Bail‐In Rules and the Pricing of Italian Bank Bonds

We analyze whether the introduction of the bail&#8208;in tool in January 2016 affected the pricing of Italian bank bonds. Using a unique dataset of 1,798 fixed&#8208;rate bonds issued during the period 2013&#8211;2016, we find an increase of the spread at issuance of bail&#8208;inable bonds compared to non&#8208;bail&#8208;inable bonds. This increase also depends on the intrinsic characteristics of each bank. Large institutions, banks with lower ratings, profitability, capitalization, and higher liquidity faced a higher cost of issuing bail&#8208;inable bonds. Overall, our results seem to support the hypothesis of an improved market discipline for the bank bond primary market.

Second, Italian banks place most of their bonds (around 80%) directly with their customers over the counter (OTC) (Coletta & Santioni, 2016;Gentile & Siciliano, 2009), thus avoiding the costs associated with the listing of these securities and, more importantly, this gives rise to an obvious conflict of interest. As a consequence, unsophisticated retail investors are the main holders of the banks' debt securities (Grasso et al., 2010). Italian banks have benefited from their placing power, which allowed them to raise funding via bond issuances quite cheaply. In fact, Del Giudice (2017) finds that bank bonds were issued at a negative spread (compared to risk-free securities) prior to the entry into force of the Market in Financial Instruments Directive (MiFID) I regulation. Interestingly, the same author observes that bank bond spreads turn positive as the MiFID I increases the competition among banks and trading venues.
Third, the National Resolution Authority (i.e., the Bank of Italy) applied a quasi bail-in to four small banks (Banca delle Marche, Banca Popolare dell'Etruria e del Lazio, Cassa di Risparmio della Provincia di Chieti, and Cassa di Risparmio di Ferrara) in November 2015, a few weeks before the official implementation of the bail-in resolution tool. 4 Notably, although those four regionally chartered banks accounted for less than 2% of the Italian banking market, the news of their quasi bail-in became the hot topic in the news for a while, which could have enhanced investors' awareness about the negative consequences of bail-ins on their savings and investments. In addition, a vast information campaign was properly carried out by the banks in Italy. Indeed, the Italian stock market regulator (CONSOB) late in November 2015 required all the intermediaries to appropriately inform investors about the consequences of the implementation of the BRRD. 5 A survey conducted by the Italian Bankers' Association (Associazione Bancaria Italiana, ABI) in January 2016 also reveals that several initiatives were adopted to massively inform customers about the switch to the new bail-in regime via, for instance, a specific leaflet enclosed with the monthly/quarterly bank statements, as well as through fliers handed out at bank branches. 6 Furthermore, although Italian investors are often blamed for lacking adequate financial literacy (see, among others, Bartiloro, 2011), a study by Accornero and Moscatelli (2018) reveals that at least recently the information regarding the banks' fundamentals, such as the Tier 1 capital ratio, influences Italian households' decisions. Similarly, Boccuzzi and De Lisa (2017) document that market discipline was properly working in Italy around the time the BRRD became effective. Overall, this leads us to reasonably assume that Italian investors improved their awareness about the potential negative effects arising from funding unhealthy banks.
Interestingly, the BRRD introduction might hit the banks with a different intensity depending on their intrinsic characteristics. Indeed, for those banks that in the pre-BRRD era would have benefited the most from an implicit guarantee, we should observe a greater impact of the new bail-in rules. Notably, bank-specific characteristics should influence their bond funding costs with different intensity in the pre-and post-bail-in phases. To the extent that the bail-in tool is valued as a credible mechanism by the market, riskier institutions will be the ones who will experience a higher cost of funding after the introduction of the new rules, as well as large banks that should no longer benefit from any implicit public guarantee.
At present, the existing literature mainly focuses on the effect of the new regulatory framework on bank-specific risk and financial stability. For example, using an event-study methodology, Schäfer, Schnabel, and Weder di Mauro (2016) analyze the reactions of credit default swap (CDS) spreads and share prices related to European banks after the announcement of some operations in which investors suffered a total or partial bail-in. The authors find evidence of a significant increase in CDS spreads and a reduction of share prices. More pronounced reactions take place in those countries where the weaknesses of public finances make it more difficult for them to implement the bailout of a large bank. Moreover, the authors emphasize that it is the actual occurrence of a bail-in, rather than the simple introduction of a new legislation, that produces a concrete reaction. In a similar fashion, Mikosek (2016) investigates the CDS spreads of 20 banks from six European countries, and compares them to the CDS spreads of the corresponding domestic governments. Starting from 2015, the ratio of the average bank CDS spreads over sovereign CDS increases substantially, showing a sharp misalignment between sovereign and bank risk perceptions. This phenomenon demonstrates that the players of this specific market (CDS market) start at one point to discount the fact that governments will not rescue banks any longer. As regards the secondary bond market, a recent study from Giuliana (2018) using a difference-in-difference approach finds that bail-in events amplified the difference in daily yield between bail-inable (non-secured) and non-bail-inable (secured) bonds. These findings support the notion that the authorities' efforts to introduce the bail-in regime increased the bail-in expectations in the secondary market. Moreover, Giuliana (2018) provides evidence that the bail-in events reinforced the relationship between a bank's default probability and the price of its securities.
Similar to Giuliana (2018), we investigate the effectiveness of the new banking regulation by comparing the pricing reaction of unsecured and secured bonds. Indeed, as pointed out by Chan-Lau and Oura (2016), the bail-in increases the cost difference between bail-inable and non-bail-inable bonds.
However, a distinct feature of our paper is that we focus on the primary market of Italian bonds. Notably, we contribute to the literature by measuring the impact of the BRRD introduction on the bank cost of issuing bonds. Indeed, the majority (80%) of bonds issued by Italian banks are not publicly listed. Hence, by focusing on the secondary market we would be disregarding a substantial share of the market. Additionally, we would not be able to offer evidence of the actual cost of funding borne by the banks which, instead, is observable from the primary market (i.e., Chan-Lau & Oura, 2016;Sironi, 2003;Zaghini, 2014). To this extent, our research work aims to provide further evidence on the credibility of the bail-in mechanism, which adds to current findings from the secondary market.
This paper is also related to the empirical literature analyzing market discipline as a regulatory tool (see, among others, Bliss & Flannery, 2002;Calomiris, 1999;Calomiris & Kahn, 1991;Flannery, 2001;Hellwig, 2005). As Bliss and Flannery (2002) highlight, market discipline consists of two distinct components: (i) monitoring, which refers to how market prices reflect the financial condition of a single bank; and (ii) influence, which, instead, describes how such market information affects the incentives for managers to engage in risk-taking behavior. Because we analyze whether the riskiness of banks has an impact on their cost of issuing bonds pre-and post-BRRD, our work relates to the first category in this classification. Some studies in this field of the literature show that bondholders are quite sensitive to bank-specific risks (Covitz, Hancock, & Kwast, 2000;DeYoung, Flannery, Lang, & Sorescu, 2001;Flannery & Sorescu, 1996;Hancock & Kwast, 2001;Jagtiani, Kaufman, & Lemieux, 1999). With respect to the European market, Sironi (2003) analyzes the primary market spread of a sample of subordinate bonds issued by European banks in the period 1991-2000. The results of this study show that investors properly price the specific risk factors of each issuer, and that in the second half of 1990s, the too-big-to-fail (TBTF) effect tends to disappear. More recently, Balasubramnian and Cyree (2011) pointed out that, in the United States, the hypothesis to reintroduce a government guarantee for large banks, after the default of the LTCM (Long-Term Capital Management) hedge fund, reduced the sensitivity of the spread to bank-specific characteristics. Similar conclusions were reached by Santos (2014) who shows the existence of a TBTF effect both in the banking as well as in the non-financial sectors, as large non-financial corporations also enjoyed the advantages due to their size. Yet, the TBTF effect is quite strong for large banks, suggesting that investors believe that the probability of a bailout for these intermediaries in time of crisis is very high. A more recent analysis by Acharya, Anginer, and Warburton (2016) indicates the presence of a TBTF effect in showing that the bond spreads of small and medium banks are more risk-specific related, while this characteristic tends to disappear for bonds issued by large banks.
Using a sample of 1,798 fixed-rate senior bonds issued by 28 Italian banks from January 2013 to December 2016, we first document an increase of bond funding cost in 2016 upon the adoption of the BRRD. Precisely, the average spread (i.e., difference between the yield to maturity at issuance offered by the bank bonds and the yield offered by government bonds with corresponding maturities) followed an interesting path during the observed period, going from 0.56% in the pre- bail-in period (2013-2015) to 0.70% for bonds issued after 2016. Notably, the difference between the average spreads in the two sub-periods appears to be statistically significant.
Our regression analysis confirms that, even when we account for bond and bank characteristics as well as bank fixed effects and bank-time fixed effects, banks faced a higher cost when issuing bailinable bonds compared to bonds not subject to the new regulation.
Consistent with the existing literature, we find that banks characterized by lower ratings, profitability, capitalization, and higher liquidity were forced to pay higher spreads to place their bonds with the new bail-in regime. In addition, we observe that large banks (and especially the largest ones) were able to pay a lower spread until 2015. In contrast, after 2015 they face, ceteris paribus, an increase in the cost of funding. Eventually, the implementation of the new Directive might have mitigated the too-big-too-fail effect, which is consistent with the findings of, among others, Zaghini (2014).
Moreover, following Schäfer et al. (2016), we show that the implementation of the quasi bail-in just a few weeks before the effective entry into force of the BRRD contributed to lead retail investors to demand a higher premium for their investment in bank bonds.
We confirm the robustness of our results by conducting several additional analyses. To mitigate the concern that the rise in the spread in the primary market is a consequence of a generalized increase of banks' risk, rather than a result of the BRRD approval, we carry out the following robustness tests. First, following Schäfer et al. (2016) and Giuliana (2018), we show that the spread between bail-inable and non-bail-inable bonds reacts also to events that do not produce a significant increase in banks' risk, such as when the EU Legislative Bodies voted the resolution to adopt uniform rules for the resolution of banks in April 2014. Second, we exclude from our analysis those banks that, during the observed period, faced serious undercapitalization problems or were subject to any public interventions, and we show that our main findings are not driven by their inclusion in our sample.
Additionally, because the non-bail-inable bonds were issued by a limited number of intermediaries, we run our regressions on a subsample of banks that excludes those that did not issue any non-bailinable bonds, and demonstrate that our results remain consistent. Furthermore, because two banks issued almost one third of the bonds in our sample, we rerun our models on a subsample that excludes them in order to rule out the possibility that the observed increase in the spread is driven by those banks' intrinsic characteristics rather than be representing an overall consequence of the BRRD implementation. Results from this subsample corroborate again our hypothesis about the existence of a regulation effect. Finally, we show the robustness of our findings even when we employ different proxies of the bank characteristics.
Overall, our analysis offers interesting policy implications because, as prescribed by the new legislation, authorities should also count on market discipline to improve their prudential supervision. Higher risk sensitivity, indeed, is especially needed in countries like Italy where, historically, market discipline has not been working properly.
The remainder of the article proceeds as follows. The next section summarizes the relevant features of the new regulatory framework. Section 3 describes our data and provides a discussion of key summary statistics, and the empirical research methods we employ in this study. Results are presented in section 4, while additional analyses are offered in section 5. Finally, section 6 concludes.

| REGULATORY FRAMEWORK
Motivated by the need to design a common toolkit for bank resolutions across the globe, as well as addressing the TBTF issue − thus, to prevent new bailouts at taxpayers' expense − the Financial Stability Board (FSB) set a framework of rules aimed at ensuring ordered resolutions of banks and limiting the use of state coffers. These principles were transposed in the European Legislation via the adoption of the so-called BRRD in May 2014. The most known resolution tool introduced by the European Directive is the 'bail-in' (as opposed to bailout), which was made available at each EU member state's Resolution Authority since January 2016. 7 In essence, under given circumstances, the National Resolution Authority (or the Single Resolution Board, under its power over the national bodies) is allowed to impose the losses of a failing bank on its owners and creditors, according to a pre-defined hierarchy. Specifically, loss absorption via 'bail-in' is achieved by first writing down Common equity Tier 1, then Additional Tier 1 capital, Tier 2 capital, all other Subordinated Liabilities, all other Senior Unsecured Liabilities, and finally eligible deposits over 100,000 euros. Notably, the bail-in tool may be applied to recapitalize an institution, as well as to convert to equity or reduce the principal amount of claims that are transferred either to a 'bridge institution,' or under the 'sale of business' or 'asset separation' tools (BRRD, Art. 43). This should enable a fairer resolution process of the banks, by excluding (or at least limiting) the injection of taxpayers' money. Furthermore, some liabilities are explicitly excluded from the bail-in scope, such as covered deposits, secured bonds, liabilities to other financial institutions with an original maturity of less than 7 days, employee remuneration, liabilities to commercial or trade creditors relating to the provision of critical goods or services, liabilities to tax and social security authorities that are preferred by law, and liabilities for contributions to deposit guarantee schemes. 8 These exclusions would, de facto, make the excluded claims senior to bail-inable debt.
Overall, by designing the bail-in tool, the policymaker aimed at increasing the incentive, for creditors, to monitor the health of banks during normal times, thus limiting the occurrence of new bank failures. 9 7 Apart from the bail-in, the BRRD toolkit also includes: sale of business, bridge bank, and asset separation. See also: https://srb.europa.eu/en/content/tasks-tools 8 More information can be found here: https://srb.europa.eu/en/content/resolution-qa 9 In this regard, see recital No. 67 of the BRRD: http://eur-lex.europa.eu/legal-content/EN/TXT/HTML/? uri=CELEX:32014L0059&from=EN

| DATA AND EMPIRICAL METHODOLOGY
To select our sample of bonds, we start out with all the banks in the Thomson Reuters Eikon database that issued fixed-rate bonds from January 2013 to December 2016. We further rely on the following criteria and select: (i) banks that issued fixed-rate bonds both before and after the introduction of the BRRD in order to avoid potential issues related to sample selection that could bias our main results; (ii) banks that are deposit-taking and loan-making institutions (Beltratti & Stulz, 2012); 10 and (iii) banks with at least one bond issuance for every given year in the sample period. All in all, these criteria lead us to a final sample of 28 banks (see Table A1 in the appendix for a list of the sampled banks). 11 For each selected bank, we then supplement our sample of bonds by manually searching, in each bank's website, potential bond issuances − especially non-listed ones − that are missing from the Thomson Reuters Eikon database.
Data concerning the characteristics of the bonds are obtained from Thomson Reuters Eikon or hand-collected directly from the final terms of each issuance when they are not available. Following Sironi (2003), we collect data regarding the coupon offered by fixed-rate bonds, the frequency of coupon payment, the size of the issuance, the maturity, and the listing venue, if any. Consistent with previous literature (e.g., Gabbi & Sironi, 2005;Iannotta, 2011;Iannotta, Nocera, & Resti, 2013a;Sironi, 2003), we exclude perpetual bonds, while we include bonds denominated in euro, with no optional component (e.g., call or put option). Overall, this selection procedure leads us to a unique dataset of 1,798 bonds issued by a sample of 28 Italian banks during the period January 2013-December 2016. Table A1 in the appendix also shows the distribution of bonds, by bank.
Notably, our sample only includes senior bonds in the fixed interest category (fixed coupon or zero coupon). This choice is motivated by several reasons. First, as pointed out by Santos (2014), 'unique' features of bonds such as floating rates and call options can affect bond pricing. Second, non-structured senior bonds are the most common type of bonds in Italian households' portfolios (Coletta & Santioni, 2016) and, more importantly, this allows us to measure the yield to maturity at issuance, which ultimately represents the effective cost of funding borne by the banks. Since our goal is to study the impact that the adoption of the BRRD has generated on the cost of funding borne by the banks when issuing bonds, we refer exclusively to the returns offered in the primary market − following the approach used, among others, by Morgan and Stiroh (2000) and Sironi (2003). Indeed, as pointed out by Zaghini (2014), the fluctuations of the returns observed in the secondary market do not influence the cost of funding for the issuing banks. Moreover, for many securities included in our sample, the secondary market is non-existent. Indeed, 80% of the bonds in our dataset are not publicly traded on a regulated stock exchange or multilateral trading facility (MTF). In any case, because our aim is to investigate the effect of regulation on the bond funding costs, this would be precluded when analyzing the secondary market.
It must be noticed, as well, that the yield to maturity at issuance is not available in Thomson Reuters Eikon. Therefore, we hand-collected this information from the final terms of each bond issuance. The yield to maturity at issuance is then compared to the yield offered by Italian 10 As in Beltratti and Stulz (2012) we require a deposit to assets ratio above 20% and a loan to assets ratio above 10%. 11 Because bank holding companies (BHCs) operate internal capital markets, shocks to one part of the organization are likely to be transmitted to other subsidiaries (Houston, James, & Marcus, 1997). We therefore use the BHC's information to summarize the condition of all its subsidiary lenders. We link each relationship lender's information with its ultimate BHC parent.
sovereign bonds with similar maturity in order to construct our variable of interest, which is Spread. 12 A first summary of the data shows that the average number of issuances, per year, is slightly greater than 520 during the first three years of the investigated period, with a peak of 610 and 608 in 2013 and 2014, respectively. In 2016, conversely, the issuances fall to 212, of which 70% are in the first semester of the year. This decrease is probably due to the expansionary monetary policies adopted by the European Central Bank (ECB) (Bufacchi, 2017), which reduced the necessity of funding from retail investors.
In the first months of the observed period Spread takes negative values, reaching a minimum of −3.13%, as reported in Panel A of Table 1. This appears to contradict theoretical assumptions according to which sovereign bonds should be considered risk free by domestic investors, or at least as a basis rate to price non-sovereign bonds. Yet this phenomenon, well known by practitioners, is not new also to academics: even if on the basis of different samples and periods, other researchers have indeed indicated the presence of negative spreads in the Italian bank bond market (i.e., Del Giudice, 2017; Grasso et al., 2010). As mentioned above, the motivation could lie in the commercial skills of the bank salesforces and the bank placing power (Del Giudice, 2017). The variable Spread also shows a wide dispersion over the sample period; this could be due to bank-and bond-specific characteristics; therefore, we explore how those variables are related to each other by employing a multivariate setup in section 4.
In Panel A of Table 1 we also report the descriptive statistics − for the entire period − of the variables related to the characteristics of each bond (BOND VAR) utilized in our analysis. In contrast, in Panel B we report a snapshot of the summary statistics of the same variables in two distinct sub-periods, namely pre-bail-in (2013)(2014)(2015) and post-bail-in adoption (2016).
The average spread for the entire period is 0.58%. However, as reported in Panel B of Table 1, the analysis of the two sub-periods reveals that, prior to the switch to the bail-in regime, the average spread was 0.56%, while after January 2016 the average spread increased by about 14 basis points reaching the value of 0.70%. This is not a minor increase if we think that it represents a surge of about 25% in the interest rate spreads from the bailout era to the new bail-in regime. 13 Concerning the other BOND VAR, we note that the average maturity for the entire sample is less than 4 years, the average amount issued is 53.74 million euros, and less than 20% of the bonds are listed on a regulated stock exchange or MTF. We do not observe any significant change of the main bond characteristics when comparing the pre-and post-2016 period, as the mean values of such BOND VARs in the two sub-periods are not statistically different.
To ascertain whether the introduction of the BRRD increases the cost of funding of bank bonds, we estimate a regression model that includes: (i) a dummy, labelled Post 2016, which equals one when a bond was issued after 1 January 2016 and zero otherwise; (ii) a dummy, labelled Non_Bailinable, which equals one when a bond is either secured, senior secured, or asset backed, and zero otherwise (i.e., senior unsecured, unsecured); as well as (iii) an interaction term between Post 2016 and Non_Bailinable dummies. Particularly, this interaction is essential for us to test the differential impact of the new bail-in regime on investors in bail-inable bonds and non-bail-inable securities, which are de 12 We use the Rendistato − as a proxy of the Italian sovereign bond yield −, which is an average return of sovereign bonds computed for a variety of maturities and published each month by the Bank of Italy. See: https://www.bancaditalia.it/ compiti/operazioni-mef/rendistato-rendiob 13 Please note that the means from the two subsamples are statistically different (unreported t-tests).
facto excluded from the bail-in scope. Specifically, we estimate the following regression model: where BOND VAR is a vector of bond characteristics, which includes the following variables: Maturity = the time to maturity, in years, measured at the issuing date (as in Iannotta et al., 2013a); Size = the log of the amount issued (as in Iannotta & Navone, 2008); Listed = a dummy equal to one if the bond is listed on a regulated stock exchange or MTF; Step-up = a dummy equal to one if the bond has a step-up structure. BANK VAR is a vector of lagged variables 14 related to the bank characteristics extracted from the Bureau van Dijk 'Orbis-Bank Focus' database, which includes the following: Bank Rating = a variable that, following the rating scale provided by Iannotta, Nocera, and Sironi (2013b), associates numerical values to the mean of the ratings assigned (to each issuance) by S&P, Moody's, and Fitch, with higher values denoting greater risk; Bank Size = the natural logarithm of total assets; TBTF (TBTF2) = a dummy that equals one if the total assets of the issuing bank are higher than the average (75th percentile) total assets of the sample in a given year; Tier 1 ratio = ratio of Tier 1 capital to risk-weighted assets; ROAA = return on average assets (computed by Orbis as net income divided by the average of total assets at the beginning and at the end of the period); Liquid Assets ratio = ratio of liquid assets and the sum of customer deposits and short-term funding (measured by Orbis); NPL ratio = ratio of non-performing loans and the total amount of outstanding credits (measured by Orbis). TIME FE are year dummies that we add to control for changing market conditions that could influence the value of the spread.
BANK FE are bank fixed effects that we include in order to control for unobservable, time-invariant, bank characteristics that might influence the bond yields. Additionally, in one of our specifications, we replace TIME FE and BANK FE with BANK * TIME fixed effect, which allows time-varying bank unobserved heterogeneity to be controlled for. Standard errors are clustered at the bank level.
We expect β 1 to be positive as investors should ask for a higher risk premium upon the entry into force of the BRRD. In contrast, β 2 should exhibit a negative sign provided that non-bail-inable securities, on average, would offer lower returns than bail-inable ones, due to their lower riskiness. The coefficient of the interaction term, β 3 , should also be negative, as we expect holders of bail-inable securities to ask for a higher risk premium, compared to non-bail-inable debtholders, once they realize that their investments are potentially at risk due to the implementation of the new bail-in regime.
The second regression model adds the interactions between bank characteristics and the Post 2016 dummy, thus, to investigate if significant differences in investors' reaction, due to the bank characteristics, emerge after the BRRD became effective.
Specifically, we estimate the following regression model: As regards the bond features, we expect the size of the issuance to be negatively correlated with Spread. Indeed, by issuing larger amounts of bonds, banks can benefit from a decrease in the cost of funding due to better economies of scale. In addition, greater issues are usually offered by larger banks that can more easily target a broader share of least bargaining retail investors, given their wider distribution network (Sironi,14 We utilize banks' balance sheets data related to the year preceding the issuance of the bond as that is the most recent public information available, to investors, at the issuing date (as in Zaghini, 2014). In other words, the data from banks' statements in year t−1 is utilized to generate our BANK VAR(s) that are matched to the BOND VAR(s) related to bonds issued in year t. 2003). The maturity, at least theoretically, should be positively correlated with our dependent variable, as higher yields should be offered to bonds with longer redemption horizons (Zaghini, 2014). Furthermore, we expect listed bonds to be cheaper for banks with respect to non-listed ones because the access to capital markets should guarantee a liquid investment to the investor. However, we are conscious that investors may not price the liquidity premium because of the low level of bank transparency. Similarly, step-up securities could be overall cheaper for banks, as investors might not be completely able to fairly price such kind of securities because of the appealing structure of their increasing coupons.
Moving to the bank characteristics, we expect that investors demand greater returns from bonds issued by riskier banks. For this reason, the variable Bank Rating − that takes, by construction, greater values when the creditworthiness is lower − should be positively correlated to our dependent variable, thus implying an increase in the cost of funding. In contrast, we expect Bank Size to reduce the spread, given that large banks should benefit from the implicit guarantee of the TBTF phenomenon, but also because they should be able to better diversify their business and have a more skilled management (Santos, 2014;Sironi, 2003;Ueda & Weder di Mauro, 2013). Tier 1 ratio should have a negative effect on spreads, because more capitalized banks normally enjoy a good reputation in the market and can raise money at a lower rate of return. Nevertheless, Herring (2010) shows that those banks that required a government intervention during the financial crisis had, on average, more regulatory capital than those not requiring it. It would follow that regulatory capital is not always an efficient regulatory tool. With respect to the issuer's profitability, as measured through ROAA, we expect it to exert a negative impact on spreads, considering that a greater profitability should signal a greater efficiency (Sironi, 2003). Alternatively, a higher ROAA could also reflect a greater risk propensity by banks; in such a case, we should find a positive correlation with the spread (Flannery & Sorescu, 1996). The effect of the Liquid Assets ratio can be either negative or positive. Indeed, on the one hand, bank liquidity can have a positive impact on bank solvency. On the other hand, greater values of the Liquid Assets ratio can be perceived as a sign of inefficiency in the management of the liquidity or, alternatively, they could inspire managers to take on more risk (Myers & Rajan, 1998), and increase conflicts of interest between managers and shareholders (Jensen, 1986). Lastly, it is reasonable to expect that a greater exposure to impaired loans (as measured by the Non-performing Loans (NPL) ratio) increases the spread at issuance because of a greater credit risk exposure and a higher uncertainty about future performances (Flannery & Sorescu, 1996). Table 2 offers descriptive statistics of the abovementioned variables. Notably, we observe that the average amount of impaired loans over total loans is about 15%, with a maximum value of 36%. The This table presents summary statistics of the bank characteristics. Bank Rating is a variable that associates numerical values to the mean of the ratings assigned (to each issuance) by S&P, Moody's, and Fitch, with higher values denoting greater risk. Bank Size is the natural logarithm of total assets. Tier 1 ratio is the ratio of Tier 1 capital over risk-weighted assets. ROAA is the return on average assets. Liquid Assets ratio is the ratio of liquid assets over the sum of customer deposits and short-term funding. NPL ratio is the ratio of non-performing loans over the total amount of outstanding credit. average Tier 1 ratio is a little greater than 11% across the whole period, while the mean of the Liquid Assets ratio is 12.5%. Moreover, it is worth noting the poor profitability of the banks, as measured by the ROAA, during the observed period. Table A2 in the appendix reports the correlation matrix related to the regressors employed in our analyses, which reveals that multicollinearity is unlikely to be a concern.

| RESULTS
The estimates of our regression model (1) are reported in Table 3. The first test − reported in column 1-is carried out on a simplified version of model (1) where we do not include the interaction term between Post 2016 and the Non-bailinable dummies. Such an interaction is then included from column 2 onwards. In columns 3-5 we add bank fixed effects. Additionally, from column 4 we add a control variable for the bank riskiness. Whereas, in columns 5-7 we include alternative measures of the bank size to check its effect on the spread at issuance. Finally, in column 8 we report a test conducted by adding bank-year fixed effects.
Overall, results confirm our earlier evidence from the univariate analysis. Indeed, the dummy Post 2016 is positively and highly significantly (at the 1% level) correlated with Spread suggesting that, after the introduction of the bail-in framework, issuing bonds became costlier for Italian banks. As expected, the Non-bailinable dummy has a negative and statistically significant coefficient (at the 1% level) in all the specifications, thus corroborating our predictions that non-bail-inable bonds carry lower yields due to their intrinsic lower level of riskiness − as being explicitly excluded from the scope of the bail-in tool. Starting from column 2, we then introduce the interaction term between Post 2016 and the Non-bailinable dummies. Its coefficient consistently exhibits a negative and significant sign across the various model specifications and is also robust to the inclusion of bank fixed effect and bank-year fixed effects in columns 3-5 and 8, respectively. Overall, this finding confirms our expectations that, upon the introduction of the BRRD, investors ask for a higher return compared to non-bail-inable bonds. Although we cannot completely rule out any alternative explanations, results from Table 3 seem to provide evidence of an improved market discipline.
Among the bond characteristics, we observe that the spread is negatively correlated with Size, which could be explained by the economies of scale and the liquidity that the bank gains when placing larger issuances. Indeed, the greater the amount issued, the lower the return that needs to be offered to the potential bondholder. This could also be explained by the fact that only larger banks and better capitalized ones are usually involved in big issuances (Zaghini, 2014). This alternative explanation seems to be supported by the fact that the coefficient is no longer significant once we control for bank fixed effects. Consistent with Rokkanen (2009) and Grasso et al. (2010), Maturity has a negative and significant coefficient. This result, apparently surprising, might be due to the significant concentration (85%) of short-medium term issuances (3-5 years) in our sample. Therefore, we repeat the analysis using different measures of Maturity (untabulated results), but the results do not change considerably. 15 An alternative hypothesis is that banks that are more creditworthy find it easier to issue longer-term bonds (Zaghini, 2014). Surprisingly, the dummy Listed enters with a positive sign, suggesting that bonds that are traded on stock exchanges are costlier than bonds traded OTC. This 15 In particular, we used three dummies for bonds with a maturity lower than 3 years, from 3 to 5 years, and more than 5 years.

TABLE 3 Bail-in regime and the spread at the issuing date
This table reports regression results related to model (1). The estimation period is January 2013-December 2016. The dependent variable is the spread between the bond yield at issuance and the yield offered by sovereign bonds with corresponding maturity. Post 2016 is a dummy that equals one when a bond was issued after 1 January 2016, and zero otherwise.
Non-bailinable is a dummy that equals one when a bond is excluded from the scope of the bail-in tool, and zero otherwise. Size is the logarithm of the amount issued. Maturity is the time to maturity, in years, measured at the issuing date. Listed is a dummy equal to one if the bond is listed on a regulated stock exchange or multilateral trading facility (MTF), and zero otherwise.
Step-up is a dummy equal to one if the bond has a step-up structure, and zero otherwise. Bank Rating is a variable that associates numerical values to the mean of the ratings assigned (to each issuance) by S&P, Moody's, and Fitch, with higher values denoting greater risk. Bank Size is the natural logarithm of total assets. TBTF (TBTF2) is a dummy equal to one if the total assets of the issuing bank are higher than the average (75th percentile) total assets of the sample in a given year. Columns 1-7 include year fixed effects. Columns 3-5 include bank fixed effects. Column 8 includes bank * year fixed effects. Heteroskedasticity-robust standard errors, clustered at the bank level, appear in parentheses. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% level. (1) (3)  No Yes result appears contradictory from a theoretical perspective, because listed bonds should be more liquid for investors. However, this coefficient is no longer statistically significant once we control for bank heterogeneity in columns 3-5 and 8. Finally, Step-up bonds show a lower return at the issuing date; it is indeed possible that investors accept a lower spread because they are influenced by the promise of an increasing coupon. Moving to column 4, we test the effect of BANK VAR on the spread for the entire sample period. Consistent with previous literature, we find that Bank Rating has a negative influence on the spread at launch: that is, the lower the issuer's rating the higher the return granted to investors and thus the cost of funding.
Further evidence about the issuer's characteristics arises from columns 5 to 7, where we focus on the importance of the bank size (Bank Size,TBTF,and TBTF2) for the spread at issuance. The results show that all the three variables enter with negative signs, supporting the idea that investors expect that large banks enjoy an implicit TBTF guarantee (Anginer & Warburton, 2014;Santos, 2014).
Interesting results also emerge from our regression model (2) whose estimates are reported in Table 4. Column 1 shows the effects of bank-specific factors on the spread. Consistently with our expectations, banks characterized by better ratings, larger size, and a lower level of non-performing loans, are able to issue bonds at lower spread. Higher Tier 1 ratio does not seem to be a selective factor for investors over the whole sample period. This evidence is in line with some literature which points out that book equity measures do not capture the banks' true ability to absorb losses (see Flannery & Giacomini, 2015).
In columns 2-7, we introduce the interaction terms between the bank risk-specific variables and the dummy Post 2016, which allows us to investigate the effectiveness of the new legislation in increasing the awareness, among investors, of the greater risks they might face in case a resolution action is undertaken against the bond issuers. The results of our analysis are indeed consistent with a higher risk sensitivity of the spread at issuance after the implementation of the BRRD. Notably, we observe that since 2016 the majority of the bank variables report statistically significant coefficients, thus confirming that the information regarding the banks' fundamentals incrementally influences investors' decisions. Specifically, banks characterized by lower ratings, and lower profitability, were forced to increase the returns of their bonds in comparison to the pre-2016 period. Interestingly, the interaction of Bank Size with Post 2016 highlights that, following the entry into force of the BDDR, large banks issued bonds ensuring, on average, higher spreads (see also Acharya et al., 2016;Zaghini, 2014) as the market figures out that the hypothesis of a public implicit guarantee is no longer reliable (Flannery & Sorescu, 1996). 16 Furthermore, we find that banks with a lower Tier 1 ratio were forced to pay a higher spread on their bonds, suggesting that retail investors started to pay more attention to this solvency indicator − which is in line with recent findings by Accornero and Moscatelli (2018). 17 Some researchers (e.g., Haldane, 2012) point out that the business of the largest banks is often tilted to trading, investment banking, and other market-related activities, so that it turns out to be less transparent, especially after the blast of the financial crisis. Therefore, the greater risk perceived by investors, which translates into a greater spread, could be due to the complexity and opacity of these 16 Although not reported, the coefficients of the interactions of Post 2016 with TBTF and TBTF2 are both positive and statistically significant. 17 After the introduction of the bail-in framework, many Italian banks increased the marketing and advertising communication about this indicator, which little by little became of common knowledge also for retail investors.  (2). The estimation period is January 2013-December 2016. The dependent variable is the spread between the bond yield at issuance and the yield offered by sovereign bonds with corresponding maturity. Post 2016 is a dummy that equals one when a bond was issued after 1 January 2016, and zero otherwise.
Non-bailinable is a dummy that equals one when a bond is excluded from the scope of the bail-in tool, and zero otherwise. Size is the logarithm of the amount issued. Maturity is the time to maturity, in years, measured at the issuing date. Listed is a dummy equal to one if the bond is listed on a regulated stock exchange or multilateral trading facility (MTF), and zero otherwise.
Step-up is a dummy equal to one if the bond has a step-up structure, and zero otherwise. Bank Rating is a variable that associates numerical values to the mean of the ratings assigned (to each issuance) by S&P, Moody's, and Fitch, with higher values denoting greater risk. Bank Size is the natural logarithm of total assets. Tier 1 ratio is the ratio of Tier 1 capital over risk-weighted assets. ROAA is the return on average assets. Liquid Assets ratio is the ratio of liquid assets over the sum of customer deposits and short-term funding. NPL ratio is the ratio of non-performing loans over the total amount of outstanding credit. All regressions include year and bank fixed effects. Heteroskedasticity-robust standard errors, clustered at the bank level, appear in parentheses. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% level. (1) (3)   banks rather than to the change in legislation. Even though the two effects (complexity and regulation) could coexist, we believe that no particular increase in the level of complexity has emerged since 2016. Therefore, the post-2016 effect can be credibly attributable to the new regulatory framework. Please also note that, having considered in our regression model both variables that are proxies of the business models and fixed effects, we are implicitly controlling for the different complexity levels at each bank.

| ADDITIONAL ANALYSES
In this section we present further sets of robustness tests. Most tables referred to in this section are found in the online appendix.

| Anticipation effect
Some researchers (see, i.e., Schäfer et al., 2016) highlight that a new legislation could produce an effect on the market also before its entry into force, more precisely when decisive resolutory actions are taken by preluding a credible implementation of the new rules. This is what happened in Italy at the end of 2015 when a quasi bail-in was implemented on four small regionally chartered banks. This event represents a unique opportunity to test whether such cases of default could have improved investors' awareness about the negative consequences of future bail-ins, possibly leading them to be more careful about their investment decisions. Therefore, we rerun our entire analysis utilizing a dummy equal to one for bonds issued after 16 November 2015, namely when the Italian government published Law 180/2015 aimed at exerting a burden sharing approach to resolve the aforementioned four small banks.
Our main results − reported in Table OA1 in the online appendix − do not change appreciably, thus supporting the idea that the market actually started to properly price bonds even before the scheduled entry into force of the new bail-in regime.

| Endogeneity issues
One potential concern that might arise from our study is that the observed increase in the spread could be due to an overall increase of banks' riskiness, rather than to an actual increased perception of bail-in occurrences. While we believe that by comparing bail-inable to non-bail-inable bonds we are mainly capturing the effect of the regulatory change, we cannot rule out this alternative explanation. Therefore, we address this issue in a variety of ways. First of all, our main regression model adds bank fixed effects that control for bank-specific, timeinvariant characteristics that account for other unobserved features that might affect our results. Our within bank analysis mitigates the possibility that sources of distress at the bank level could drive our results. In a more refined version of our model we also include bank-year fixed effects, which allow bank-specific, time-variant characteristics to be controlled for, and our main results hold.
Second, to further reduce this concern, we exclude those banks in the sample that experienced severe problems during the observed period. Specifically, the ECB required Banca Carige to boost its total capital, which deteriorated over the years as a consequence of the unresolved problem of high NPLs. Veneto Banca, instead, was orderly liquidated, in 2017, by the National Resolution Authority (Banca d'Italia) after serious undercapitalization problems came out in 2015. Therefore, in order to alleviate the concern that the observed increase in Spread is led by the relatively high riskiness of these few banks in our sample, we decided to run again our estimates by excluding Banca Carige, as well as Veneto Banca and its controlled BancApulia. Table OA2 in the online appendix reports the results arising from the analysis conducted on such a subsample. Overall, this modification broadly corroborates our conclusions. 18 Third, following Schäfer et al. (2016) and Giuliana (2018), we show that the spread between bailinable and non-bail-inable bonds reacts also to events that do not produce a significant increase in banks' risk such as, for instance, the events linked to the legislative process of the BRRD. More specifically, we repeat our tests by employing a dummy, labelled Post 2014, that equals one when a bond was issued after 15 April 2014-that is the date when (according to Schäfer et al., 2016) the EU Parliament backed the Commission's proposal on the Single Resolution Mechanism. Table OA3 in the online appendix displays the results of such a test. The results are in line with our main findings in Table 3 and are robust to a variety of subsamples. Indeed, regressions are run on the entire sample (see column 1); on a sample that excludes the bonds issued during the bail-in era (i.e., 2016) (see column 2) in order to avoid concerns related to the fact that there was an increase of generic risk among banks; and on a sample that only includes bond issuances made 1 year before and 1 year after 15 April 2014 (i.e., from April 2013 to April 2015) (see column 3). Overall, our evidence corroborates the idea that the shock on the yield spread is not necessarily the consequence of a generalized increase of banks' risk.

| Other robustness tests
Another possible issue that might affect our main analyses is that not all the banks in our sample issued non-bail-inable bonds. Hence, one might wonder whether the observed increase in Spread between bail-inable and non-bail-inable bonds is driven by the securities issued from those banks that have not issued non-bail-inable bonds across the period. To address this concern, we decide to run again our models by including only the observations related to the intermediaries that have issued at least one non-bail-inable bond over our sample period. This leads us to a sample of 991 bonds issued by 11 banks. Results from this subsample are reported in Table OA4 in the online appendix and confirm our main findings.
Moreover, because two banking groups in our sample (i.e., UBI Banca and Crédit Agricole Cariparma) have issued a significantly high number of bonds compared to the rest of the banks in the sample (see Table A1 in the appendix), one could argue that our results are driven by the intrinsic characteristics of the two aforementioned issuers. To rule out this possibility, we run again our regression model (1) on a sample that excludes the observations related to UBI Banca and Crédit Agricole Cariparma. Results from this subsample are reported in Table OA5 and confirm our main findings.
Finally, we examine how the main results are affected by the choice of the regressors by employing alternative proxies for the bank characteristics as included in the BANK VAR vector. More specifically, we use a dummy that tracks if a bank is listed (instead of the logarithm of total assets), 19 Total Capital ratio (in lieu of Tier 1), ROAE (rather than ROAA), Liquid assets over total deposits and borrowings (instead of Liquid assets over customer deposits and short term funding), Net charge off over average gross loans (as opposed to the NPL ratio), as alternative proxies for the bank size, capitalization, profitability, liquidity, and credit quality, respectively. Furthermore, we add the Cost-to-income ratio as a measure of the bank's efficiency. Table OA6 in the online appendix reports the results from this robustness check, which widely confirms our findings. 18 The Italian bank Monte dei Paschi di Siena has not issued any fixed-rate bond after the entry into force of the BRRD, therefore it is not in our sample. 19 Because the Bank FE absorb the Listed Bank variable − since this is constant over our sample period −, we first run our model without Bank FE (column 1); whereas from column 2 onwards we use Bank Size − measured as log of total assets − in order to control for unobserved bank heterogeneity.

| CONCLUSIONS
The Global Financial Crisis that started in 2007 represented a big challenge for regulators around the world. A large amount of taxpayers' money was utilized to resolve troubled banks in order to limit the negative spillovers that eventual bank defaults would have generated to interconnected financial systems. In light of this, European policymakers decided to agree, as proposed by the FSB, on a 'revolutionary' change about how to resolve stressed banks. That is why in May 2014 the European Legislative Bodies adopted the BRRD, which led to a big change from a bailout regime to an 'internal' way of rescuing banks − so-called bail-in. In essence, this radical switch in regulation aimed at transferring the bank risk from taxpayers to the banks' shareholders and unsecured bondholders. Because of this increased risk, holders of bail-inable liabilities are expected to ask, ceteris paribus, for higher returns compared to holders of liabilities that are excluded from the bail-in mechanism.
Therefore, in this paper we assess the effect of the new rules on the cost paid by Italian banks when issuing bonds, by comparing bonds that are subject to the new regulatory framework compared to bonds that are specifically excluded. To test this, we rely on a unique dataset of 1,798 fixed-rate senior bonds offered by a sample of 28 Italian banks during the years 2013-2016. The Italian market is a good laboratory for testing our hypothesis for several reasons. Italian banks are quite dependent on bonds as a form of funding, compared to banks in other European countries; also, the portion of bank bonds in Italian retail investors' portfolios is rather greater than the average of developed countries. Second, Italian banks mainly place their bonds directly (i.e., OTC) with their unsophisticated customers, giving rise to obvious conflicts of interest, which makes even more important the analysis of the effectiveness of the new regulation. Third, the National Resolution Authority (i.e., the Bank of Italy) applied a quasi bail-in to four small banks in November 2015, a few weeks before the official implementation of the new bail-in framework. Notably, although those four regionally chartered banks accounted for less than 2% of the Italian banking market, the news of their quasi bail-in became the hot topic in the news for a while, which could have enhanced investors' awareness about the negative consequences of bailins on their savings and investments.
Overall, we find that Italian banks experience, on average, a higher bond funding cost upon the introduction of the new regulatory framework. Indeed, while controlling for bond and bank characteristics, as well as bank fixed effects and bank-time fixed effects, our findings reveal that − since the entry into force of the BRRD − the average cost borne by the issuing banks increases compared to the cost borne by the government when issuing bonds with similar maturities.
Consistent with the existing literature, we observe that banks characterized by lower ratings, profitability, capitalization, and higher liquidity were forced to pay higher spreads to place their bonds with the new bail-in regime. Additional analyses also highlight that such an effect significantly emerged right after the decision of the Italian Resolution Authority to resolve four small banks in November 2015 by exerting a so-called burden sharing on some of those banks' bondholders. This confirms the existence of a bail-in effect on the cost of funding borne by the banks through the issuance of bonds. Moreover, we confirm the robustness of our findings by conducting several tests in section 5.
Overall, our main results offer interesting policy implications because, as prescribed by the new legislation, authorities should also count on market discipline to improve their prudential supervision. Higher risk sensitivity, indeed, is especially needed in countries like Italy where, historically, market discipline has not been working properly.

SUPPORTING INFORMATION
Additional Supporting Information may be found in the online version of this article. This table reports the names of the sampled banks (first column), along with the corresponding number of bonds issued (second column). The third column provides each bank's total assets in thousands of euros as of December 2016 (last observed year in our sample). Main owner's name and related share held are reported in the fourth and fifth columns, respectively. Column 6 shows if a bank is listed on a stock exchange, while column 7 provides information about the bank's type. The eighth column provides the total number of a bank's branches, whereas the ninth column reports whether a bank has branches abroad. Finally, the authority in charge of each bank's supervision − either European Central Bank (ECB)     is a dummy that equals one when a bond is excluded from the scope of the bail-in tool, and zero otherwise. Size is the logarithm of the amount issued. Maturity is the time to maturity, in years, measured at the issuing date. Listed is a dummy equal to one if the bond is listed on a regulated stock exchange or multilateral trading facility (MTF), and zero otherwise.
Step-up is a dummy equal to one if the bond has a step-up structure, and zero otherwise. Bank Rating is a variable that associates numerical values to the mean of the ratings assigned (to each issuance) by S&P, Moody's, and Fitch, with higher values denoting greater risk. Bank Size is the natural logarithm of total assets. Tier 1 ratio is the ratio of Tier 1 capital over risk-weighted assets. ROAA is the return on average assets. Liquid Assets ratio is the ratio of liquid assets over the sum of customer deposits and short-term funding. NPL ratio is the ratio of non-performing loans over the total amount of outstanding credit.

Size Maturity Listed
Stepup Bank Rating

Bank Size
Tier 1 ratio

ROAA Liquid Assets ratio NPL ratio
Non-bailinable