THE TRANSMISSION OF NEGATIVE INTEREST RATES: EVIDENCE FROM SWISS BANKS *

Transcription

1 THE TRANSMISSION OF NEGATIVE INTEREST RATES: EVIDENCE FROM SWISS BANKS * CHRISTOPH BASTEN AND MIKE MARIATHASAN DECEMBER 2018 Studying monthly supervisory data and the differential exposure of Swiss banks to a foreign-induced negative interest rate policy, we identify the transmission of negative rates as special in the following ways: First, rates stop being transmitted to depositors, which also breaks the transmission to mortgage rates. Second, banks only divert some of their liquid assets into riskier and longer-maturity assets, and otherwise retire longer-term liabilities. This implies an expansionary effect on mortgages, but a contraction and more deposit-dependence of the balance sheet. Banks ultimately remain profitable, also through fees, but become more exposed to credit, market, and interest rate risk. JEL Codes: E43, E44, E52, E58, G20, G21 Keywords: negative interest rate policy, monetary policy transmission, deposit rates, bank profitability, credit risk, interest rate risk * We would like to thank Andreas Barth (discussant), Christoph Bertsch, Jef Boeckx, Frederic Boissay, Diana Bonfim, Martin Brown, Raymond Chaudron (discussant), Jean-Pierre Danthine, Hans Degryse, Olivier De Jonghe, Narly Dwarkasing, Robin Döttling, Jens Eisenschmidt (discussant), Mara Faccio, Leonardo Gambacorta, Hans Gersbach, Denis Gorea, Christian Gourieroux, Iftekhar Hasan (discussant), Florian Heider, Johan Hombert, Robert Horat, Matthias Jüttner, Catherine Koch (discussant), Frederic Malherbe, Klaas Mulier, Philip Molyneux, Emanuel Mönch, Friederike Niepmann, Steven Ongena, Jonas Rohrer, Kasper Roszbach, Farzad Saidi, Glenn Schepens, Eva Schliephake, Bernd Schwaab (discussant), Piet Sercu, Enrico Sette, Joao Sousa, Johannes Ströbel, Ariane Szafarz, Dominik Thaler, Lena Tonzer, Benoit d Udekem, Greg Udell, Xin Zhang, as well as seminar/conference participants at ACPR-Banque de France, FINMA, Danmarks Nationalbank, National Bank of Belgium, Norges Bank, SNB, Sveriges Riksbank, Université Libre de Bruxelles, Université de Neuchâtel, Université Paris-Nanterre/EconomiX, the Annual CEBRA Meeting 2017, the CESifo Area Conference on Macro, Money & International Finance, the 14 th Christmas Meeting of German Economists Abroad, the 16 th CREDIT Conference, the ECB Workshop on Monetary Policy in Non-Standard Times, EFA 2017, the 3 rd EUI Alumni Conference, the CEPR-ESSFM 2018, the 18 th FDIC/JFSR Annual Fall Research Conference, the FINEST Winter Workshop, the 6 th Research Workshop in Financial Economics (University of Bonn), the SSES Annual Congress 2018, and the 10 th Swiss Winter Conference on Financial Intermediation for their valuable comments. All remaining errors are our own. Work using supervisory data was completed while C. Basten worked for the Swiss Financial Market Supervisory Authority (FINMA). The authors are grateful for this opportunity and for the thoughtful comments from Michael Schoch and Christian Capuano. Any views expressed in this paper remain the sole responsibility of the authors and need not reflect the official views of FINMA. University of Zürich, Department of Banking & Finance, christoph.basten@bf.uzh.ch; KU Leuven, Department of Accounting, Finance & Insurance; mike.mariathasan@kuleuven.be Work in progress. Please cite or circulate only with the authors permission.

2 1. Introduction Negative nominal interest rates have long been considered impossible. 1 Research has therefore focused on understanding monetary policy transmission at or above the zero lower bound (ZLB), while paying less attention to the dynamics when rates go negative. Following Denmark in 2012, however, central banks in the Euro Area, Japan, Sweden, and Switzerland have recently adopted negative rates as instruments of unconventional monetary policy. 2 This has made it necessary and empirically possible to also investigate the transmission channels below zero. Using detailed supervisory data from Switzerland, we contribute to this investigation by providing novel evidence on banks response to negative interest rate policies (NIRP). We study a largely unanticipated decision by the Swiss National Bank (SNB) to cut its deposit facility rate from zero to -75 basis points (bps) the lowest in all currency areas and to exempt bank-specific fractions of reserves. Specifically, and presumably to target the marginal rather than the total cost of reserves, the SNB continued to charge no interest on central bank deposits below twenty times each bank s minimum reserve requirement (MRR). 3 This policy design provides us with a measure of adverse NIRP-exposure for each bank (total minus exempted central bank deposits prior to implementation in January 2015) and enables us to identify the effects of the rate cut by comparing the behaviour of differentially affected banks over time. 4 We identify three differences with the transmission under positive rates. First, a reluctance to transmit negative rates to depositors, which is compensated for with fees and an incomplete downward transmission to lending rates. Second, despite an expansionary effect on mortgages 1 Paul Krugman, for instance, wrote as late as 2013 that the zero lower bound isn t a theory, it s a fact ( accessed: September 14, 2017). 2 Eggertson et al. (2017) go as far as calling negative interest rate policies (NIRP) a radical new policy experiment. 3 Aggregate reserves at the time were equal to 24 times the sum of all banks minimum reserve requirements (MRR), so that marginal reserves were affected independent of how total reserves were distributed across banks. 4 For the identification of adverse NIRP exposure, it does not matter whether banks initially held reserves above or below their exemption. Higher reserves above the exemption have a direct cost of -75 bps, while higher reserves below the threshold have an opportunity cost of similar size (banks essentially loose a profitable arbitrage opportunity by which they borrow on the interbank market, at negative rates, and lend to the central bank for free). We discuss the assumptions underlying our identification in more detail in Section 5. 2

3 and a rebalancing of the portfolio towards riskier assets, a contractionary effect on the balance sheet as a whole. And third, an increase in maturity mismatch and therefore interest rate risk. 5 Since banks willingness to retain their depositors constrains deposit rates at zero, the rate cut does not reduce banks funding costs as much as under positive rates. This creates pressure on profit margins and explains the weaker transmission to lending rates and the higher fee income. The contractionary effect on the balance sheet, instead, occurs first, because the rapid transmission of the reserve rate quickly turns other liquid assets into unattractive substitutes and second, because allocating reserves to other asset classes is particularly difficult for banks that are more exposed to the NIRP. Their breaking of the transmission to loan rates means that they either attract riskier or fewer borrowers, as higher rates either reflect a higher riskpremium or reduced demand, or a combination of both. A reallocation to other assets therefore becomes unattractive earlier for these banks. Since continuing to lend to the central bank or the interbank market is different from positive rate environments a loss-making strategy as well, they therefore prefer to cut outstanding liabilities. Since the ZLB on deposit rates further implies that banks cannot easily adjust their deposit intake, the contraction of the balance sheet is achieved by not rolling over or repurchasing outstanding long-term liabilities. The result is a shorter average maturity of the remaining liabilities and together with the reallocation from liquid assets to mortgages a stronger maturity mismatch. Through these adjustments, adverse NIRP exposure does not ultimately harm banks profitability. It does, however, impair financial stability: Portfolio rebalancing leads to additional credit and market risk, and thus to a reduction in risk-weighted capital. The stronger maturity mismatch, instead, implies more exposure to (traditional measures of) interest rate risk and a reduced liquidity coverage. Lowering policy rates below zero, in other words, does 5 Drechsler et al. (2018) have recently argued that maturity mismatch in banks does not need to imply interest rate risk if the rate-invariant cost of maintaining a deposit franchise is taken into consideration. We are sympathetic to this view and believe that the notion of a deposit franchise is helpful for the interpretation of our results. Here, however, we refer to interest rate risk as it is traditionally measured. 3

4 not only lead to an overall contraction of banks balance sheets, it also requires a careful balancing of any (perceived) benefits with potentially conflicting financial stability objectives. In understanding how our results differ from the transmission of rate cuts above zero, we find banks reluctance to introduce negative deposit rates to play an important role. We do not empirically investigate the origins of this reluctance, but our preferred explanation is banks concern that deposit funding becomes particularly elastic around zero. 6 To protect their deposit franchise (Drechsler et al., 2018) and the relationship with valuable clients, banks seem willing to maintain non-negative deposit rates even if this implies higher funding costs and a greater maturity mismatch. 2. Related Literature With the ZLB on deposit rates as an additional constraint in mind, we can relate our findings to the existing literature on monetary policy transmission through banks. Traditionally, this literature has focused on the bank-lending channel (e.g., Bernanke & Gertler, 1995; Kashyap & Stein, 2000; Jimenez et al., 2012), which predicts a contractionary effect on credit supply from monetary tightening; at least in circumstances in which banks are unable to raise additional deposits, or alternative external financing. We find that the rate cut into negative territory has an expansionary effect on mortgage lending by more NIRP-exposed banks, but a contractionary effect on the size of the balance sheet. More recently, the focus has instead been on lending standards and the riskiness of banks credit supply. Evidence on the risk-taking channel is abundant and generally suggests that a rate cut induces banks to lend more to exante riskier borrowers (e.g., Ioannidou et al., 2014; Jimenez et al., 2014). 7 Bonfim & Soares (2018) provide an updated summary of this literature and distinguish between risk-taking that 6 Recent commentary (Cecchetti & Schoenholtz, 2016; Danthine, 2016) also suggests that deposits rates are constrained by the return on cash. 7 Dell Ariccia et al. (2016) analyse a rate hike and find a symmetric effect: ex-ante risk-taking by banks decreases as short-term rates increase. 4

5 is motivated by a search for yield and incentives for risk-shifting. 8 This distinction provides helpful guidance. When deposit rates are downward flexible e.g., because banks have invested in their deposit franchise and face a relatively inelastic supply of deposit funding (Drechsler et al., 2018) a reduction of the policy rate is likely to affect banks funding costs faster than the return on longer-term assets. This increases the value of banks franchise in the short-term and generates incentives to protect it by taking fewer risks (Dell Ariccia & Marquez, 2013; Dell Ariccia et al., 2014). When the cost of deposit funding is constrained by the ZLB, instead, this franchise value effect is mitigated, and we expect stronger risk-taking incentives. At the same time, a rate cut also reduces the return to safe and liquid short-term assets, creating incentives for risk-taking in pursuit of higher returns. In principle, these incentives exist both under positive and under negative rates, although banks might find it harder under negative rates to expand their lending (since the transmission to loan rates is incomplete). In summary, this suggests that the additional constraint on deposit rates should increase risk-taking incentives, while limiting the feasibility and scale of a credit expansion. This matches our evidence that banks become riskier in response to adverse NIRP-exposure and they do not reallocate all of their liquidity to the mortgage market. The special role of non-negative deposit rates is also a recurrent element in the emerging literature on the transmission of negative policy rates. Heider et al. (forthcoming), in fact, exploit it for identification and compare changes in the lending behaviour of banks with different deposit ratios. They study the effect of the European Central Bank s NIRP specifically on the syndicated loan market and find that banks with higher pre-nirp deposit ratios lend more to riskier borrowers after the policy rate was gradually lowered from zero to -20bps. Our findings are consistent with their observation that the credit expansion is directed towards 8 They also discuss incentives for risk-taking during prolonged periods with low interest rates (e.g., Gambacorta, 2009; Acharya & Naqvi, 2012), which are different from our analysis of banks response to a rate cut. 5

6 riskier borrowers. 9 In addition, however, we also show that banks balance sheets become riskier overall, and that market, liquidity and interest rate risk increase along with credit risk. Eggertson et al. (2017) add a ZLB on deposit rates to their macro-model and also use deposit dependence to study the Swedish case, using bank-level information. 10 Like us, they observe an impaired transmission to banks loan rates and explain it with banks incentives to keep profit margins stable. Bottero et al. (2018), instead, argue that the retail deposit channel is inactive in their sample, because compressed deposit margins are offset with higher fees. 11 They observe an expansionary effect on the supply of credit to Italian firms, and attribute this to a portfolio rebalancing channel, i.e. a search for yield by banks that are confronted with negative interest rates on their central bank reserves. The offsetting effect on fees is also present in our analysis and we too identify a strong rebalancing of banks portfolios towards riskier assets. In addition, however, we also show that the ZLB on deposit rates is not only relevant because it squeezes banks margins, but also because it forces banks that are unable to reallocate all of their reserves to alternative assets to deleverage, and specifically to reduce their longer-term liabilities. 12 This causes the maturity mismatch to increase and exposes banks to additional interest rate risk. It also interferes with banks liquidity coverage ratio and implies that the expansionary effect on loans (mortgages in our case), is accompanied by an effect on the overall balance sheet that is actually contractionary. 13 Finally, we also show that banks profits are not only preserved with fees, but also with an interrupted transmission to loan rates. 9 We conduct our analysis at the bank-level and do not observe borrower characteristics directly. We do, however, observe an relatively stronger growth in mortgages and a more impaired downward transmission to mortgage rates among banks that are more exposed to the NIRP, and after controlling for demand through Google search volumes (see Sections 4 and 5 for more detail). To the extent that relatively higher rates either reflect a higher risk-premium or imply higher default probabilities for borrowers with otherwise identical characteristics, this is consistent with the expansion being directed towards riskier borrowers. 10 Urbschat (2018) also uses banks deposit-dependence for identification to study the response by German banks. 11 Comparing evidence on bank profitability from 27 countries, Lopez et al. (2018) also conclude that high deposit banks do not seem disproportionately vulnerable to negative rates. 12 This effect relates to Demiralp et al. (2017), who find that negative rates seem to induce banks to reduce the intake of wholesale funding. 13 While a contractionary effect of low (not necessarily negative) interest rates has been predicted by Brunnermeier & Koby (2017), their model could not explain the expansion in mortgage lending that we also observe. Since their mechanism relies on banks being capital constrained, it is likely not applicable in our sample of very well capitalized banks. 6

7 3. The Swiss Context 3.1. The Negative Interest Rate Policy (NIRP) Prior to January 2015, monetary policy in Switzerland was conducted primarily via open market operations. The SNB defined upper and lower bounds for the target interbank rate and injected or extracted liquidity from the market to navigate the 3-month CHF LIBOR within these bounds. It paid no interest on central bank reserves. On December 12, 2008, the lower target bound was reduced to zero, while the upper bound was subsequently lowered from 100 bps to 75 bps on March 12, 2009, and to 25 bps on August 03, On December 18, 2014 the SNB then moved the lower bound to -75 bps and announced a return of -25 bps on banks sight deposit account balances for January 22, On January 15, 2015, this rate announcement was then corrected to -75 bps and the target bounds for the LIBOR rate were lowered to -125 bps and -25 bps respectively. Presumably to ensure interbank transmission while limiting the strain on the system at large, the SNB further chose to exempt all central bank reserves below 20 times the minimum reserve requirement for the reporting period 20 October 2014 to 19 November 2014 (static component), minus any increase/plus any decrease in the amount of cash held (dynamic component). 14 Importantly, this ultimately bank-specific exemption was designed to manage aggregate liquidity and not to target specific banks. The Swiss NIRP also differs, for example from the Euro Zone, as it was primarily motivated by the exchange rate. Since 2011, the SNB had continuously acquired assets in foreign currency to moderate pressure on the Swiss Franc, and to defend an exchange rate of 1.2 CHF vis-à-vis the Euro. Despite having communicated a renewed commitment to this exchange rate on December 18, the SNB then decided to unpeg the Franc on January As a consequence, Some commentators have attributed this decision to concerns that a further expansion of the SNB's holdings of foreign-currency assets could at some point cause huge losses and thereby erode its equity and credibility. Others, instead, have posited that even negative equity need not be an issue for a central bank. 7

8 the introduction of the NIRP was accompanied by an appreciation of the Swiss currency from 1.20 CHF/EUR in December 2014 to 1.04 CHF/EUR in April 2015 (Figure 1). For an exportdependent economy, this appreciation was an adverse shock, and exports fell by 3.9% between 2014 Q4 and 2015 Q1. Possibly aided by schemes to subsidize a temporary reduction of working hours and the international price-setting power of many Swiss exporters, however, they quickly recovered and economic growth remained largely unaffected. That monetary policy was largely exogenous to domestic credit growth in Switzerland, that the SNB s decision to charge negative interest on reserves had no precedent in Swiss monetary policy, and that the NIRP was implemented with relatively short notice between December 2014 and January 2015 supports our identification. The simultaneous unpegging of the exchange rate, instead, constitutes a challenge that we address in more detail in Sections 4 and 5. In short, we first remark that economic growth as a proxy for credit demand did not react strongly, as a large share of Swiss exports are highly specialized products with relatively inelastic demand. Second, and more importantly, we focus our analysis on domestically owned commercial banks, with insignificant foreign-currency assets or liabilities, and irrelevant foreign-currency income or costs. 16 Third, although our measure of banks NIRP-exposure is unlikely to be correlated with exposure to the exchange rate, we also explicitly control for proximity to the border, as a proxy for a possible exchange rate dependence of banks clients The Transmission to Interest Rates and Margins Before discussing our identification further, we complete our description of the Swiss case by documenting the evolution of key interest rates from July 2013 to June 2016: Figure 2 illustrates the evolution of the Swiss monetary policy target between July 2013 and June 2016, and the corresponding interest rates for overnight (SARON), 3-month and 12-month interbank 16 This is different for Switzerland-specific wealth management banks, which we therefore analyze separately. 8

9 (LIBOR) loans, as well as federal government bonds with one-year maturity. All short-term rates drop to a level around -75 bps immediately in January The 3-month LIBOR rate and the overnight lending rate stay close to the target, while the return on one-year government bonds is more volatile and initially even below the target. Consistent with a standard yield curve, the return on 12-month interbank loans, instead, is on average higher than the target rate. The main take-away, for our purposes, is the immediate transmission of the negative reserve rate to assets that may serve as close substitutes. 17 The return on longer-term assets, instead, exhibits a weaker reaction to the introduction of the negative policy rate (Figure 3). Government bonds, covered bonds, cantonal bonds, and bank bonds with an 8-year maturity continue an almost uninterrupted downward trend that approaches -75 bps only around June In view of the effect on banks balance sheets, these trends suggest that safer financial assets with longer maturities became relatively more attractive after the policy change. The imperfect pass through to the return on bank bonds, however, which remains positive until June 2016, also suggests an effect on banks funding costs. While we analyse these effects at the bank-level in our regression analysis, Figures 4 through 7 already illustrate the implications for banks profit margins. On the liability side of the balance sheet, the ZLB on deposit rates is evidently at work. The sight deposit rate (Figure 4) approaches 1 bp after the policy change, while banks are only able to earn a return close to -75 bps on reserves or close substitutes (e.g., SARON or 3-month LIBOR). The sight deposit margin consequently drops from -3 bps to -75 bps between December 2014 February This policy-induced drop in interest margins would not occur in positive rate environments in which banks can reduce the return on deposits. 17 Recall, that banks could not hold cash instead of reserves, as any changes in their cash balances were also charged negative interest rates, under the dynamic component of the Swiss NIRP. 18 A notable exception is the return on non-financial corporation (NFC) bonds with the same 8-year maturity, which does not drop further after January 2015 and subsequently approaches 1% from below. 19 During the same period, the margin for demand deposits drops from -17 bps to -99 bps. 9

10 Even in the aggregate, however, we observe that banks do not bear the full cost of decreasing deposit margins; instead, they also disrupt the transmission to loan rates. Figure 5 depicts the margin between the average adjustable rate mortgage (with rates resetting based on the 3- months CHF LIBOR every 3 months, and a contract duration of 3 years) and the 3-month CHF LIBOR rate itself. While the LIBOR rate dropped to -75 bps after January 2015, banks kept the rate received on adjustable rate mortgages largely unchanged, which implied an increase in the corresponding margin from 118 bps in December 2014 to 203 bps in February Over the same period, the margin on 10-year fixed-rate mortgages similarly jumped from 122 bps to 174 bps (Figure 6). 20 Since the average bank in our sample invests about 70% of its balance sheet in mortgages, this suggests an important compensation for squeezed liability margins. In anticipation of our econometric identification, different characteristics of the Swiss case matter. First, the quick succession of events and the lack of a precedent implied that banks could not easily foresee (and adjust to) the implementation of the NIRP. Second, the ultimately bank-specific exemption did not target individual banks. Third, the simultaneous unpegging of the exchange rate challenges our identification of the effects of NIRP-exposure. Fourth, the pass-through to the interbank market remains intact for negative interest rates. Fifth, banks deposit margins are squeezed, but there appears to be an offsetting effect on asset margins. Below, we explain how we integrate these elements into our econometric analysis. 20 Consistent with evidence in Bech & Malkhozov (2016), Figure 6 also shows that the interest rate on 10-year mortgages itself initially increased after January

11 4. Data Our panel comprises symmetric pre- and post-treatment periods and is constructed from comprehensive monthly balance sheet information that FINMA and the SNB jointly collect for supervisory purposes. It begins 18 months before the introduction of the NIRP in Switzerland (July 2013) and ends 18 months thereafter (June 2016). The original sample comprises all [b]anks whose balance sheet total and fiduciary business combined exceed CHF 150 million and whose balance sheet total amounts to at least CHF 100 million. 21 Of the 237 banks that originally satisfy these criteria, we retain 50 domestically-owned commercial and 46 wealth management (WM) banks and focus in our interpretation primarily on the commercial banks. 22 These banks account for about 40% of all banks Swiss assets and have negligible fractions of their assets and liabilities denominated in foreign currencies. 23 With deposit and mortgage ratios of around 70% they also have a relatively transparent business model, which allows us to identify the transmission channels more clearly and contributes to the external validity of our findings. WM banks, instead, are more exposed to foreign currencies and operate a business model that is fairly specific to the legal environment in Switzerland. As the second largest group in our original sample we nonetheless retain them for key regressions, to illustrate that our main insights are robust to a potential confounding from the exchange rate. Our focus on commercial and WM banks notably eliminates the two large universal banks UBS and Credit Suisse, as their Swiss commercial banks did not yet report separate financial statements to the authorities at the time, as well as all trading banks, which do not hold What we call commercial banks are banks that satisfy FINMA s definition of retail banks, which requires them to generate at least 55% of their income from balance-sheet effective activities (primarily net interest income and fees on loans) on average during the three years preceding June 2013 (i.e., the last month before the start of our pre-treatment period; income shares, however, are stable over time so that the group composition would not change if we chose a different selection date). 68 banks from the original sample satisfy this definition and we drop the 18 among them that are foreign-owned (because the relationships with their owners may expose them to the simultaneous exchange rate shock; robustness tests show that results are qualitatively robust if we include the foreign-owned banks). FINMA uses this classification for internal peer group analysis. It is different from the SNB s definition of retail banks, which takes into account banks ownership structure. We believe that a classification that is purely based on banks business models is preferable for our purposes; especially, since the FINMA classification also provides us with a larger sample size. We worked with the SNB s definition in an earlier version of this paper and found our results to be qualitatively robust. Other banks including WM banks generate significantly larger shares of their income from advisory fees and trading income. 23 The pooled sample averages are 2.73% (assets) and 4.38% (liabilities) respectively 11

12 meaningful amounts of deposits, loans or mortgages. 24 Finally, we also have to drop cooperative banks, as they hold reserves at shared clearing banks and not individually at the SNB, and for consistency eliminate all banks that are not present during the 36 months of our baseline period and the 36 months of an earlier (positive rate) period for which we report comparative results in Table Table 1 provides pooled summary statistics for the sample of commercial banks; Table 2 provides statistics for the pre- and post-treatment periods, separately for banks that experience treatment intensity below (Panel A) and at or above (Panel B) the sample median. 26 Summary statistics are provided for different balance sheet items, income, measures of risk-taking, and bank capitalization. Table 1 shows that the average bank in our sample invests 72.78% of total assets in mortgages, 8.49% in uncollateralized loans, and 4.70% in financial assets. Liquid assets amount to 8.34% and are dominated by central bank reserves (7.77% of total assets). On the liability side, deposit funding constitutes the largest fraction (67.59%), followed by bond funding (13.04%). 27 The sample banks hold few assets in foreign currency (2.73%) and raise 95.62% (= 100% %) of their funding in CHF. On average, they exceed their risk-weighted capital requirement by 8.21% of risk-weighted assets and hold a weakly negative net position on the interbank market (-0.86% of total assets). The share of required equity that is attributed to credit risk amounts to 94% and is significantly higher than the share that can be attributed to market (1%) or operational risk (6%). 28 In short, we focus on simple commercial banks that 24 Nucera et al. (2017) suggest that traditional commercial banks, with incomes that are not as diversified as those of large universal banks, are more affected by NIRPs. While this means that we are likely to observe stronger effects than in the full sample, it also means that we are able to observe the effects more clearly. 25 Our results are qualitatively robust to including foreign-owned banks. The earlier period lasts from February 2010 to January 2013 and is centered around a previous rate cut; our results are robust to running the regressions in a sample that is not balanced across the two periods. 26 We do not report figures and summary statistics for groups above and below the exemption threshold, to maintain equal group size. As can be seen in Figure 8, the group of banks with positive levels of exposed reserves is smaller, implying that subsample statistics are less likely to provide reliable values. 27 Medium Term Notes, which we use to illustrate the effect on longer-term borrowing costs, account for about 3.7% of total assets. 28 We use those values that FINMA and SNB collect and report for regulatory purposes. Notice that required equity is calculated before deductions, so that individual fractions (or the sum of different fractions) can exceed 100%. 12

13 deal primarily with local households and firms. They are well capitalized and their main exposure stems from traditional services, such as credit provision and maturity transformation. In Table 2, we consider the change in sample characteristics from the period before January 2015 to the period of and after that month. We observe that the average bank held more liquid assets, as well as fewer claims on other banks and a lower share of cash in foreign currency. It also generated less net interest and fee income, invested in safer portfolios and more strongly exceeded its regulatory capital requirement. Because of the simultaneous exchange rate shock and because banks were differently exposed to the NIRP, however, we cannot necessarily attribute these changes to the effect of negative interest rates. To isolate the marginal effect of adverse NIRP-exposure, we need to compare changes of banks with different degrees of exposure over time. The difference between banks with exposed reserves below (Panel A) and at or above (Panel B) the median provides first insights. We observe an increase in average SNB reserves in both Panels, but a stronger change in Panel B (from 4.06% to 9.14%, compared to a change from 8.30% to 9.59%); at the same time, the net position on the interbank market changes from -0.35% to in Panel A, and from 0.16% to -0.51% in Panel B. For banks below the median, this reflects the existence of an arbitrage opportunity. Their spare capacity allows these banks to deposit more interest-exempt reserves at the central bank while receiving a negative rate on loans from other banks. When exposed reserves are above the median, instead, the group consists of banks with both positive and negative levels of exposed reserves so that effects might cancel out. It is therefore plausible that the lower panel exhibits a qualitatively identical, but weaker change in ratios. The same pattern is present in other key variables as well: below-median banks reduce the share of mortgages on their balance sheet from 74.81% to 72.89%, for example, while above-median banks do not reduce it significantly. 13

14 To capture potentially differential changes in mortgage demand, we complement our supervisory data with Google Trend data for the topic Mortgage. 29 These data are available at the cantonal (state) level and in 7-day intervals; the sample is normalized to 100 by the highest search volume (which comes from the canton Obwalden in the 7 days preceding December 2013). 30 We match these search volumes to banks by using weighted sums across cantons, with weights equal to banks allocation of mortgages across cantons. 31 In this way, we obtain a bank-specific measure that captures differential changes in mortgage demand across cantons, as well as banks potentially differential exposure to these changes (depending, for instance, on their pre-existing branch network, name recognition among potential customers, or local expertise). Figure 7 plots search volumes for all of Switzerland over time. Consistent with the interpretation as a proxy for demand, it exhibits a distinct increase in search volumes after the rate cut in January 2015, i.e. after mortgages became cheaper for households. 32 While summary statistics are indicative, they do not offer conclusive evidence. To gauge the effects of adverse NIRP-exposure more carefully, we proceed with our regression analysis. 29 We use Hypothek, which is German for Mortgage. Topic searches automatically include relevant related search terms; these include in particular identical search terms in English or either of the country s languages. We obtain our data from the Swiss Google Trends site. 30 To match the frequency of our balance sheet data, we sum up the weekly data to obtain monthly frequencies. 31 Note that we do not observe the issuance of new mortgages. We therefore calculate the weight based on the stock and cantonal distribution of mortgages on banks balance sheets (prior to the introduction of the NIRP). 32 In our Online Appendix we provide robustness checks that do not include Google Search volumes; as expected, it turns out that controlling for mortgage demand is not essential for our results. 14

15 5. Empirical Strategy 5.1. Model Our goal is to identify the effect of adverse NIRP-exposure on banks investment, lending and funding choices, and to understand implications for income and risk-taking. To this end, we rely on a difference-in-difference (DiD) design with a continuous rather than binary measure of treatment intensity. The treatment period is characterised by a dummy variable (Postt) that is equal to one from January 2015, and equal to zero before. 33 Our baseline measure of treatment intensity is equal to average SNB reserves prior to December 2014 minus the bankspecific exemption, in percent of a bank's average total assets. 34 We refer to this measure as Exposed Reserves (ERi): ER i = SNB i Reserves SNB Exemption i i Total Assets Denoting a generic dependent variable in period t by Yi,t, our benchmark regression then takes the following form: Y i,t = α + β ER i + γ Post t + δ (ER i Post t ) + ε i,t. (1) The coefficient of interest, δ, captures the effect of adverse NIRP-exposure on Yi,t. A plausible control group in a DiD setup with binary treatment variable would require us to observe banks that are completely unaffected by the SNB s policy. Such banks do not exist. With a continuous treatment variable, instead, we can identify the effect of a more adverse NIRP-exposure, without explicitly observing unaffected banks (i.e., banks with ERi = 0). This identification assumes that the effect we are after is independent of ERi being positive or negative. To see 33 We use quarterly data when we analyse the effect on risk-taking and bi-annual data when we study bank income. Treatment dummies in these cases are equal to one for all quarters (semesters) following and including 2015Q1 (2015H1), and zero before. 34 Rather than using the values of SNB Reserves and Total Assets in December 2014, we take average values over the entire pre-treatment period. While our results are robust across measures, we believe that the average measure is better suited to capture banks actual NIRPexposure (as opposed to a potentially random fluctuation in reserves in December 2014). 15

16 that this is indeed the case, it is instructive to notice that one unit more of positive ERi requires banks to pay 75 bps more interest to the central bank, while a unit more when ERi is negative means that banks lose the opportunity to borrow one unit on the interbank market (at approximately -75 bps) that they could lend to the SNB for free. As long as the transmission to the interbank market is intact (and fast), the cost in both cases is 75 bps. Building on Model (1), we consider several extensions: First, for our main regression, we saturate the model with bank and time fixed effects (FE) to control for time-invariant, bankspecific heterogeneity and for period-specific factors. Next, we allow banks to react differently to their NIRP-exposure if they are headquartered in a canton with a foreign border (Borderi is a dummy variable that is equal to one in this case). This is primarily to capture any potential confounding from the simultaneous effect on the exchange rate, which for the locally active banks that we consider has a primarily geographic dimension. Finally, we add Google search volumes (Mortg.Demandi,t) to control for bank-level variation in the demand for mortgages: 35 Y i,t = δ (ER i Post t ) + η (ER i Border i Post t ) + κ (Border i Post t ) +λ Mortg. Demand i,t + FE i + FE t + u i,t (2) Next, to capture not only the average treatment effect for the post-treatment period, we also estimate monthly effects by interacting our treatment variable with indicators for 35 of the 36 sample months, using the first month of the entire sample as the reference date: 06 Y i,t = α 16 + δ s (ER i FE s ) s= FE i + FE t + e i,t. (3) 35 Since we do not know exactly how Google generates its search volumes for topic searches, we also provide robustness checks that do not feature any measure of mortgage demand (in the Appendix). As expected, it turns out that changes in demand do not covary in any meaningful way with our exposure measure, so that coefficients are largely unchanged. 16

17 The coefficients of interest (δ s) provide evidence of the difference in the intertemporal change in Yi,t between our initial sample date (July 2013) and each subsequent month. Over the 17 pretreatment months this constitutes an implicit placebo test, which should return insignificant interaction effects under the parallel trend assumption. Over the 18 post-treatment months, instead, Model (3) provides additional insights into the evolution of the negative rate effect over time. We estimate our models using ordinary least squares and cluster our standard errors at the bank level (Bertrand et al., 2004) Identification An important identifying assumption is that absent the NIRP the time trends in our outcome variables would be parallel for banks with different levels of exposed reserves. To support this assumption, we first plot the histogram of our treatment variable in Figure 8, which confirms that neither banks nor the SNB targeted any specific cut-off level. In addition, we also provide graphic illustrations of pre- and post-treatment trends in key balance sheet variables (Figures 9, 11 and 13) and of the coefficients on the interaction terms from Model (3) (Figures 10, 12 and 14). Consistent with the assumption that pre-treatment growth rates are not systematically different, Figures 9, 11 and 13 exhibit parallel trends over the first 18 months of our sample. In addition, Figures 10, 12 and 14 suggest that growth differences between more and less exposed banks are insignificant during the pre-treatment, but not the post-treatment, period. Overall, we conclude that this visual evidence supports our identifying assumption. The same evidence also suggests that we do not necessarily need to include additional control variables in our regression, so that we can abstract, among other things, from concerns related to the estimation of dynamic panel models Recall that we do nonetheless include bank and period fixed effects in Model (2) and that we also control for whether a bank is headquartered in a canton with a foreign border, as well as for the time-variation in Google search volumes to capture changes in mortgage demand. 17

18 A second challenge to our identification arises from the removal of the exchange rate peg that occurred simultaneously with the implementation of negative interest rates in Switzerland. The unpegging came as a surprise to financial markets and lead to heavy losses among currency traders betting on a depreciation of the CHF. Their losses transmitted to direct brokers, both foreign and domestic, who had financed the currency traders bets with Lombard Loans. 37 The parallel trend assumption would then be violated if these losses were systematically related to the level of exposed reserves, e.g. because direct brokers have on average lower deposit ratios. In response to this challenge, we do not include direct brokers in our sample and focus entirely on domestically owned commercial banks. We further ensure isolation from exchange rate effects by eliminating from our sample the two big universal banks (UBS and Credit Suisse), which are more heavily engaged in foreign borrowing and lending than retail banks, and by separately controlling for WM banks and for commercial banks that are headquartered in border locations. For commercial banks themselves, the exchange rate shock may initially have been expected to matter insofar as the more export-oriented of the corporate clients may have suffered from reductions in competitiveness. In hindsight, these clients have by and large coped well, aided inter alia by tax-financed schemes to support shorter working hours as well as by the international price setting power of many Swiss exporters. 38 Even if changes in competitiveness did play a role, however, our conclusions would only be affected if the resulting differences in corporate credit demand would be systematically related to the level of banks negative NIRP-exposure See, for example, Swiss central bank moves to negative deposit rate (Financial Times; ) and Swiss franc storm claims scalp of top forex broker (Financial Times; ), where the latter is referring to a UK entity. 38 To avoid bankruptcies or heavy losses of employers, as well as lay-offs in the face of temporarily lower demand for a firm s products, short-term work schemes have employees work only e.g. 50% of regular hours but receive 80% of their full wage, where the difference is paid by the government. For the government this is cheaper than the unemployment benefits due if the person were laid off entirely. See for example 39 For additional robustness, we also ran our models using various alternative treatment definitions (unreported, but available on request). First, we considered the difference between total liquid assets required of each bank in 2015 and the Swiss exemption threshold (scaled by total assets), as an additional measure of banks dependence on, and need for, liquid assets and hence central bank reserves. Second, we also considered the deposit ratio (the fraction of total assets financed through deposits) as in Heider et al. (forthcoming), to capture banks exposure to the NIRP indirectly, i.e. by assuming the limited downward flexibility of their deposit rates. Both alternatives yield consistent results. 18

19 6. Results Having argued that our empirical setup allows us to develop causal insights into banks responses to adverse NIRP-exposure, we now proceed to discuss our results. When the central bank lowers the return on reserves, we first expect banks to reallocate resources to comparable assets. Since changes in cash holdings are also charged negative interest, these are primarily loans to the interbank market and liquid assets in non-chf currencies. Next, we expect a reallocation towards riskier and longer-term assets, and once this seizes to be profitable a reduction in banks liabilities and thus the size of the balance sheet. We further expect the transmission to deposit and lending rates to be interrupted, and so a tendency for banks to cut longer-term financing rather than their deposit intake. In the remainder of this section, we test these expectations empirically and analyse implications for bank profits and financial stability Central Bank Reserves & Liquid Assets Table 3 shows how banks that are more adversely exposed to the NIRP respond to the introduction of negative interest rates by reallocating their central bank reserves towards the Swiss interbank market and towards liquid assets in non-chf currencies. They also show how more exposed banks ultimately and despite this reallocation reduce their liquid asset holdings. The results are consistent with the lower/negative return on central bank reserves and illustrate the transmission to assets that can serve as close substitutes. They are also robust to controlling for changes in mortgage demand and banks exposure to the exchange rate, and present for commercial as well as WM banks. 40 Throughout, our estimates are also robust to model specifications that do not include fixed effects or Google search volumes (Table 3A). 40 We see that WM reduce their central bank reserves and liquid asset holdings, and that they increase their exposure to non-chf liquid assets, although the effects are smaller in magnitude than for commercial banks. We also see that WM banks do not seem to move their liquidity to the interbank market, suggesting that their specific business model provides them with an easier access to alternative assets. Because of these differences, and because we are worried that the observed response by WM banks may be confounded by the simultaneous exchange rate effect and affected by the special relationships with their clients, we will primarily focus on commercial banks. Similar to the results in Table 3, however, the evidence for WM that we present in the remainder of the paper suggests that our main conclusions apply beyond the more focused sample of commercial banks. 19

20 6.2. Loans & Investments Monetary policy transmission, in normal times, would predict an expansionary effect of a rate cut and incentives to invest in riskier and potentially more profitable assets. Consistent with these predictions, Table 4 shows that relatively more NIRP-exposed banks do indeed exhibit relatively higher growth in mortgages after December 2014, and a stronger exposure to financial and non-chf assets. 41 In some contrast to these expansionary dynamics, however, we also observe that banks are either unable or unwilling to reallocate all of their liquid assets towards riskier and longer-term assets. As a consequence, they end up with a contraction of their balance sheet, which because loans, mortgages and investments in financial and foreign currency assets are not proportionally reduced is accompanied by an increase in the balance sheet shares of riskier assets. That is, we find an expansion of banks activities in the mortgage market that is consistent with the risk-taking channel in normal times. In addition, however, we also observe that banks do not allocate all of their excess liquidity towards other assets and in some contrast to the expansionary effects on credit reduce the size of their balance sheet. While this may limit risk-taking in terms of new loans, it increases banks overall exposure to riskier assets. Because growth in corporate lending and investments in financial assets are not reduced proportionally with the reduction in liquid and total assets, these asset categories end up accounting for a larger share of the balance sheet. This not only increases the average riskiness of banks portfolios, but also their average maturity. At the balance sheet level, one could argue that our findings are consistent with the results in Heider et al. (forthcoming), who simultaneously identify a contraction in lending (in balance sheet size, in our case) and an expansion towards riskier borrowers (towards the mortgage 41 The effect on non-chf assets is robust across specifications within the sample of commercial banks, but not when we estimate it in the sample that also includes the WM banks. We therefore interpret this result more cautiously. 20