26 banks and group them into three banking business models: retail, investment, and wholesale. Using a larger sample of European banks, Mergaerts and Vander Vennet (2016) apply factor analysis to financial data and find two main factors: retail and diversification. More recently, Marques and Alves (2020, 2021) combine both approaches by using the retained principal components as inputs to an ensemble of three alternative clustering techniques (fuzzy c-means, self-organizing maps, partitioning around medoids) yielding a total of four business models: retail focused, retail diversified funding, retail diversified assets, and large diversified. While such developments may be noteworthy, it is also striking that none of the cited works use non-financial data to identify business models (e.g., business lines or customer experience offered). We argue that, while such shortcomings may be less severe when mapping a sample of traditional banks, it becomes more striking when the goal is to identify the ‘neobanks’ business model, which, as described above, is focused on specific segments and on offering an innovative digital customer experience. However, gaining access to non-financial data can be very costly, especially for large samples.

As such, it is not surprising that the identification of ‘neobanks’ has seldomly been attempted by the literature. For instance, Delgado et al. (2007) study the performance of 15 ‘primarily internet banks’ in Europe between 1994 and 2002, which are characterised as being “heavily reliant on the Internet as their most important delivery channel” (p.650). Similarly, DeYoung (2005) identifies 12 internet banks operating in the US, applying a set of criteria related to age, size and range of banking products offered through the Internet. More recently, the ‘Cambridge Centre for Alternative Finance’ (CCAF) set up the ‘Cambridge FinTech Ecosystem Atlas’ with the goal to “systematically identify, classify and visualize FinTech entities” (CCAF 2022). With reference to the end of 2021, the database covered 2915 entities, from 108 countries, classified in 14 market segments, 63 sub-segments and 118 categories. Such classification was done in a collaborative way by academics and industry participants, using data from a variety of sources such as company websites, company statements, interviews, and news articles. One of the classifications presented by CCAF is the ‘fully digitally native banks’, which seems to fit our definition of neobanks. However, when checking the entities operating in EU-28 countries, only 23 entities are currently classified as ‘fully digitally native bank’, in 7 countries. Perhaps more importantly, we find evidence that not all these entities hold a banking license.⁴ In our view, while the collaborative and evolving nature of the project is bound to eliminate such imprecisions, this contributes to the perception that a gap effectively exists in the banking literature regarding the identification of neobanks.

The performance of ‘neobanks’ may be understood using two conceptual frameworks from the management and banking literature: bank intermediation theory, and strategic management theory.

A key theoretical framework for our study is bank intermediation theory. In general, when performing its role in the efficient allocation of savings in investment opportunities (Merton 1995),⁵ banks address two types of frictions: (i) information frictions, which are related to moral hazard and adverse selection, and are mitigated via the screening and monitoring of risky investments on behalf of savers (Diamond 1984); and (ii) communication frictions, which are linked to search, switching and transportation costs, and historically have been overcome by setting up physical branches which enabled customer relationships to arise (Boot et al. 2021).

Regarding the information frictions, a theoretical result that seems particularly timely for our empirical context is the notion of ‘winner’s curse’. Namely, this phenomenon relates to effects, for de novo banks, that emerge from the possibility that loan applicants may indefinitely apply for loans, after being previously rejected. The model by Broecker (1990) suggests that this feature of the loan market makes new entrants prone to have a riskier loan portfolio than those of incumbents unless their screening abilities⁶ are, in fact, superior – which is corroborated by the empirical findings from Shaffer (1998). An alternative explanation is put forward by the literature on ‘relationship-transactional lending’, wherein neobanks are likely to adopt a transactional banking model. Under this framework, on one hand, neobanks are likely to offer arms-length contractual relationship, with less access to soft information on borrower creditworthiness, and hence greater exposure to information asymmetries (Herpfer 2021) and lower ability to price discriminate (Degryse and Ongena 2005); on the other hand, the relatively simpler hierarchical structure of neobanks, coupled with superior IT capabilities, allows neobanks to shorten the decision process, resulting in a more timely response to customer requests, for which they are willing to pay an interest rate premium (Tanda and Schena 2019; Williams 2021; Berg et al. 2022). In order for the latter narrative to hold true in our empirical setting, we would expect neobanks to record higher interest income (net of impairment charges), while showing a similar level of riskiness; on the contrary, the ‘winner’s curse’ hypothesis would be consistent with a higher level of risk recorded by the new entrants (neobanks).

As for communication frictions, theoretical and empirical works have historically used the term “spatial capture” to depict the relationship banks’ ability to profit from their physical proximity to customers via cross-selling (Puri and Rocholl 2008) and access to cheap and stable funding (Drechsler et al. 2021). However, according to Boot et al. (2021) the current wave of digitalisation may have shifted the “spatial capture” of customers from the physical to the digital domain. Particularly, according to Boot et al. (2021) such ‘digital spatial capture’ could result from the neobanks’ ability to “set up efficient communication channels via web portals and mobile apps at very low cost (…), reach targeted audiences via direct marketing tools, including social media (…), source flexible and cost-effective IT infrastructure through cloud services (…), [and] facilitate payments for online purchases” (p.6). As such, whether neobanks have been able to ‘digitally capture’ bank customers, remains an open empirical question. In our analysis, such ‘digital capture’ would be backed by findings consistent with the ability of neobanks to charge higher fees and pay lower interest expenses for customer funding.

Another framework that may help us explain the performance of neobanks is provided by strategic management literature – namely, strategic groups theory (SGT) (Caves and Porter 1977) and resource-based theory (RBT) (Wernerfelt 1984). According to SGT, the firms of a given market are likely to make decisions regarding a set of strategic dimensions, such as the distribution channel, the type of products offered, or the level of value chain integration, resulting in the creation of groups of firms that exhibit similarity of strategic choices within each group, and dissimilarity from other groups. Such strategic choices are often related to costly investments which (i) could impede firms from easily changing their group membership in the short run, i.e., the so-called ‘barriers to mobility’; and (ii) may protect incumbent firms from new entrants (McGee 2006). As such, one of the key propositions of SGT is that mobility barriers may play an important role in explaining intra-industry performance heterogeneity (Porter 1979). In the same vein, according to RBT an incumbent firm’s performance strongly depends on the uniqueness of its resources and capabilities (Wernerfelt 1984), i.e., whether they are rare, difficult to imitate, and difficult to transfer between firms, at least in the short-run (Barney 1991). In banking, some examples of ‘unique resources and capabilities’ that are difficult to transfer between banks include risk management processes and customer knowledge – in line with the ‘soft information’ concept, previously discussed. In general, such theories provide us with mixed predictions regarding the performance of neobanks: on one hand, the branch network and longstanding customer relationships of traditional peers may be seen as an entry barrier (or an unique resource) impeding neobanks from tapping into certain potentially lucrative markets (e.g., SME lending); on the other hand, the ability to take full advantage of innovative technologies may be seen a resource uniquely available to neobanks, or a difficult barrier for traditional peers to overcome, particularly given that many of these banks face IT legacy issues (Stulz 2019).

The emergence of internet banks in the late 1990s and early 2000s ignited a strand of literature studying their characteristics and especially focused on analysing their performance relative to traditional peers (Claessens et al. 2002). The seminal work in this subject was developed by DeYoung (2005), wherein the performance of 12 internet-only US banks, incorporated between 1997 and 2001, is compared to that of 644 branching banks set up in the same period. According to the author, the focus on banks born as internet-only banks allows “a clean test of the internet-only business model (…) unaffected by any production structure or client relationships left over from a preexisting business model” (DeYoung 2005: p.894). The results show that, on average, internet-only banks underperform relative to the traditional peers, due to overhead inefficiencies that more than offset the better pricing abilities. According to the author, such efficiencies may be linked to the fact that “internet-only banks have access to deeper scale economies than branching banks” (DeYoung 2005: p.895) – an indication that internet-only banks could become more profitable than their traditional peers, after reaching a certain size.

Bearing this aspect in mind, Delgado et al. (2007) extend DeYoung’s (2005) framework to a sample of 15 ‘primarily internet banks’ operating in 6 European countries, between 1994 and 2002. The authors find that “internet banks performance lies below both the newly chartered traditional banks and the small established traditional banks for all size categories, but this performance gap diminishes for larger Internet banks” (Delgado et al. 2007: p.654). Moreover, the results regarding the drivers for the underperformance of US internet banks are confirmed for European banks. Namely, it is found that such underperformance is mainly related to the overhead costs. Importantly, however, the authors find evidence that the rate at which overhead costs reduce with size is greater for internet banks than for traditional peers, and that the net margin increases with size only for internet banks. Respectively, such results are interpreted as consistent with the existence of ‘economies of scale’ in internet banking.

Strikingly, however, the analysis of the cost of risk of internet banks has fallen outside the scope of existing studies on internet bank performance. As such, next we draw on the literature regarding the performance of fintech lenders, where emphasis has been placed on their screening abilities.

Overall, the literature on the performance of fintech lenders provides mixed results regarding the merits and shortcomings of fintech lenders, depending on the type of business line. On one hand, using a large sample of mortgage loans in the US between 2007 and 2015, Buchak et al. (2018) find that default rates are statistically similar between fintech lenders and traditional banks. Moreover, fintech lenders tend to charge higher spreads than traditional banks for comparable borrowers, which is interpreted as a premium that borrowers are willing to pay for the superior customer experience offered by fintech lenders. In the same vein, using a similar database to Buchak et al. (2018), Fuster et al. (2019) also document similar default rates between loans extended by fintech and non-fintech lenders. Also, the evidence collected points towards a significantly faster pace at which mortgage applications are processed in fintech lenders, which seems to provide additional reasoning for the ‘better customer experience-higher rates’ nexus.

On the other hand, Di Maggio and Yao (2021) focus on a large sample of US personal loans and show that fintech lenders enter the market by lending to higher-risk borrowers. Also, the evidence suggests that borrowers with similar characteristics are more likely to default when borrowing from fintech lenders than from traditional banks. Despite this, the results suggest a high correlation between interest rates and default rates, which is interpreted as evidence of fintech lenders’ pricing ability. Similarly, Balyuk et al. (2020) analyse a large sample of US small business loans and find evidence suggesting that fintech lending tends to replace the riskier loans granted by large and out-of-market banks. Finally, Carmichael (2017) find that fintech borrowers in the US personal loans market who were previously rejected by another fintech competitor are twice as likely to default as borrowers who were not rejected, which is seen as evidence of the ‘winners curse’ (Shaffer 1998) in the online lending market.

One feature of this strand of literature is the lack of research studies regarding European fintech lenders. The few exceptions that provide a picture of European fintech lenders (e.g., Milne and Parboteeah 2016; Claessens et al. 2018) are focused mostly on country-level data and do not offer a comparative performance of fintech and non-fintech lenders. We argue that this phenomenon is presumably linked to the relative lack of microdata in Europe vis-à-vis the US, where some fintech lenders provide open access to their data. For instance, the US fintech ‘Prosper’ provides access to monthly loan-level data, regarding the characteristics and performance of loans. In this context, while the use of bank-level data is not ideal, we argue that it remains a relevant contribution to the scant literature on emerging digital business models in Europe.

Our method to identify neobanks consists of four sequential filters applied to the data, wherein each step uses different sources of data and is supported by the literature. As presented in Fig. 1, in the first filter, our goal is to find banks that share a similar business profile in terms of geography, size, business model and ownership type. Regarding geography, we focus on supervised banks operating in EU-28 countries. This is done by retrieving the list of supervised banks from each national supervisory authority, including the ECB for the Euro Area, with reference to 2019. With respect to size, we expect neobanks to be newcomers and, as such, to exhibit a relatively small size. Hence, we only consider banks with total assets smaller than EUR 10 bn. The ecosystem of neobanks includes a variety of business lines, including retail lending, payments, trading, asset management, and B2B (ECB 2020). To improve the comparability of our sample, we narrow our focus on retail-oriented neobanks by requiring banks in our sample to have at least 5% of total assets dedicated to customer lending and customer deposits. Finally, we exclude stakeholder banks (i.e., cooperative and savings) from our sample, given that these are ex-ante more likely to operate a ‘brick-and-mortar’ model due to their strong presence in rural areas, where customers are typically less digitally mature.⁷

In the second filter, our aim is to identify banks with a high propensity to adopt a digital banking model, typical of neobanks. This is achieved by analysing the distribution channel and the age of banks. Regarding the distribution channel, as reported by Ehrentraud et al. (2020: p.8), digital banks may be seen as “delivering banking services primarily through electronic channels instead of physical branches”. Hence, we consider that banks with more than 5 branches⁸ have a low propensity to adopt a digital banking model, and as such exclude them from the neobanks sample. With respect to the age of the bank, we expect that customer preferences towards digital banking services as well as the technology necessary to offer such services, became more intense after the internet diffusion reached a critical mass. According to Boot et al. (2021) such tipping point occurred during the early 2000s. As such, we consider that banks that were incorporated before the year 2000 are less likely to adopt a digital banking model, and hence exclude them from this sample.⁹

The third filter identifies banks that stakeholders perceive as being digital. This is done via the systematic analysis of news in Factiva.¹⁰ Importantly, this procedure allows us to complement the analysis of hard data (filter #2) with information on the perception (of peers, industry associations, media) regarding which banks may be considered as “neobanks” (FT 2019, 2020). Theoretically, such type of analysis bears support from the cognitive perspective of strategic groups (Reger and Huff 1993: p.103), according to which “industry participants share perceptions about strategic commonalities among firms” which are likely to influence decision making. To undergo the systematic analysis of news in Factiva, we run individual searches for each bank with a high propensity to operate a digital banking model, combining the name of the bank with twelve alternative keywords related to digital banking (vide list of keywords in Fig. 1). Also, when analysing the news reports, in some cases we identify new candidate digital banks (i.e., banks not included in the initial list of banks with high propensity to operate a digital banking model). For such banks, we run a reverse analysis, wherein we go back to filter #2 and re-check the reason for exclusion. Such analysis allows us to correct several issues in the BankFocus database (e.g., incorrect year of incorporation, specialisation, or number of branches).

The fourth and final filter is related to the availability of certain online functionalities that may be seen as typical of a neobank. We view this step as complementing the previous two (hard data and stakeholder perception). In other words, even if the number of branches of a bank is very low, and/or stakeholders identify the bank as a neobank, one could argue that neither of these indicators constitute direct evidence of the ability of the bank’s IT systems to offer “customer-friendly interfaces and employ efficient IT processes” that enable it to reduce communication frictions – an important feature of digital banking, as suggested by Boot et al. (2021: p. 6). In line with this, Buchak et al. (2018: pp.458–459) classify as fintech lenders those that show “a strong online presence and if nearly all of the mortgage application process takes place online with no human involvement from the lender”. Our approach in this regard is to check the website of each bank for two specific features: (i) the ability to open an account fully online and (ii) the ability to apply for a loan online. Given that some banks in our sample have only a small orientation towards credit granting, we impose as sufficient condition for a bank to be considered as digital that it offers at least one of these online functionalities.

The estimation of the impact of adopting a digital business model on the performance of a bank is bound to be affected by endogeneity issues (Clougherty et al. 2016; Shipman et al. 2017). Namely, the difficulty is to find a suitable counterfactual that allows us to estimate the effect of being born with a fully digital business model (i.e., being a neobank) vis-à-vis operating a traditional banking model. First, given that we cannot directly check what the performance of the neobank would be had it not been born as a digital bank. Secondly, because it is likely that bank-specific features may simultaneously determine the likelihood of a bank being born as digital and its potential performance.

To mitigate these issues, we conduct ‘Propensity Score Matching’ (PSM) on the full sample of banks, wherein the PSM score is computed using the logit model and the matching is performed using the Epanechnikov-Kernel matching function. The propensity score (p) is defined as “the conditional probability of assignment to a particular treatment given a vector of observed covariates” (Rosenbaum and Rubin 1983: p.41) for each bank i:wherein Ω_i is a dummy that takes the value 1 if the bank is classified as ‘neobank’ and 0 otherwise; and X_i is the set of matching variables which we expect to cumulatively bear some impact on the treatment (neobank dummy) and outcome variables (performance and riskiness), namely: size, gross loans to customers, customer deposits, income diversification, liquid assets, and total equity. Moreover, we force exact matching on bank specialisation and a dummy which takes on the value 1 if the bank cumulatively meets the criteria referred in filter #1 of the neobank identification process (Fig. 1), i.e., total assets <10 bn, gross loans to customers / total assets >5%, customer deposits / total assets >5%, and non-cooperative or savings bank.

Alternatively, one may consider the possibility of matching neobanks with traditional peers also on the number of branches, age, and digital orientation. However, when doing so, the number of matched observations is heavily reduced, thereby raising model overfitting issues; perhaps more importantly, the matched counterfactuals become populated by banks with very specialised business models (e.g., corporate banks, consumer credit banks, subsidiaries of foreign owned banks), hence failing to ensure the external validity of our sample. The latter issue is discussed by Shipman et al. (2017: p.216): “in settings with limited overlap, PSM systematically excludes observations that lack counterfactuals, compromising the degree to which [results] can be generalized outside of the sample”. Effectively, in our empirical setting these critical features (branches, age, online functionalities) exhibit a limited overlap between the treated (neobanks) and the pool of potential controls (traditional peers), making it challenging to find similar banks on all relevant covariates without sacrificing external validity. In this context, following the suggestion made by Shipman et al. (2017), we privilege external validity while also ensuring that “alternative specifications yield similar inferences” (p.216) – namely, we iteratively include age and branches as additional matching variables in the PSM, and remove control banks with age or branches beyond specific data-driven thresholds (see Sections 6.2 and 6.3).

Finally, to assess the quality of our matching process, we test whether the mean differences between the treated and controls are significantly reduced in the post- versus pre-matching samples, hence providing evidence on the suitability of the counterfactuals (Rosenbaum and Rubin 1983).

We assess the performance of neobanks vis-à-vis traditional peers by estimating the following model on the matched sample, using cross-section OLS with White (1980) cluster-robust standard errors:wherein Y_i is the outcome variable, including the Return on Assets (ROA) and the sub-elements of ROA (interest income, interest expenses, net interest income, non-interest income, staff expenses, other non-staff expenses, impairment charges); α₀ is the model constant; Neobank_i is a dummy which takes on the value 1 if bank i is defined as a “neobank” according to our methodology¹¹; BC_i is the mean vector of individual bank control variables of bank i (size, income diversification, gross loans to customers, liquid assets, customer deposits, total equity, non-performing loans, number of products); CL_i is a vector of country-level controls (GDP growth, GDP per capita, concentration ratio); β, γ and δ are the regression coefficients’ vectors; and ε_i is the disturbance term. We also perform several alternative specifications concerning the choice of outcome variables (Return on Equity - ROE) and country controls (country fixed effects).

Given the relatively small size of our matched sample (n = 376), concerns may be raised regarding potential model overfitting and, hence, the validity of statistical inference (Harrel 2001). In this regard, the standard rule of thumb is that a “fitted regression model is likely to be reliable when the number of predictors (…) is less than n/10 or n/20” (Harrel 2001: p.61). This means that the n/p ratio in our baseline regressions (376/13 = 28.9) is above the minimum threshold set in the literature (10).

Our financial data corresponds to yearly financial statements information, at the unconsolidated level, obtained via Orbis BankFocus. We retrieve data for 2019 and 2020. Regarding non-financial data, we collect the information from business lines and online functionalities by checking each bank’s website in October 2021. The Factiva news search for media references equating each bank as a ‘neobank’ is also done during this period. To narrow the period mismatch and mitigate the difference in data frequency between the financial and non-financial data, we perform two data treatments. First, following the approach by Buchak et al. (2018) we assess the historical accuracy of the website analysis using the ‘Wayback Machine’, which provides access to archived webpages from 2019 and 2020; this allows us to confirm our initial results. Second, we compute the average values of financial data across 2019 and 2020, resulting in a 2019–2020 cross-section database. The country-level data is obtained from the World Bank database for 2019. We winsorise bank-level data at the 1st and 99th percentile.

We follow the literature on bank performance and include bank-level controls related to the size, diversification, asset and liability structures, leverage, and risk. Below we provide the reasoning for our choices of proxies.¹²

According to the literature, country level factors may influence the performance of banks in a variety of ways, and as such we control for their effects.

Table 1 shows the step-by-step results of our identification process. We start with a total of 3342 supervised banks in EU-28 countries as of 2019, with available financial data. In the first step we apply filters on bank’s size (total assets), level of retail-orientation (gross loans to customer on total assets, customer deposits on total assets) and specialisation in order built a sample of banks with similar business profile and ownership type. These criteria lead to a sample of 553 banks – corresponding to a reduction of approximately 80% of the initial sample – mainly due to the exclusion of cooperative and saving banks, which contribute to a reduction of 70% of the starting list of banks. Second, we further restrict the sample excluding banks with more than 5 branches or incorporated before the year 2000. Step 2 then allows us to define a group of potential neobanks comprised of 135 financial institutions. The group of 68 banks identified in the third step reflects both the direct and the reverse analysis performed via Factiva. Namely, in the direct analysis of the market perception we find 43 neobanks; whereas the reverse analysis yields an additional 25 neobanks. Finally, to further confirm the digital footprint of these banks, we assess the bank’s online functionalities by excluding banks that do not offer the possibility to open an account or apply for a loan via their website. In the end we obtain a total of 65 neobanks operating in Europe.¹⁴

Regarding the identification of controls, the implementation of the PSM produces a sample of 313 traditional peers, significantly reducing the original pool of potential controls (n = 3277). On the other hand, the matching process leads us to drop 2 unmatched neobanks, resulting in 63 neobanks in the post-PSM sample. Moreover, the results for the quality of matching, presented in Table 2, show that the PSM improves the overall quality of controls. Namely, the comparison of mean values of the matching variables of neobanks and traditional peers, pre- and post-matching, shows that PSM reduces the mean differences in all variables, almost all of which becoming statistically insignificant, except one (income diversification). In general, we take the results of the PSM procedure as providing an important source of confidence regarding the similarity of the neobanks and traditional peers with respect to key determinants of bank performance and riskiness.

In Table 3 we provide descriptive statistics for the key features of neobanks and traditional peers. As expected, on average, neobanks exhibit almost no branches (0.7 branches), are relatively young (10.8 years), and they all provide online functionalities. Moreover, neobanks have a relatively small balance sheet (EUR 2.5 billion of total assets) and are mostly oriented towards lending (54.0% of total assets) and deposit-taking (71.9%). Coherently, they show a low product diversification (4.5 products offered). When comparing neobanks to traditional peers, as expected we do not find significant differences in most of the key features – vide column (1). An important exception is related to the three variables used in the neobanks’ identification process: branches, age, and online functionalities. In particular, traditional peers record a significantly larger branch network (18.2 branches) and are significantly older (47.8 years) than neobanks. Additionally, only few traditional peers offer online functionalities (9.3%). Importantly, we do not find any statistically significant difference in the non-performing loans of neobanks vis-à-vis traditional peers – which is consistent with the findings provided by Buchak et al. (2018) and Fuster et al. (2019), according to which default rates for US borrowers are statistically similar between fintech and traditional lenders.

Table 4 shows the mean values of neobanks and traditional peers for Return on Assets (ROA) and the sub-components of ROA. In general, the results suggest that neobanks perform worse than their traditional peers. This is shown by the negative average ROA recorded by neobanks (−0.125%), which is significantly lower than the ROA registered by traditional banks (0.396%). This result is backed by the regressions results reported in Panel A of Table 5, particularly by the negative and significant coefficient of the ‘neobank dummy’ on ROA, in column (1).

To further shed light on the reasons that determine such underperformance, we analyse the relative performance of neobanks in each of the ROA sub-components – in Table 4 and Table 5, Panel A, columns (2–8). First, we find that neobanks record significantly higher interest income than their traditional peers (shown by the positive and significant coefficient of interest income in Table 5, Panel A, as well as the higher mean value in Table 4), while also recording a significantly higher cost of risk (as per mean value and coefficient of impairment charges). Additionally, we find that such ‘high interest-high impairment charges’ nexus is driven solely by the online functionalities dummy (Table 5, Panel B). As discussed in the literature review, we envision two alternative explanations for this phenomena: the ‘winners curse’ narrative (Broecker 1990; Shaffer 1998; Di Maggio and Yao 2021), according to which neobanks may be tapping into a pool of risky borrowers previously rejected by traditional peers; or, alternatively, the ‘transactional model’ narrative, wherein neobanks may be charging higher interests and recording more loan provisions as a reflection of the information asymmetries that arise from the adoption of a transactional banking model, which impairs the exploitation of soft information on borrower creditworthiness (Herpfer 2021). In our view, the fact that we control for the riskiness of the loan portfolios in the regressions, coupled with the similar levels of non-performing loans in neobanks and traditional peers, suggests that the ‘transactional model’ narrative is a more suitable explanation for the ‘higher interests–higher impairments’ recorded by neobanks.

Relatedly, such ‘high interest-high impairments’ approach to lending may generate concerns regarding the sustainability of the neobanks’ business model. To shed light on this aspect, we build a new dependent variable, ‘net interest income minus impairment charges’ and still find a positive and significant coefficient of the neobanks dummy at the 5% significance level (untabulated) – suggesting that neobanks seem to charge sufficient interests to cover the higher cost of risk, thereby alleviating the concerns on the potential lack of sustainability. Importantly, we note that these results seem to contradict the standard literature on lending technology (i.e., relationship vs transactional lending) (Boot 2000; DeYoung 2010; van Ewijk and Arnold 2014), which typically relates transactional lending to lower net interest margins, due to the lack of product differentiation. However, an emerging strand of the literature suggests that in the digital era customers place significant value on the quality of the customer experience (e.g., accessibility, usability, flexibility) and are willing to pay an interest rate premium, especially for faster loan decisions (Buchak et al. 2018; Tanda and Schena 2019; Di Maggio and Yao 2021; Williams 2021; Berg et al. 2022). For instance, Di Maggio and Yao (2021) study a large sample of US personal loans and show that personal loans from fintech firms are approximately 3 p.p. more expensive than traditional peers. Also, Buchak et al. (2018) find that fintech lenders tend to charge higher spreads than traditional banks for comparable borrowers and interpret this finding as evidence that borrowers are willing to pay for the superior customer experience offered by fintech lenders. In general, our results seem to align with the predictions put forward by this emerging strand of the literature.

With respect to the mitigation of ‘communication frictions’, the findings reported in Tables 4 and 5, indicate that, overall, neobanks exhibit a similar funding cost and ability to generate fees to that of traditional peers – hence, falling to provide support for the ‘digital capture’ hypothesis (Boot et al. 2021). Furthermore, the results show that neobanks record significantly higher non-staff expenses (e.g., IT costs, advertising costs, reporting costs) than traditional banks (as shown by the negative and significant coefficient of the neobank dummy in Table 5, column (7), as well as the statistically significant mean difference in Table 4). This result seems in line with the empirical literature which relates the relative underperformance of internet banks vis-à-vis traditional peers to cost inefficiencies (DeYoung 2005; Delgado et al. 2007). Note, however, that this strand also argues that digital banks seem to show a greater ‘depth’ of economies of scale than traditional peers – as will be discussed in the next sub-section.

Finally, in Panel B of Table 5 we report the results of the regressions with the criteria used to identify neobanks (i.e., number of branches below or equal to 5, age below or equal to 20 years and online functionalities) as the main independent variables, instead of the neobank dummy. In general, the findings suggest that the baseline results (higher interest income, non-staff expenses and impairment charges) are driven by the online functionalities, and not the other criteria. Moreover, this analysis allows us to uncover several results that are coherent with our prior expectations. For instance, banks with smaller branch networks and younger banks seem to record relatively higher funding costs than other banks. The fact that no such effect is found for the neobank dummy (Panel A) suggests the potential presence of a ‘digital discount’ on funding costs, possibly linked to the superior online customer experience, as previously discussed. Another interesting result is related to staff expenses, as we uncover two apparently contradicting effects for neobanks: while operating with smaller branch networks seems to reduce staff expenses, the implementation of online functionalities seems to require a greater investment in such expenses, which may be linked to the need to invest relatively more in specialised human resources (Tanda and Schena 2019; Williams 2021).

To examine this question, we re-run the baseline regressions including two interaction terms with the neobank dummy, representing banks with low values (below the 10th percentile) and high values (above the 90th percentile) of size (Table 6, Panel A), age¹⁵ (Panel B), and income diversification (Panel C); and we split the original sample into buckets of size (Table 7, Panel A), age (Panel B) and number of products (Panel C), and run sub-sample regressions using ‘non-staff expenses’ as the dependent variable (DeYoung 2005).

Regarding the potential presence of economies of scale, the results in Table 6 show that neobanks in the top decile of size significantly outperform other neobanks, and this can be traced to their ability to record lower non-staff expenses, as reported in column (7). Moreover, the net effect on non-staff expenses (i.e., the sum of coefficients for the neobank dummy and neobank dummy*top decile of size) becomes negative, suggesting that larger neobanks may be more efficient than traditional peers. With respect to the sub-sample regressions, the results reported in Table 7 (Panel A) suggest that the underperformance of neobanks in terms of non-staff expenses fades away as we remove very small banks (less than EUR 600 million) from the sample. In general, the results suggest that in our sample of relatively small banks (total assets below EUR 10 bn), neobanks enjoy some ‘depth’ of economies of scale, in line with earlier evidence collected by DeYoung (2005) and Delgado et al. (2007).

With respect to economies of experience, the results in Panel B of Tables 6 and 7 suggest the performance gap towards traditional peers is not smaller for older neobanks. As a case in point, the regression results reported in columns (7–12) of Table 7 indicate that the underperformance of neobanks persists as we exclude younger banks, at different age thresholds. Notwithstanding, we notice that older neobanks seem to charge higher fees and commissions – as observed in column (13) of Table 6. In our view this finding is consistent with the ‘digital spatial capture’ narrative (Boot et al. 2021), according to which as customers become more acquainted with a bank’s technology, their switching costs become higher, hence allowing digital banks to charge fees for their services.

As for economies of scope, in general, we find similar results to the economies of scale. Namely, the coefficients reported in Table 6 (Panel C) show that (i) neobanks in the top decile of income diversification record a better performance in terms of ROA than less diversified neobanks, (ii) this effect seems to be exclusively driven by efficiencies in terms of non-staff expenses, and (iii) the overall effect (coefficient of neobanks plus the coefficient of neobanks*top decile of diversification) is negative and sizable – suggesting that more diversified neobanks may be more efficient than traditional peers. Coherently, the findings in Table 7 (Panel C) reveal that neobanks no longer exhibit non-staff inefficiencies after broadening the scope of products offered (namely 6 or more). In our view, such results provide support regarding the presence of economies of scope in neobanking.

We provide answers to this question by re-running the baseline regressions in three sub-samples of banks based on whether they offer payments (credit and debit cards), loans (personal and mortgage loans) and services (broker & advisory services and insurance). Moreover, to improve the comparability of coefficients on the interaction terms, we remove from each sub-sample the banks that offer all products of that specific business line. For instance, in column (1) we remove 147 banks from the baseline sample (of which, 17 neobanks) that offer both credit and debit cards. The results are presented in Table 8.

Overall, we only find statistically significant results in the ‘Loans’ sub-sample. Namely, the results in column (2) indicate that neobanks offering personal loans perform significantly worse than other neobanks, whereas the same does not occur for those offering mortgage loans. This result is backed by untabulated results, which show that top performing neobanks (i.e., with above median ROA) are significantly less likely to offer personal loans than worse performing neobanks (41.9% vs 67.7%), whereas such difference is much smaller for mortgage loans (29.0% vs 38.7%). Moreover, the overall effect (i.e., the sum of coefficients for the neobank dummy and neobank dummy*personal loans) becomes negative, suggesting that neobanks that offer personal loans perform significantly worse than traditional peers. In our view this finding seems in line with the literature suggesting that hard information can be used more effectively to assess borrower creditworthiness in collateralised loans (e.g., mortgage loans) than in non-collateralised loans (e.g., personal loans) (Stein 2002; DeYoung et al. 2007).

A tangible concern regarding our empirical setting is related to the possibility that some unobserved features of banks may simultaneously affect their propensity to follow a business model and their performance (Clougherty et al. 2016). To address this issue, we apply 2SLS estimation using two instrumental variables (IVs) that reveal the banks’ access to the knowledge necessary to purse certain digital strategies; and such knowledge spillovers are expected to impact the performance of banks mainly via the digital business model channel.

Our first IV is the proximity to knowledge centres, calculated as the natural log of the road distance between the bank’s headquarters and the nearest top50 university,¹⁶ according to the ‘2021 Scimago Institutions Ranking’. The distance is measured in number of road hours and refers to the distance between the NUTS2 regions of the bank and the university (Persyn et al. 2020). Our rationale for the choice of this IV is that banks with headquarters positioned closer to knowledge centres are more likely to access the specialised resources (human and technological) necessary to adopt certain digital strategies (Tanda and Schena 2019; Williams 2021). An opposing argument, however, could be made that the location of certain knowledge centres may, in turn, be a function of the proximity to employers, such as banks, which would entail a problem of reverse causality. While to the best of our knowledge there is no work that studies the relationship between the location of bank headquarters and knowledge centres, anecdotal evidence would suggest that most top universities are centenary-old institutions, which face very significant relocation costs (and would hardly make such costly decision based on the proximity, or lack thereof, to specific employers). As such, we are confident that this may be considered a suitable instrument.

The results of the 2SLS estimations are presented in Table 9. The post-estimation tests show that the Cragg-Donald F-test is 11.30 which is above the rule of thumb of 10, as suggested by Staiger and Stock (1997) and just below the 15% TSLS bias critical value of 11.59 as defined by Stock and Yogo (2005). Additionally, we report the Anderson LM statistic of 22.094, which suggests the rejection of underidentification at the 1% level. Regarding the first-stage regression, the proximity to knowledge centres increases the likelihood to follow a digital banking model. We also find that the inclination to adopt a digital model is positively affected by the level of excellence of the nearest knowledge centre. As for the second stage result, the neobanks dummy is found to negatively affect ROA, lending support to our baseline regressions.

Given that we do not include age or branches as matching criteria between neobanks and traditional peers, one legitimate concern is whether the performance differences are chiefly related to the fact that neobanks are younger and have relatively less branches than traditional peers, rather than to the fact that they adopt a digitally native business model. Indeed, as shown by the distribution of banks per age and branch buckets (Fig. 2), one can observe that, as expected: (i) with reference to age, neobanks are all in the ‘0–20 years’ bucket, whereas the interquartile (Q3-Q1) of traditional peers is located in the ‘20–50 years’ bucket, with particular concentration in three age buckets: ‘15–20’, ‘20–25’, and ‘25–30’ years; and (ii) regarding branches, all neobanks are located in the ‘0–5’ branches bucket, whereas 54.0% of traditional peers in the matched sample (n = 169) are also positioned in this bucket.

To address this potential concern, we adopt several strategies that allow us to compare the performance of neobanks with similarly young and relatively branchless traditional peers. Regarding age, the results of the tests are presented in columns (1–4) of Table 10. First, we include age as an additional matching variable in the PSM. As expected, this step significantly reduces the age gap between neobanks and controls, even though we still find a statistically significant difference. Namely, the mean age gap moves from 37.0 years (neobanks: 10.8 years vs traditional: 47.8 years) to 9.3 years (10.7 years vs 20.0 years, respectively). Also, as expected the number of matched observations significantly reduces (from n = 376 to n = 226). Second, we retain in the sample only banks born in 2000 or after and re-run the PSM. Again, this narrows the age gap significantly to just 2.2 years, at the cost of a significantly smaller matched sample (n = 132). Third, based on the notion that the first neobanks were born in 1995 (namely, Nest bank, Boursorama and Skandiebanken) we remove from the original matched sample (n = 376) banks older than 25 years. In line with the previous strategies, this reduces the age gap in an important way (to 4.9 years) at the cost of a significantly smaller sample (n = 157). Fourth, we re-run the baseline regression removing banks that are less than 5 years old, and the baseline results are confirmed. Overall, the baseline results hold for all specifications.

With respect to branches, the results are shown in columns (5–7) of Table 10. In line with the age-related tests, we use branches as a matching variable in the PSM (column 5), and we remove banks with more than 5 branches before and after running the PSM (column 6 and 7). These procedures significantly reduce the gap between neobanks and traditional peers with respect to the number of branches, from 17.5 branches (neobanks: 0.7 vs traditional: 18.2) to 0.5 branches (1.0 vs 1.5), respectively. Our main findings are confirmed.

An additional possible concern is related to the sensitiveness of the neobanks identification to the thresholds of size, asset and funding structures, age, and branches. To address this potential issue, we adjust the thresholds of size (from 0-10b to 0-30b), the share of loans to assets and deposits to assets, (from a minimum of 5% to 10%), the age (from maximum of 20 years to 30 years) and the number of branches (from maximum of 5 to 20). We check for changes in the pool of neobanks by applying the new thresholds to each variable individually, and to all variables cumulatively.

Overall, we find that the identification of neobanks remains notably stable after changing the referred thresholds. Namely, the criteria which affects the identification the most is related to customer lending and deposits (4 neobanks are removed); increasing the size threshold allows the inclusion of 2 additional neobanks; and increasing the age to 30 years facilitates the inclusion of 3 neobanks. Interestingly, changing the number of branches does not affect the total number of neobanks identified. Alternatively, applying the four changes to criteria cumulatively leads to off-setting effects: we identify 7 additional neobanks, but also remove 4 neobanks.¹⁷

Moreover, we re-run the baseline regressions for each of the new samples of neobanks (Table 10, columns 9–12), and find that the baseline results remain unchanged. Relatedly, we apply the new thresholds on size and asset and funding structure to the sample of traditional banks and find that the main baseline results hold – as reported in column (13). In our view, both results (stability of neobank identification and robustness of ‘neobank-performance’ linkage) provide a strong case for the robustness of our approach.

In this sub-section we report four additional robustness checks, which provide further evidence on the reliability of our results. First, our sample includes 2019 and 2020 data. This period can raise concerns given that the Covid-2019 pandemic (which developed in 2020) significantly accelerated the development and usage of digital banking channels (Abramson et al. 2020; Dadoukis et al. 2021). We tackle this potential issue by running a robustness check using financial data exclusively from 2019 (and not the mean value of 2019–2020). By doing so, we drop one neobank from the sample, which is born in 2020. We re-run the baseline regressions and find similar results (Table 10, column 14).

Second, our baseline regressions include three country-level controls. One possible issue is the presence of omitted factors that are not captured in our country-level controls, hence affecting our results. To test whether this is the case, we re-run the baseline model including country fixed effects, and no changes occur to the baseline results (see Table 10, column 15). Nonetheless, our preferred specification remains the inclusion of country controls, rather than fixed effects, due to model overfitting issues for some of the sub-sample tests we perform throughout the paper.

Third, in our baseline approach we winsorise the dataset to tackle data quality issues. However, in general, winsorisation can create bias in the results, by affecting relatively more one of the groups than the other. To address such concerns, we analyse the distribution of neobanks and traditional peers below (above) the p1(p99) thresholds, for each of the bank controls. Overall, in the 1.4% of the full sample that is changed across the control variables, only 0.3% of values are related to neobanks – clearly indicating that most changes are related to traditional peers. As an additional test, we re-run the baseline regression using non-winsorised data and, as expected, the baseline results hold (column 16).

Fourth and final, while our preferred dependent variables, ROA and its decomposed elements, have been widely studied in the banking literature (e.g., Mergaerts and Vander Vennet 2016; Marques and Alves 2021), other proxies have also been used to provide insights into the performance of banks. As such, next we re-run our baseline regressions using Return on Equity (ROE) as an alternative dependent variable. Untabulated results show a negative coefficient of the neobank dummy (−0.080), significant at the 5% level, which confirms the previous findings.