Who Can Tell Which Banks Will Fail?

Abstract

In this paper, we study the run on the German banking system in the spring of 1931 and ask: in light of a major financial shock, can depositors tell which banks will fail? Are banks better than regular depositors in anticipating which banks are more likely to fail? And, to the extent that banks can tell which other banks will fail, are they willing to provide liquidity to otherwise healthy banks via the interbank market?

Understanding depositor behavior and interbank market dynamics during bank runs is important for designing optimal policy measures to prevent or contain banking crises. Empirical studies of bank runs have made great progress in understanding depositor behavior during runs on individual institutions (see, e.g., Iyer and Puri 2012; Iyer, Puri, and Ryan 2016; Martin, Puri, and Ufier forthcoming). Still, to study whether depositors differentiate between weaker and stronger banks in their withdrawal behavior, it is desirable to observe depositor behavior across heterogeneous banks during the same macroeconomic shock. Likewise, the ability of the interbank market to provide liquidity is best tested during a major financial shock, when the entire banking system is stressed. However, the empirical analysis of systemwide bank runs is typically constrained as government interventions distort depositor withdrawal decisions (Iyer et al. 2019) and make systemwide runs infrequent in modern times (Baron, Verner, and Xiong 2020).

In this paper, we use newly digitized historical micro-level data to study the run on the German banking system from May through July 1931—one of the largest bank runs in economic history and a key event of the Great Depression (see, e.g., Kindleberger 1973; James 1984). There are three key advantages in studying this historical episode. First, the German system was lightly regulated with no capital or liquidity requirements and no deposit insurance. Thus, all types of depositors could plausibly expect to realize losses in the case of a bank failure. Second, we can obtain and digitize granular and comprehensive bank balance sheets at a monthly frequency, allowing us to study run dynamics in more detail than previous work. Third, the systemwide nature of the bank run provides an empirical laboratory with a large number of failing banks—15 of around 120 banks in our sample failed during the crisis—exposed to the same macroeconomic environment. Taken together, this allows us to establish whether and which depositors anticipate which banks will fail.

Our findings are simple but striking. We find that deposits decline by around 20% over the two months from the start of the run to the end, when the government declared a bank holiday. There is no difference in total deposit outflows between failing and surviving banks. This implies that, on average, depositors seem unable to successfully identify which banks will fail. However, the average obscures the distinct behaviors of different types of depositors. We find that outflows of regular deposits—retail and nonfinancial wholesale deposits—are the same across failing and surviving banks. In contrast, interbank deposits decline mostly for failing banks. Banks themselves withdraw almost exclusively from banks that end up failing, but not from banks that end up surviving. Thus, failing banks start to lose access to the interbank market throughout the run and by the end of the run, they are essentially shut out of the interbank market. This pattern, in turn, implies that interbank deposit flows during the run can be used to predict which banks will fail. The area under the receiver operating characteristic curve (AUC), a common measure of diagnostic performance for a binary classifier, can exceed 80% using a simple regression model in which failure is predicted by interbank deposit flows.

We further study interbank dynamics during the run in more detail. We show that in normal times—before the run—banks subject to deposit outflows respond by borrowing from banks subject to deposit inflows, exactly as predicted by theories of optimal risk-sharing via interbank markets (Allen and Gale 2000). While aggregate deposit funding of the banking system contracts during the run, we document that there is also substantial heterogeneity in deposit flows across banks and some banks obtain deposit inflows. Notably, we find that banks that experience deposit inflows do not hoard liquid funds as suggested by theories of interbank market freezes (see, e.g., Caballero and Krishnamurthy 2008; Allen, Carletti, and Gale 2009). In contrast, they continue to intermediate funds to banks that are subject to deposit outflows. However, only surviving, not failing, banks subject to deposit outflows are able to borrow from other banks. Thus, while the interbank market collapses for failing banks, it remains functioning for surviving banks even in times of severe financial distress.

At least two plausible and not mutually exclusive explanations can help explain our main findings. Either banks are well informed about which banks will fail and withdraw from failing banks to protect themselves. Or, alternatively, failing banks fail because they are losing access to interbank funding. While both explanations are a priori reasonable, we argue that the former explanation is more plausible than the latter for the following two reasons. First, we find that total deposit funding (both regular and interbank deposit funding) contracts by about the same percentage share for failing and surviving banks. This equal outflow of total funding is possible despite the striking differences in interbank deposits flows because interbank deposits represent on average only a small share of total deposit funding. Theories of bank runs and fire sales, however, imply that the primary determinant of whether a bank becomes illiquid first and then insolvent is the total shortfall of funding (see, e.g., Diamond and Dybvig 1983; Goldstein and Pauzner 2005) rather than the composition of funding. Consequently, the scope for the shortfall of interbank funding alone to cause failure is reduced. Second, we show that our findings are unchanged when restricting our sample to banks that rely very little on interbank funding to begin with and for which interbank deposit outflows cannot plausibly be the immediate cause of failure. Altogether, we argue that losing access to funding from other banks alone is unlikely to have caused the failure, similar to findings by Perignon, Thesmar, and Vuillemey (2018).

It is important to highlight that our empirical approach does not allow us to determine between the causes of bank failures more generally. In particular, we cannot distinguish between whether withdrawals are primarily caused by the prospect of fundamental insolvency (as in a solvency run) or whether overall withdrawals are the primary cause of default (as in a panic-based run).¹ Our approach instead focuses on the information structure: which depositors understand which banks will ultimately fail? The differential response between regular and interbank deposits allows us to conclude that domestic banks are well informed about which banks will fail due to the crisis, while regular deposit flows are uninformative about failure. However, we cannot identify what information banks are acting on. Our findings allow for different possibilities. Banks may have information about a specific bank’s solvency that is independent of deposit withdrawals. Or, banks may have information about which banks are more likely to be perceived as fragile by regular depositors and fail as a consequence of their withdrawals.

We also provide a set of additional important empirical findings that support the interpretation that banks are well-informed and able to tell which banks will fail. First, we show that interbank funding declines more at banks that relied on foreign-currency-denominated deposits. In light of concerns about Germany’s ability to maintain the Gold Standard (see e.g., Ferguson and Temin 2003), the incentives to withdraw foreign-currency-denominated deposits during the run were likely to be stronger than for regular deposits. Importantly, whether a bank had such an exposure to this more risky type of funding was not public information but only reported in confidential filings at the Reichsbank. Thus, given that we find that interbank funding declines more at banks that relied on foreign-currency-denominated deposits reinforces the idea that bank had a good sense of which banks were at risk during the run.

Second, we study the extent to which the stock market identifies failing banks for a small subset of our original sample. We find that the stock prices of failing banks start to decline in the first month of the run and fall by around 25% more than those of surviving banks over the course of the crisis. These stock returns can be predicted by interbank flows in the early phase of the run.

Finally, we document that failing banks are at times subject to demand deposit inflows during the run, but these inflows are mirrored by outflows of time deposits. Thus, depositors are more likely to take a cautious stance in failing banks and convert deposits with longer maturity into those that can be withdrawn easily, indicative of maturity shortening (Brunnermeier and Oehmke 2013). This pattern of outflows of time deposits rather than demand deposits also confirms the finding by Martin, Puri, and Ufier (forthcoming) that flows of term deposits in failing banks tend to be more sensitive to bank risk than demandable transaction deposits.

Our findings have important policy implications. Our conclusion that banks are well-informed about which banks will fail suggests that the interbank market can be effective in providing liquidity to otherwise financially sound banks even in the midst of large financial shocks. Thus, our findings support the view that the interbank market can be surprisingly resilient (Afonso, Kovner, and Schoar 2011). Further, the existence of a functioning interbank market can be valuable beyond standard risk-sharing rationales. Central bank actions that make interbank markets redundant—such as an abundant reserves regime—should consider the cost of losing the valuable information contained in the interbank market and the potential to provide discipline through interbank flows.

Our findings also lend support to the view that banking crises are not just sudden and unpredictable events. Existing research shows that crises typically follow credit booms (see, e.g., Schularick and Taylor 2012; Baron, Verner, and Xiong 2020) and whether a crisis will happen can thus in part be predicted (Greenwood et al. 2022). We find that interbank deposit growth—unlike total deposit funding growth—contains information that allows us to predict which banks will fail. Given that banks themselves seem to know “where the bodies are buried,” our findings suggests that, conditional on being in a banking crisis, it is possible to predict which banks are more likely to fail.

Taken together, our evidence also provides an empirical reconciliation between the two standard rationales for the existence of short-term debt. On the one hand, short-term debt can be a means to provide valuable liquidity services to depositors (see, e.g., Diamond and Dybvig 1983; Gorton and Pennachi 1990). On the other hand, it can also be an instrument that allows depositors to discipline bank behavior (see, e.g., Calomiris and Kahn 1991; Diamond and Rajan 2000, 2001). The two types of rationales differ considerably in how they view depositors and may thus be in conflict (Admati and Hellwig 2013): while the former regards them as liquidity demanders, the latter considers them to be informed providers of discipline. While we do not test either theory directly, our finding that different types of depositors vary widely in how informed they are offers a resolution: interbank depositors get rewarded for being informed and attentive by withdrawing first from failing banks—comparable to the informed depositors in the model from Calomiris and Kahn (1991)—and are thus able to discipline other banks. Regular deposit flows, in contrast, are not informative about failure. Hence, regular depositors may not be well positioned to provide discipline but rather value liquidity benefits associated with holding deposits.

1 Literature

Our paper contributes to a rich literature on bank runs and banking crises. Seminal work by Diamond and Dybvig (1983) shows under which conditions demand deposit contracts can insure depositors against idiosyncratic liquidity risk, but also how demand deposit contracts set the stage for coordination failures and self-fulfilling runs.² A complementary rationale for the existence of short-term funding of banks and bank runs is provided by Calomiris and Kahn (1991) and Diamond and Rajan (2000, 2001), who argue that demand deposit contracts are an instrument to discipline the behavior of the bank’s management. In this line of argument, bank runs are equilibrium outcomes as a response to information about nondiligent behavior of bankers as well as the aggregate state of the economy.³

While the theoretical literature on systemwide bank runs has made great progress, the empirical study of the subject is often constrained by the lack of adequate settings and data. Either government intervenes before a systemwide bank run fully plays out (Baron, Verner, and Xiong 2020), or, when it does occur, data are only available at a low frequency. Thus, existing empirical work focuses on either bank runs in settings in which deposit insurance or related government interventions affect depositors’ incentives (Iyer et al. 2019), or the analysis is concerned with banking crises from prior to or during the Great Depression, when data are typically not available at a high frequency. The key advantage of our setting is that our granular monthly data allow us to analyze depositors’ behavior—including interbank depositors—during a major financial shock in a setting in which depositors of all types had to expect to realize losses if their bank failed.

Evidence on the importance on the heterogeneity of depositor behavior during bank runs in contemporary settings is provided by Iyer and Puri (2012), Iyer, Puri, and Ryan (2016); Iyer et al. (2019), Martin, Puri, and Ufier (forthcoming), and Artavanis et al. (2022). Iyer and Puri (2012) establish that depositors that have stronger ties to the banks, either socially or financially, are less likely to withdraw. Iyer, Puri, and Ryan (2016) provide evidence that sophisticated and uninsured depositors are more sensitive to solvency risk. Martin, Puri, and Ufier (forthcoming) emphasize the importance of gross funding inflows in troubled banks and show that, prior to bank failure, outflows of uninsured deposits are offset with inflows of insured deposits. Moreover, they establish that time deposits tend to be less “sticky” than demand deposits, possibly reflecting depositor sophistication and the forward-looking nature of term deposits. Artavanis et al. (2022) use deposit level to study the case of a slow run on a Greek bank in 2015. All the above settings, while providing valuable information about depositor behaviour, focus on runs on individual institutions. Our paper complements these important papers by studying the behavior of depositors in a lightly regulated banking system that featured a systemwide run.⁴

By studying the dynamics in the interbank market in a systemwide run, our work also directly relates to empirical studies of interbank market dynamics. Iyer and Peydró (2011) test financial contagion due to interbank linkages and Iyer et al. (2014) study the real effects of interbank market distress. Similarly, Craig and Ma (2022) study systemic risk in the contemporary German interbank market. Afonso, Kovner, and Schoar (2011) study the interbank market in the United States during the 2007–2009 financial crisis. Like Afonso, Kovner, and Schoar (2011) we find evidence that the interbank market continues to function during a major financial shock. Banks do not hoard liquidity but only stop lending to failing banks. Surviving banks continue to be able to borrow. Our findings are also in line with evidence from Perignon, Thesmar, and Vuillemey (2018), who study wholesale funding dry-ups for European banks around the European Debt Crisis and stress the role of informed and uninformed investors.

Other papers that have studied systemwide banking panics in lightly regulated banking systems are largely confined to historical episodes in which data are available at a much lower frequency or aggregate rather than bank-level data. In classic studies, Gorton (1988) and Calomiris and Gorton (1991) show that systemwide banking panics during the National Banking Era typically occurred when economic activity peaked. Calomiris and Mason (1997) also provides an account of the bank failures in Chicago during 1932 and supports the view that weaker banks were more likely to fail. Saunders and Wilson (1996) and Calomiris and Mason (2003b) study causes of bank failures during the Great Depression using biannual data and find evidence that the causes of the bank runs were related to fundamental solvency concerns. Calomiris and Mason (2003a) study the real effects associated with the Great Depression and Frydman, Hilt, and Zhou (2015) study the real effects of the Panic of 1907. Kelly and Ó Gráda (2000) and Ó Gráda and White (2003) study depositor runs using depositor-level data in a case study of a New York bank during the Panics of 1854 and 1857. Closest to our work is Ó Gráda and White (2003) who find that less sophisticated depositors withdrew during the nonsystemic run of 1854, but more educated depositors were withdrawing their deposits during the systemwide crisis of 1857. However, their study is confined to studying a single bank while we can study the entire banking system, including the interbank dynamics.

An important exception in the empirical bank run literature is the paper by Schumacher (2000), which studies the cross-sectional variation in bank stability during a banking panic that took place in Argentina in 1995 following the Mexican “Tequila shock.” Importantly, Argentina at the time had no deposit insurance scheme and no wider safety net. However, the crucial advantage of our empirical approach is that we have information on the different types of deposit flows and thus the richness of our data allows us to test for heterogeneity in depositor information explicitly.

Another important exception is Mitchener and Richardson (2019), who use weekly city-level data to show how interbank market crashes amplified the credit crunch during the Great Depression. Besides the fact that our paper allows to provide bank-level as opposed to city-level evidence at a relatively high frequency, note that the German banking system in the 1930s featured geographically diversified banks. The U.S. banking system before and during the Great Depression, in contrast, featured mostly local banking markets and thus a hierarchical interbanking system, with very different roles for country banks and reserve city banks. Thus, in contrast to Mitchener and Richardson (2019), our study provides evidence from a setting without deposit insurance in which the overall structure of the banking system is more similar to that of contemporary banking systems than the historical U.S. banking system is. In contrast to Mitchener and Richardson (2019), we find that the interbank market need not necessarily be a source of financial contagion during times of distress but can provide liquidity.

Further, our paper also relates to a set of papers that evaluate the role of deposit insurance on market discipline across various empirical settings. For instance, Anderson, Richardson, and Yang (2023) study the effect of the creation of the Federal Deposit Insurance Corporation (FDIC) and find that while deposit insurance reduced monitoring, it did not entirely eliminate it. Calomiris and Jaremski (2019) show that the introduction of deposit insurance schemes in U.S. states during the early 20th century removed market discipline. Karas, Pyle, and Schoors (2013) study the effects of deposit insurance in Russia during the 1990s and also find declines in market discipline. Peria and Schmukler (2001) study the cases of Argentina, Chile, and Mexico during the 1980s and 1990s and find that depositors discipline banks by withdrawing deposits and by requiring higher interest rates but only limited effects of deposit insurance on market discipline. Our paper, by utilizing a unique historical setting and novel data, complements these papers by providing additional evidence that banks themselves are the most informed depositors and thus best positioned to provide discipline. However, it is important to highlight that while our paper provides some insights on how depositors can behave in absence of deposit insurance, unlike the papers mentioned above, it does not allow to test the effectiveness of depositor discipline itself and cannot directly speak to the issue of moral hazard stemming from deposit insurance.

Finally, our paper also contributes to the literature on the Great Depression in Germany. Papers studying the more general role of the economic and political crisis and the rise of political extremism and the Nazi party are provided by Galofré-Vilà et al. (2021) and Doerr et al. (2022), with the latter focusing on the impact of the failure of the Danatbank—the second largest bank at the time and discussed in more detail below—on the rise of fascism. Two important accounts of the crisis episode, interpreting it primarily as a banking crisis, are provided by Born (1967) and James (1984). In contrast, Temin (1971, 2008) and Ferguson and Temin (2003) emphasize the actions of the German government. Kindleberger (1973) and Eichengreen (1995) emphasize the international dimensions of both, currency and banking crisis. Schnabel (2004) emphasized the role of “too big to fail” guarantees for the large Berlin banks that may have led to excessive risk-taking. Moreover, Schnabel (2009) also studies the effect of liquidity and government guarantees on bank stability during the crisis.

2 Data and Setting

Our main data source is a set of detailed monthly bank balance sheets that were collected by the central bank—henceforth “Reichsbank”—and made publicly available via the contemporary newspaper Deutscher Staats- und Preussischer Reichsanzeiger. Digital versions of the newspaper are made available by the University of Mannheim (Kling 2016) and complemented by hand-collected data from the archives of the Reichsbank held at the Federal German Archives (“Bundesarchiv”) in Berlin and Koblenz.⁵

Bank balance sheets for large commercial banks are available monthly between 1928 and 1933, excluding balance sheets as of December and January. Banks that report their balance sheets to the Reichsbank include the very largest banks with a nationwide branch network—so-called “Berlin banks”—as well as the smaller regional credit banks with a local or no branch network. Further, our sample also includes clearing banks and brokers for savings banks (“Girozentralen”) and publicly owned banks (“Landesbanken”).

Note that our data do not include information on local savings banks, mortgage banks, or private investment banks and brokers. Altogether, our data cover an average of more than 120 banks per month which constitute more than 50% of the entire German banking sector’s assets (Schnabel 2004) and more than 75% of total C&I lending. Importantly, the banks in the parts of the banking system for which no micro data are available are not offering the exact same services and have slightly different business models. For instance, local savings banks were largely financed by retail depositors and invested in mortgages or intermediated funds to the Landesbanken and Girozentralen (which are part of our sample). Private banks tended to be investment banks or brokers. The Berlin banks and regional credit banks raised deposit funding from both retail and wholesale depositors. Note that the sample selection implies that the banks in our sample are likely to be financed by depositors that are more sophisticated and larger than the average depositor.

The data are fairly granular with more than 70 balance-sheet items reported. Among other things, the data distinguish between domestic interbank and regular deposits, demand and time deposits, loans and covered bonds, as well as high- and low-quality liquid assets.⁶Table 1 gives an overview of the observable characteristics in our sample. The table reports the average of assets (panel A) and liability (panel B) items as a share of total assets and liabilities for 126 banks that report balance sheets between February and April 1931. We report the respective shares as averages for the entire banking sector as well as for the four different types of banks mentioned above. In the columns far left of panel A, we also report the average bank size and number of banks in each category.⁷

The largest banks in our sample are the six Berlin banks (four of which had nation-wide branch systems) with an average balance-sheet size of around two billion Reichsmarks (RM). In contrast, regional credit banks are much smaller, with an average balance-sheet size of only 50 million RM. Girozentralen are considerably larger than the regional banks but also smaller than the Berlin banks, with an average asset balance of 300 million and 240 million RM, respectively.

The average bank in our sample has around 73% of its funds invested in illiquid assets. Illiquid assets, in turn consist of 53% commercial and industrial loans and 15% covered bonds, such as mortgages and municipal bonds.⁸ Around 26% of banks’ funds are invested in liquid assets. Liquid assets can broadly be categorized into liquid assets of higher and lower quality as well as interbank claims. High-quality liquid assets consist of cash, reserves, or government bonds. Lower-quality liquid assets are bills of exchange from private nonfinancial firms. Around 5% of assets are in high-quality liquid assets and around 12% in low-quality, and 9% in interbank claims. Note that for interbank claims, we can also distinguish between those due within 7 days. On average around 45% of interbank claims are short term.⁹

On the liability side, we can distinguish between different types of deposits. The balance sheet splits deposits into three different categories: deposits from domestic banks, regular deposits (which combines retail and nonfinancial wholesale deposits—including those denominated in foreign currency) and other types. Further, the reporting form distinguishes between those deposits that are due within 7 days (which we refer to as demand deposits) and those with a specified maturity of more than 7 days (which we refer to as time deposits). Note that the distinction by maturity is only applied for the sum of domestic interbank deposits and regular deposits. That is, we cannot distinguish by maturity within domestic interbank and regular deposits. Given that interbank claims have to be equal to interbank deposits in the aggregate, it’s fair to assume that a little less than half of interbank deposits are due within 7 days. Interbank lending was typically unsecured.

On average, banks finance 66% of their assets with deposits of which the majority are regular deposits: 51% of assets are financed by regular deposits and only 12% by domestic interbank deposits. Further, one can observe that equity finance is relatively higher at the smaller regional banks (23%), since these banks are not diversified geographically. In contrast, equity finance is lowest at the Berlin banks and Landesbanken (7% and 5%, respectively).

Note that there is considerable variation across the different types of banks, with Girozentralen and Landesbanken having a different business model than the large Berlin banks and the smaller regional banks. Berlin banks and regional banks were largely in the business of financing nonfinancial firms, in part by discounting their trade credit claims. In contrast, Girozentralen and Landesbanken intermediated investments from local savings banks, investing in mortgages and municipal bonds. Hence, interbank deposits are much more common at the Girozentralen and the Landesbanken.¹⁰ The main focus of our analysis is on regional banks and Berlin banks, which resemble a textbook banking business model of financing loans with deposits. All cross-sectional and panel estimations thus include bank-type or bank-type-time fixed effects, respectively. Further, throughout our analysis we also show that all main findings are not bank-type dependent and hold when using a sample of only the smaller regional banks.

Further, we also obtain data that was confidentially filed with the Reichsbank—and thus not publicly available during the run—and allow us to approximate the use of deposits denominated in foreign currencies. Information on the exposure to deposits denominated in foreign currency is crucial as many observers stress the role of deposits denominated in foreign currency in the run (see, e.g., Schnabel 2004; Temin 2008).¹¹ The information on the use of foreign-currency-denominated deposits available to us is limited to the summers of 1929 and 1930. We use it as a proxy for which banks issue foreign deposits and make foreign investments. Specifically, we approximate foreign-currency-denominated deposits by multiplying the maximum share of those deposits observed between 1929 and 1930 with the amount of overall deposits net of domestic interbank deposits. Table 1 shows that deposits denominated in foreign currency are highly concentrated in the large Berlin banks and a few of the larger regional banks. Foreign funding is essentially nonexistent in the smaller regional banks and uncommon for Girozentralen and Landesbanken.

We also use the Reichsbank records as well as information from Born (1967) and Schnabel (2009) to determine which banks fail, which are merged, and which are actively bailed out by the state, see Table A.1 in the Internet Appendix. We identify 15 banks that fail, and another 7 banks that did not fail but were distressed and received some form of government aid or were subjected to a distressed merger. We focus on contrasting deposit flows in failing versus surviving banks but also show that our results are robust to using a more general version of distressed banks.

We supplement the balance sheet data of banks with additional data sources. We hand-collect data on daily stock prices for the banks that were traded from the Monatskursblatt, published by the Berliner Börsenpapiere for 1931. These are monthly publications that contain daily stock- and bond-price information for stocks traded on the Berlin Stock Exchange. It tracks closing trading prices for each day of the month. Not all the banks in our sample are publicly traded or listed on the Berlin exchange. We are able to match daily stock prices with balance sheet information from 28 banks covered in the Reichsanzeiger. We also hand-collect the weekly balance sheets of the Reichsbank for the entire year of 1931. The balance sheet includes information on the amount of notes outstanding as well as the amount of gold held by the Reichsbank in its vaults which we use in Appendix A.3 in the Internet Appendix to provide more background on the Reichsbank’s actions. We also follow Doerr et al. (2022) and measure firm-bank relationships based on information on the banks that paid out a firm’s dividends (Zahlstellen). This information is reported in the investor manual Saling’s Börsen-Jahrbuch and allow to proxy for whether a bank was connected to the nonfinancial firm “Nordwolle,” which failed just before the height of the crisis and is closely related to the failure of Danatbank.

A key advantage in studying the German Crisis of 1931 is that the bank run took place in a banking system that had very little government interventions. Specifically, there was no capital or liquidity regulation and most importantly no deposit insurance. The German banking system was following a German tradition of “self-regulation” in which the only interventions came from the Reichsbank with its only real power stemming from the ability to refuse to act as a lender of last resort (James 1984).¹² Given our research objective, it is important to establish that depositors—regular depositors and interbank depositors alike—had a reason to believe that they would realize losses on their deposits in case of a bank failure. Thus, in Internet Appendix A.1, we provide evidence that bank failures were quite common before the run in 1931 and in those bank failures, depositors typically realized losses. Thus, depositors of any type had reasons to expect that they would realize losses if their respective bank failed.

3 The German Crisis of 1931

In this section, we first provide a brief discussion of the key events of the crisis and then discuss how the run presents itself in our data. Note that we keep the description of the historical events to a minimum and provide a more detailed description of the crisis and its circumstances in Internet Appendix A.2 and refer to existing work that provides detailed narratives of the crisis (Born 1967; James 1984; Schnabel 2004).

The run on the German banking system in 1931 was preceded by a 2 year period of contraction in output and employment, deflation, and a high degree of political uncertainty. The run on the German banking system can be broadly categorized into three phases from early May 1931 through July 1931. In the first phase in May 1931, the interbank market shows signs of distress. The distress started when the failure of the largest Austrian bank, the “Creditanstalt”, was announced on May 11, 1931 (Born 1967; Kindleberger 1973; James 1984). German banks were not contractually linked to the Creditanstalt. Although bank failures were quite common in interwar Germany, as discussed in Internet Appendix A.1, the failure of the Creditanstalt was remarkable. It was the largest Austrian bank and its failure was widely unanticipated by the public. Thus, the failure of the Creditanstalt is sometimes interpreted as a “Minsky moment” that triggered a banking crisis without revealing any additional information about the state of the German banking system (James 1984).

The second phase of the run coincides with the German government’s announcement on June 6 that it was unwilling/unable to continue reparations payments, thus raising doubts about Germany’s ability to maintain the Gold Standard. During this second phase in June and early July, withdrawals continued with varying intensity. For instance, withdrawals picked up when a major creditor of Danatbank called “Nordwolle” announced heavy losses, leading to speculation about Danatbank’s imminent failure. Similarly, withdrawals started to slow noticeably after the announcement of the “Hoover moratorium” on June 19, a suggestion by U.S. President Herbert Hoover to pause all war-related debt payments for 1 year. However, when French opposition to the arrangement became clear throughout the end of June, withdrawals intensified again.

The third and final phase of the crisis was reached on July 10-13, when the Reichsbank’s gold reserves had fallen below the legally mandated 40% gold-to-notes coverage ratio. In anticipation, the Reichsbank had started a last attempt to obtain emergency loans from the Bank of England and the Banque de France.¹³ When this attempt was unsuccessful, the Reichsbank decided to further increase the discount rate and tighten its already restricted liquidity provision to the banks. This rendered the Danatbank illiquid, as it had already discounted all of the assets that qualified for Reichsbank purchases. As an additional last-minute attempt to merge Danatbank and Deutsche Bank failed, Danatbank had announced it would not open its branches again on Monday, July 13.

Following the failure of Danatbank, retail depositors started a full-blown panic, queuing at most banks to withdraw their funds. This triggered the illiquidity of “Dresdner Bank”, at the time the third largest bank, on July 14. The then full-blown run led the government to intervene by imposing a 2 day bank holiday, which was followed by an effective suspension of convertibility lasting throughout August¹⁴ and the introduction of capital controls. Further, the government ensured that illiquid banks would have access to the liquidity provision of the Reichsbank and set up a conduit that allowed banks to make their securities eligible for Reichsbank purchases. While deposits continued to contract until the end of 1931, albeit at a slower pace, the financial crisis was over when the government restructured the largest banks in spring 1932.¹⁵

How does the run present itself in the data? Figure 1 depicts the aggregate flows of a selected set of bank assets and liabilities relative to the previous month. The shaded areas depict month-to-month flows in assets, while the colored lines depict flows in liabilities. Aggregate deposits contract by around 500mil RM from April to May. From May to June as well as from June to July, the aggregate deposit outflow almost triples, to a little less than 1,500mil RM per month, representing an outflow of around 8% of the precrisis level of total deposits for 2 consecutive months. Overall, deposits fall by around 5bn RM between March and November 1931, around 25% of the precrisis level.¹⁶

Fig. 1

Aggregate dynamics of assets and liabilities

The figure shows the month-on-month aggregate changes of key balance sheet components during the crisis in 1931 reported in the Deutscher Staats- und Preussischer Reichsanzeiger. Lines depict liabilities, such as domestic interbank borrowing (in blue) and the sum of all deposits (in dark red). Solid areas represent key assets such as illiquid assets (primarily loans and covered bonds in green), liquid assets (in yellow), and interbank lending (in blue). The first vertical line, on May 11, 1931, represents the date of the failure of the Austrian Creditanstalt. The second vertical line, on July 13, 1931, represents the failure of Danatbank and the start of the banking holiday.

Open in new tabDownload slide

Figure 1 reveals that during the first month of the bank run—in the immediate aftermath of the failure of the Creditanstalt—the deposit outflow is largely accounted for by a contraction in domestic interbank deposits, which is accompanied by an equal fall in interbank claims.¹⁷ The first month of the run is therefore largely a run of banks on banks. Moreover, interbank lending and borrowing continue to contract steadily throughout the crisis.

Between May and July, deposit outflows intensify. In addition to the contraction of interbank deposits in May, more than 1bn RM of regular deposits such as retail and nonfinancial wholesale deposits are withdrawn from the banking system in both June and July. Banks meet these withdrawals largely by reducing their holdings of liquid securities (red shaded area), discounting them at the Reichsbank as discussed in more detail in Internet Appendix A.3. Illiquid assets such as loans and mortgages are also contracting (blue-shaded area) but contract much more slowly than the securities portfolio during the bank run itself.

Figure 2 shows results for the dynamics of deposits in panel A and assets in panel B. In line with the dynamics of aggregate deposits, initially only interbank deposits contract. We estimate that interbank lending falls on average by around 10% in May. The interbank market continues to collapse throughout the run and on average, interbank deposits decline by more than 30% by July. Further, the effect is also statistically significant in every month after April 1931 (see Figures A.6 and A.7 in the Internet Appendix for point estimates with confidence bands).

Fig. 2

Deposit and asset dynamics during Spring 1931

The above figures display the sequence of coefficients |$ \{\beta_{t} \} $| that results from estimating the model:

$$\displaystyle y_{b t} = \gamma_{b} + \sum \limits_{t \neq \text{April} \,1931} \beta_{t} \times \gamma_{t} + \epsilon_{b t},$$

where y_bt is the natural logarithm of one plus either a bank b’s deposits (total, interbank, demand, and time deposits) in panel A or bank b’s assets (liquid assets net of interbank claims by quality, interbank claims, and credit) in panel B. We multiply y_bt with 100 to convert the coefficients into percentage points. γ_b is a set of bank fixed effects. We weight each observation by bank size and normalize the set of time-varying coefficients |$ \{\beta_{t} \} $| to April 1931. Data are from the Deutscher Staats- und Preussischer Reichsanzeiger. Figures A.6 and A.7, in the Internet Appendix, show the estimates with confidence bands.

Open in new tabDownload slide

In contrast to interbank deposits, regular deposits are stable throughout May. However, regular deposits start to contract in June, when they fall around 10% and by July they have contracted by more than 15%, with the effect again being statistically significant. Thus, while the run starts out as a run of banks on banks in May, it turns into a broader run that includes withdrawals by other depositors in June and July. Given that interbank deposits are a relatively small share of overall deposits, there is no decline in total deposits throughout May. However, given the relative importance of regular deposits for total deposits, total deposits also start to contract together with regular deposits in June and July. We estimate that the average bank loses around 15% of its deposits by the end of June and 20% by the end of July after the breakdown of the banking system and the start of the banking holiday.

As discussed above, our data allow us to distinguish between standard demand deposits with a maturity of less than 7 days and time deposits with a maturity between 7 days and more. Note though that time deposits could also be withdrawn at any time, although these withdrawals would be subject to a penalty.

In panel A of Figure 2 we estimate that on average, demand deposits are stable in the first two months of the run and actually increase in the last months. Hence, the drop in overall deposits is entirely driven by an outflow in time deposits which decline by around 55% by July. The fact that demand deposits do not fall throughout the run is a striking finding as all deposits including demand deposits are uninsured. The finding is thus seemingly incongruent with standard bank run theories, which predict that uninsured debt claims with the shortest maturity are most likely to be withdrawn first in a crisis.

While demand deposits were most commonly held by retail depositors such as households, time deposits were more akin to modern-day wholesale funding as they carried considerably larger interest payments and tended to be held by corporations and wealthy investors. Hence, the stability of demand deposits can be rationalized by the fact that the latter type of depositor is arguably more sophisticated and more attentive.¹⁸ Retail depositors started to withdraw across the board only when Danatbank declared bankruptcy, marking the third and final phase of the run (Born 1967). However, the attempted withdrawals were immediately stopped by the bank holiday and thus not reflected in the data. The fact that retail depositors do not withdraw until the end of the Reichsbank liquidity support and the failure of Danatbank is in line with retail depositors having higher information acquisition costs and thus only paying attention in later stages of the run (He and Manela 2016). The finding is also reminiscent of the difference in the behavior of retail and institutional investors in money market funds after the Lehman failure, when retail investors were much less likely to react to the shock (see Schmidt, Timmermann, and Wermers 2016). Further, that depositors start a physical bank run once they learn that Danatbank defaulting suggests that they then revise their expectations massively once the aggregate liquidity shortage (Diamond and Rajan 2005) becomes salient in the light of the failure of Danatbank.

Figure A.8 in the Internet Appendix indeed shows that demand deposits, are in aggregate increasing slightly during the crisis, suggesting that some time deposits are being converted to shorter maturity demand deposits. The pattern of increasing rather than decreasing demand deposits can be rationalized by maturity shortening in time deposits (Brunnermeier and Oehmke 2013) in which worried depositors—to the extent that they are not leaving the banking system—convert time deposits into demand deposits. The finding is also in line with the notion that term deposits in failing banks tend to be more sensitive to bank risk than demand deposits (Martin, Puri, and Ufier forthcoming).¹⁹

Mirroring the outflows in deposits, panel B of Figure 2 provides information on the dynamics of bank assets during the run. In line with the evidence in Figure 1, interbank claims decline throughout the run. Further, we find that high-quality liquid assets are stable throughout most of the run and only start to fall slightly in July. This pattern arguably reflects that banks are anticipating a higher value of high-quality liquidity in later stages of the run and prefer to deplete their low-quality liquid assets first (Diamond and Rajan 2011). As the withdrawal of regular deposits sets in in June, banks reduce their holdings of lower-quality liquid assets such as bills of exchange. They do so by discounting the claims at the Reichsbank’s discount window in return for currency, which is then used to serve withdrawing depositors. As noted above, see Internet Appendix A.3 for more details on the behavior of the Reichsbank. We estimate that by the end of July, banks have reduced their holdings of low-quality liquid assets by around 75% compared to April, mirroring the outflow of time deposits. In contrast, banks’ illiquid assets contract much more slowly and by only around 10% from April through July.

4 Deposit Flows in Ex Post Failing and Surviving Banks

Next, we turn to our main analysis and ask which depositors are withdrawing from failing banks. Our empirical strategy exploits the fact that we can observe the ex post outcomes as to which banks fail throughout or in the aftermath of the crisis and which banks survive the crisis. While we have balance sheet information for more than 120 unique banks during the main phases of the crisis in 1931, 15 of these banks (around 12%) fail at some point during or in the aftermath of the run; see panel A of Table A.1 in the Internet Appendix for a list of failing banks and Figure A.9 for a timeline of the run and the failures.²⁰

In the following, we use both cross-sectional variation in bank failures and deposit flows to analyze whether depositors withdraw more funds from banks that end up failing compared to banks that survive the run.

4.1 The cross-section of deposit flows

We start out by establishing that there is substantial cross-sectional variation in deposit flows during the run. Figure 3 plots monthly deposit growth from March through July 1931. Just before the crisis starts, in March and April of 1931, the distribution of monthly bank-level deposit growth is centered around zero, with some banks receiving deposit inflows and some being subject to outflows. The interbank market then reallocates deposits from banks with inflow to those with outflows, and the banking system does not lose deposits, indicating successful risk-sharing (Allen and Gale 2000). However, as the crisis starts, the average deposit growth rate turns negative. Notably, between May and July, some banks lose more than 20% of their entire deposit base per month. However, throughout the entire run phase, there are always some banks that continue to receive deposit inflows.

Fig. 3

The cross-section of deposit flows through Spring/Summer 1931

This figure depicts the kernel density of monthly bank-level log-changes (in ppt) of total deposit funding from March 1931 to July 1931. We depict each month separately. We include all banks reporting in Deutscher Staats- und Preussischer Reichsanzeiger.

Open in new tabDownload slide

4.2 Who withdraws from failing banks?

We also include a set of observable bank characteristics X_b. For instance, to proxy the distance to default, we calculate the ratio of a bank’s total liabilities (calculated as total assets net of equity) over a bank’s equity. To proxy for a bank’s (il)liquidity, we calculate the ratio of liquid security holdings and interbank claims over total deposits. This addresses the concern that depositors are more likely to withdraw at banks that appear more likely to become illiquid throughout the run. Further, we control for bank size using quartile indicators for size based on total assets. We also control for whether a bank relies on foreign-denominated deposits and thus address the concern that bank failures may primarily be a by-product of the run on the currency. Given the prominent role of the interbank market, we also control for the reliance on interbank funding measured by the share of interbank deposits over total deposits. Finally, given the prominent role of the bankruptcy of “Nordwolle”, we also include a dummy of whether a bank has an observable relationship to this firm to test whether bank failures are driven by the failure of this larger borrower. We calculate all control variables as bank-level averages from February 1931 through April 1931. Note that except for the reliance on foreign currency denominated deposits, all the characteristics were in principle easily available to the depositors at the time via the sources we are using.

An important caveat is that we do not observe the ultimate cause of bank failures. Hence, we cannot identify whether withdrawal motives are based on the prospect of default or whether they are the cause of default. Said differently, failure could be the consequence of deposit flows and the interpretation of β₁ is not causal. However, to the extent that there is variation across different types of deposits, we are nonetheless able to identify heterogeneity in depositor information. Variation in the contraction of deposit flows (or the lack thereof) across failing and surviving banks allows to understand whether depositors can tell which banks will fail or not. Variation across different types of deposits can give a sense whether some depositors are better at anticipating which banks will fail or not. Thus, our research objective allows us to remain agnostic about the causes of failures. For instance, we cannot tell whether a bank would have failed even in absence of withdrawals (fundamental failure) or due to the withdrawals (panic-based failure).²⁴

Table 2 shows our results. There is no statistically significant difference in the growth of regular deposits between failing and nonfailing banks throughout the run (see columns 1 and 2). The point estimates suggest that regular deposits grow at a slightly lower rate at failing banks but the confidence bands suggest that total deposit funding at most falls by 10-percentage-points more at failing banks than at surviving banks. However, the confidence bands also allow for the possibility that deposits increase by 6-percentage-points less for failing banks. This finding is striking because in Figure 2 we estimate that regular deposits fall by more than 15% from April through July. Moreover, recall that all types of depositors should expect to realize losses if their bank were to fail. Yet, while deposits are falling substantially there is no statistically significant difference between failing and surviving banks. Hence, regular depositors either don’t withdraw at all or, to the extent that they do withdraw, they do not discriminate between weak and strong banks.

In contrast to regular deposits, there is a substantial difference in the growth of domestic interbank deposits between failing and surviving banks. Interbank deposits fall by around 71-76 percentage points more at failing banks than at surviving banks (see columns 3 and 4). The magnitude is remarkable since we estimated in Figure 2 that banks on average lose around 30% of their interbank funding. This implies that while surviving banks see essentially no changes in their domestic interbank deposits from April through July, those banks that end up failing lose approximately 70%. Thus, failing banks, while not losing more regular deposits throughout the run, effectively lose access to the interbank market.²⁵

This striking result on the difference between regular and interbank deposits can also be visualized by considering the density of the log-growth in regular and interbank deposits from April through July while splitting the sample into failing and surviving banks (see Figure 4). Panel A plots distribution of growth in regular deposits and reveals that—while deposits decline on average for both types of banks—there is no obvious difference in the flow of regular deposits across failing and surviving banks. The negative mean of the distribution shows that most both types of banks are subject to net outflows in deposits throughout the run. Notably, both failing and surviving banks that receive inflows of total deposit funding.

Fig. 4

Deposit growth from April 1931 through July 1931 for failing and surviving banks

This figure plots the kernel density functions for the log-change (in ppt) in total deposits (panel A) and interbank deposits (panel B) from April 1931 through July 1931, splitting the sample into banks that failed and those that survived. Data from Deutscher Staats- und Preussischer Reichsanzeiger.

Open in new tabDownload slide

Panel B plots the distribution of growth in interbank deposits and reveals a striking difference between failing and surviving banks. On average, there is almost no contraction in interbank deposits for surviving banks, and many surviving banks that see their interbank liabilities grow throughout the run. In contrast, almost no failing banks that increase their interbank borrowing and most density is to the far left, indicating that failing banks lose access to the interbank market.

Interbank deposits, however, are a relatively small part of overall deposit funding. Thus, their higher outflows at failing banks do not translate into a statistically significant difference and the above findings imply that there is no net difference in the outflow of total deposits—the sum of regular and interbank deposits—between failing and surviving banks (see columns 9 and 10 of Table 2).

Aside from the regular and interbank deposits, we also distinguish between time and demand deposits. We find that failing banks are subject to relatively higher growth of demand deposits and the inflow of demand deposits is mirrored by an outflow in time deposits, with only the latter being statistically significant and only if control variables are included. Unfortunately, as described in Section 2, our data do not allow us to distinguish whether the outflow in time deposits is primarily from interbank deposits or regular deposits. Thus, we distinguish whether the outflow in time deposits is driven by the outflow of interbank time deposits or outflows of regular time deposits.

4.3 Dynamics

Figure 5 shows our findings for regular, interbank, time, demand, and total deposits. First off, there are no differences in deposit flows across failing and surviving banks before the run starts after the failure of the Creditanstalt in May. Further, in line with our results from estimating Equation (2), we find that interbank deposits start to change relatively more rapidly for failing banks starting in June 1931. By June, failing banks report 40% less interbank deposits relative to surviving banks, and by July, the difference has grown to more than 70%. As before, there is virtually no difference in the change in regular deposits across failing and surviving banks throughout the run. Regular deposits if at all increase at failing banks during the early phase of the run, possibly because failing banks offer higher rates on regular deposits for raising funds to make up for the lost interbank funding as in the model by Egan, Hortaçsu, and Matvos (2017) and an empirical pattern also documented for recent failures of U.S. banks by, for example Martin, Puri, and Ufier (forthcoming).²⁷

Fig. 5

Deposit dynamics for failed banks

The figure displays the sequence of coefficients |$ \{\beta_{s} \} $| that results from estimating the model:

$$\displaystyle y_{b t} = \gamma_{b} + \gamma_{\theta t} + \sum \limits_{s \neq \text{April} \,31} \beta_{s} \times \mathbb{I} \left[ s = t \right] \times \text{Failed}_{b} + \sum \limits_{s \neq \text{April} \,31} \mu_{s} \times \mathbb{I} \left[ s = t \right] \times X_{b} + \epsilon_{b t},$$

where y_bt is the logarithm of one plus the type of deposit indicated in the figure for bank b in month t. We multiply y_bt with 100 to convert the coefficients into percentage points. |$ \text{Failed}_{b} $| is an indicator of whether a bank fails during or after the run. X_b is a set of bank-level control variables. We include a bank’s ratio of total liabilities (total assets net of equity) to equity, liquid assets (securities and interbank claims) to total deposits, interbank funding to total deposits, indicators of the size quartile based on total assets, an indicator for use of foreign-currency denominated deposits, and an indicator for whether a bank was connected to the nonfinancial firm “Nordwolle,” which declared bankruptcy in June 1931. All control variables are calculated by averaging at the bank level from February through April 1931. The model is estimated using balance sheets reported from February through July 1931 in the Deutscher Staats- und Preussischer Reichsanzeiger, dropping banks once they have failed. Finally, γ_b are bank-level fixed effects. The first vertical line, on May 11, 1931, represents the date of the failure of the Austrian Creditanstalt. The second vertical line, on July 13, 1931, represents the failure of Danatbank and the start of the banking holiday. Figure A.11 in the Internet Appendix shows the estimates for each type of deposit in separate plots including confidence bands.

Open in new tabDownload slide

Figure 5 also confirms the pattern of maturity shortening throughout the different phases of the run discussed above: in May, failing banks see 10% more growth in demand deposits while they also see around 10% lower levels of time deposits. Further, note that the effect is—unlike in Table 2—initially statistically significant, as shown in Figure A.11 in the Appendix. This findings suggests that some depositors at failing banks take a more cautious stance early in the run and convert their time deposits into demand deposits.

4.4 Predicting which banks will fail

To further corroborate the finding that interbank flows anticipate bank failures, we ask whether bank failures can indeed be predicted with deposit growth rates. To this end, we reverse the original empirical specification in Equation (2). We now relate bank failures to observable characteristics of banks but also to their deposit flows throughout the run. This approach is especially useful as it allows us to also evaluate the AUC, which quantifies by how much of both total deposit flows and interbank deposit flows help improve the econometrician’s ability to predict which banks will fail.

We estimate various versions of Equation (4). First, we estimate the model without including interbank deposit flows but just bank characteristics observable from before the run. This specification is useful as it allows to test whether observable characteristics from before the run and in principle available to contemporary depositors²⁸ allow to predict failures. Second, we include the growth of interbank deposits and total deposits from April 1931 through July 1931 each separately and together. This allows to understand to which extent interbank deposit and total deposit growth each increase the ability to predict bank failures. Finally, further below we also estimate a slightly different variant of the model that allows for a nonlinear relationship of bank failure and interbank deposit flows.

Table 3 shows results. Columns 1 and 2 show results from estimating Equation (4) when excluding deposit flows as an explanatory variable. This is a useful benchmark for what follows. We find that there is only limited explanatory power when using ex ante balance sheet characteristics. There is no systematic pattern for how balance sheets characteristics relate to bank failure.

Next, in columns 3 through 6, we include either growth in interbank funding or total deposit funding during the run as an explanatory variable. We find that interbank deposits indeed predict failure as evidenced by the statistically significant coefficient, which shows that banks with less interbank deposit outflows (higher interbank deposit growth) are less likely to fail. We find that a 10-ppt higher growth rate interbank funding is associated with an around 10-to 11-percentage-points higher chance of survival. In contrast, total deposit funding has no predictive power. This is also true when including both as explanatory variables (see columns 7 and 8). Table A.7 in the Appendix also shows that these findings are broadly robust to estimating the model using Probit rather than ordinary least squares (OLS).

Further, we also estimate a slightly modified model that takes into account that the relation between interbank deposit flows and failure can be nonlinear— with, for instance, only very substantial outflows being associated with bank failure but moderate interbank deposit outflows having little predictive power. To that end, we calculate dummies that indicate the quintile of the growth from interbank deposits from April 1931 through July 1931 for each bank and calculate the probability of failure for each quintile. We report the results of the regression in Figure 6. The pattern is clear: there is a monotonic but nonlinear relationship between interbank funding growth and the probability of failure. We find that firms with very heavy outflows (in the lowest quintile of the interbank deposits growth distribution) are most likely to fail. For instance, being in the lowest quintile implies that the chance of failure is around 40% and thus considerably higher than the unconditional probability of failure of around 10%. In contrast, being in the second or third quintile of the interbank funding growth distribution implies a probability of failure that is similar to the unconditional failure probability. Further, being in the 4th or 5th quintile implies a probability of failure that is significantly lower than the unconditional failure probability.

Fig. 6

Failure probability conditional on the quintile of interbank deposit growth during the run.

This figure shows the set of coefficients |$ \{\beta_{2,s} \} $| results from estimating the following model:

$$\displaystyle \text{Failed}_{b} = \beta_{1} \times X_{b} + \sum \limits_{s = 1}^{5} \beta_{2,s} \times \mathbb{I} \left[ \Delta \text{Interbank} \,\text{Deposits}_{\text{April} \,31 \colon \text{July} \,31} \in {Q}_{{s}} \right] \left({s} \right) + \gamma_{\theta} + \epsilon_{{b}},$$

where |$ \mathbb{I} (s) $| is an indicator that takes the value one if the growth if interbank deposits lies in quintile Q_s. Note that we suppress the constant and thus |$ \beta_{2,s} $| can be interpreted as the probability of failure conditional on being in quintile s of the interbank deposit growth distribution.

As control variables, X_b, we include a bank’s ratio of total liabilities (total assets net of equity) to equity, liquid assets (securities and interbank claims) to total deposits, interbank funding to total deposits, indicators of the size quartile based on total assets, an indicator for use of foreign-currency denominated deposits, and an indicator for whether a bank was connected to the nonfinancial firm “Nordwolle” that declared bankruptcy in June 1931. All control variables are calculated by averaging at the bank level from February through April 1931. The model is estimated using the cross-section of banks that report balance sheets in July 1931 in the Deutscher Staats- und Preussischer Reichsanzeiger and dropping banks that have failed before the banking holiday of July 1931. Confidence bands of 95% applied.

Open in new tabDownload slide

To formally understand how powerful interbank flows are in predicting bank failure, we next use a standard tool used to evaluate binary classification ability, the receiver operating characteristic curve (ROC) as is also done in, for example Schularick and Taylor (2012). To gauge the ability to predict which banks will fail, we plot the AUC. The AUC provides a simple test against the null value of 0.5 with an asymptotic normal distribution.

To visualize our main insights, we calculate and compare the AUC obtained from various versions of the regression model in Equation (4) and the specification underlying estimation of the coefficients shown in Figure 6. First, we compare the specification in column 1 in Table 3 and compare it to the specifications in columns 3 and 4 in the same table in which we control for either interbank deposit growth or total deposit growth. Finally, we compare all three to the AUC of the nonlinear model in which we use the quintile indicators of the interbank deposit growth distribution as explanatory variables. We also report the AUC for all other estimated regression models at the bottom of Table 3.

Figure 7 shows our main insights. Our baseline model that excludes interbank deposit flows has an AUC of 0.58 and hence just above 0.5. Thus, trying to predict bank failure with observable balance sheet characteristics alone is only slightly better than tossing a coin. Thus, there is next to no chance of identifying a bank that will fail from these characteristics alone without having a very large number of false positives. Further, the AUC increases only minimally to just above 0.6 once we also control for the flow of total deposits. This is particularly striking as it emphasizes that the total shortfall of funding is about the same for both failing and surviving banks.

Fig. 7

Predicting failure: Receiver operating characteristic plot

This figure plots the receiver operating characteristics (ROC) curve from estimating three versions of Equation (4) as estimated in Table 3 and Figure 6. The curve “Baseline controls” refers to the AUC corresponding to the estimates from column 1 of Table 3 and includes observable bank characteristics in which we include a bank’s ratio of total liabilities (total assets net of equity) to equity, liquid assets (securities and interbank claims) to total deposits, interbank funding to total deposits, indicators of the size quartile based on total assets, an indicator for use of foreign-currency denominated deposits, and an indicator for whether a bank was connected to the nonfinancial firm “Nordwolle” that declared bankruptcy in June 1931. The curve “Controls and total deposit growth refers (linear)” corresponds to the estimation of column 4 of Table 3, which include the bank characteristics and a measure of total deposit changes over the crisis. The curve “Controls and interbank deposit growth refers (linear)” corresponds to the estimation of column 3 of Table 3 and includes the above-mentioned controls as well as a measure of interbank deposit changes at the bank over the crisis. Finally, the curve “Controls and interbank deposit growth refers (nonlinear)” corresponds to the estimation from shown in Figure 6 and makes use of the above-mentioned controls as well as dummies for quintiles of interbank deposit growth.

Open in new tabDownload slide

However, we find that the AUC increases substantially when including interbank deposit growth as an explanatory variable. Adding the interbank deposit flows to the predictive model, that is going from column 1 to column 3 increases the AUC from 0.58 to around 0.72. This is a substantial increase in the AUC. Hence, knowledge of interbank deposits growth throughout the run substantially increases the ability to tell which banks will fail. More than that, once we allow for the relation between failures and interbank deposit flows to be nonlinear, the pattern becomes even clearer. The AUC goes up to 0.82. The model with an AUC of 0.82 performs very well in identifying failing banks. For instance, a policy maker with knowledge of contemporaneous interbank deposit flows and willing to accept a false positive rate of 0.5 could have identified almost around 90% of all failing banks during the run, see Figure 7.

Our findings suggest that interbank deposits are a predictor of whether a bank will fail.²⁹ Further, the relationship is somewhat nonlinear and substantial outflows in interbank funding are associated with an elevated chance of bank failure during the run. The outflow in regular deposits, in contrast, has little explanatory power for who will fail.

4.5 Are banks better informed or do interbank deposit outflows cause failure?

A priori, there at least two³⁰ equally plausible explanations for the finding that bank deposits decline almost exclusively for failing banks while regular deposits fall equally for both failing and surviving banks. On the one hand, banks may have been better informed than regular depositors and thus banks can discriminate between banks that end up failing and those that end up surviving and hence they withdraw from the former to protect themselves against potential losses. One the other hand, the fact that some banks lost access to the interbank market could have been the immediate cause of bank failure.

While both possible explanations are not mutually exclusive and in general both can be true at the same time, we believe, however, that the evidence suggests that the former explanation is more plausible. There are two main reasons why that is the case. First, recall that failing banks do not lose more total funding. Failing and surviving banks lose about the same percent of deposits despite the much higher interbank deposit outflows at failing banks since interbank funding is a relatively small share of total funding. However, standard models of bank runs suggest that the total shortfall of funding governs whether a bank becomes illiquid first and then insolvent (Diamond and Dybvig 1983; Goldstein and Pauzner 2005) as opposed to the composition of funding.³¹ Hence, we argue that the decline in interbank funding alone is unlikely to be the immediate cause of the failures. This finding is reminiscent of the evidence by Perignon, Thesmar, and Vuillemey (2018), who find that informed investors tend to withdraw, but not cause the lower performance of weak banks when studying wholesale funding dry-ups during the European Debt Crisis.

Second, we also provide an explicit test on whether interbank deposit flows cause failure and study whether our results are robust to excluding banks that are more reliant on interbank funding. That is, we test whether our results are robust when estimating Equation (2) but using only banks that have relatively little reliance on interbank funding before the run³². For such a subsample, it is particularly implausible for interbank deposit outflows to be the immediate cause of the failure as they are a small share of overall funding to begin with. We thus estimate Equation (2) to study the growth of regular, interbank, and total deposit funding throughout the run when restricting the sample to banks with less than 10%, 7.5%, or 5% of their total deposits funding coming from with interbank deposits.³³

The results, which can be found in Table 4, confirm our main findings: regular and total deposits fall by about the same for both failing and surviving banks. In contrast, the difference in interbank deposits remains statistically significant. For instance, we find that for the sample of banks with less than 10% of interbank funding prior to the run, interbank funding falls by around 40% on average—slightly higher than the 30% in the main sample—and the difference between failing and surviving banks is more than 100 percentage points (see column 4). In contrast, there is no difference in the decline in regular or total deposit funding (see columns 1 and 7). We interpret these findings as evidence that suggest that the outflow of interbank deposits alone is unlikely to be able to trigger a bank’s illiquidty and insolvency.³⁴

4.6 Alternative explanations for the main findings

Above, we considered what we believe are the two most obvious candidates to explain our main findings. Next, we discuss other alternative explanations.

A third alternative explanation would be that failing banks choose a different investment strategy than surviving banks during the run. This difference in asset selection, in turn, could then translate into a difference in the funding choices. We argue that this explanation is less plausible for the two following reasons. First, recall from Figure 5 that interbank funding drops very quickly during the run. Note though that a quick adjustment in the asset-side composition tends to be costly when liquidity dries up and bid-ask spreads widen. Hence, we argue that the pattern is more likely a result of depositors withdrawing funds that are available on demand as opposed to banks’ asset selection. Second, Figure A.12 in the Internet Appendix we also study the dynamic coefficients plot obtained when estimating Equation (3) using bank assets as outcome variables. We find that failing banks are—if at all—less likely to reduce their loan portfolio, in line with failing banks continuing to lend to their customers, indicating possible forbearance. Hence, we do not observe failing banks reducing their funding needs during the run.

A fourth explanation candidate are differences in opportunity costs. We do not observe differences in information sets or in opportunity costs across different types of depositors. Hence, another potential concern stems from potentially unobservable differences in opportunity costs. Specifically, it may be more costly for a retail depositor to withdraw from a bank and invest funds someplace else than for bank. Hence, differences in opportunity costs may be affecting the observed outcomes and hence our interpretation of the findings. Note, however, that opportunity costs can only explain variation in the responsiveness of different types of depositors across time, but not across failing and surviving banks. A depositor with relatively higher opportunity costs will plausibly withdraw later in the run (as distress becomes more severe). However, differences in opportunity cost cannot explain the fact that regular depositors do not distinguish between failing and surviving banks, as documented in Table 2. Thus, the fact that interbank deposits essentially collapse for failing banks but regular deposits do not can be explained by differences in information about prospective bank failure, but not by differences in opportunity costs.

5 Interbank Market

The striking difference between interbank deposit flows for failing and surviving bank raises the intriguing question whether surviving banks can borrow from other banks when subject to deposit outflows. This question is especially intriguing as Figure 3 revealed that some bank continue to receive deposit inflows during the run. Hence, we next study the interbank market in more detail and whether there is reallocation within the interbank market. We ask: do banks with deposit inflows redeposit within the interbank market or do they hoard cash and other liquid assets as suggested in theories by Caballero and Krishnamurthy (2008); Allen, Carletti, and Gale (2009)?

That is, we define the interbank exposure of a bank as the relative difference between interbank claims and interbank deposits. We can then study the correlation of change in the exposure with the change in regular deposits, both normalized by bank size as measured by a bank’s total assets.

In a normally functioning interbank market, one would expect that a bank with deposit inflows would lend out the received funds to those with deposit outflows and thus increase the interbank exposure. In contrast, a bank that is subject to deposit outflows would borrow through the interbank market and thus reduce its interbank exposure. See, for example Allen and Gale (2000) for a model of such interbank deposit flows in which the interbank market insures banks against bank-specific withdrawal shocks, suggesting a positive correlation of interbank exposure with regular deposit funding growth.

Figure 8 shows the relation of the month-to-month change in interbank exposure for failing and surviving banks for both the period before the run (Panel A) and during the run (Panel B). The positive correlation in Panel A confirms the theoretical notion of how interbank markets work and bank balance outflows by borrowing from banks with inflows. Indeed, before the run starts, both failing and surviving banks increase their interbank exposure in a month in which they receive deposit inflows and decrease their interbank exposure in a month in which they are subject to deposit outflows. Hence, the interbank market works and reallocates the funds effectively.

Fig. 8

Correlating changes in interbank exposure and regular deposit funding before and during the run

This figures relates banks’ change in regular deposit funding (total deposits net of interbank deposits), normalized by total assets, to banks’ interbank exposure defined as:

$$\displaystyle \Delta \text{Interbank} \,\text{Exposure}= \Delta [ \text{Interbank} \,\text{Lending-Interbank} \,\text{Borrowing} ].$$

The sample is split by whether or not a bank fails during or after the run. The left panel uses bank-level outcomes from January 1930 through April 1931. The right panel uses bank-level outcomes from May 1931 through July 1931.

Open in new tabDownload slide

Panel B, however, reveals a striking difference between failing and surviving banks during the run. The correlation between the change in interbank exposure and the change in regular deposits goes from close to one to zero and loses its statistical significance. Even more striking, however, is the positive correlation between interbank exposure and deposits for surviving banks, which remains close to one. Hence, while failing banks get excluded from the interbank market, surviving banks subject to deposit outflows in a given month continue to be able to borrow from those banks with deposit inflows. Said differently, banks with inflows of regular deposits during the run do not hoard the funds they receive but intermediate them to other surviving banks via the interbank market.

Here, there are three coefficients of interest. β₁ is the average correlation between the change in interbank exposure and regular deposit flows (changes in total deposits net of interbank deposits). In a functioning interbank market, this coefficient should be close to one. β₂ is the relative difference in the relationship between changes in interbank exposure and deposit flows for all banks. If the interbank collapses for all banks, it should be a negative number close to –1. Further, β₃ is the relative change in the relationship between changes in interbank exposure and deposit flows during the run for failing banks alone.

Results can be found in Table 5. First, note that β₁ is indeed close to one, confirming the notion that in normal times, banks with deposit inflows increase their interbank exposure and banks with deposit outflows decrease it. β₂ is also statistically significant and negative in some specification. Thus, the overall intermediation is less during the run than before the run. However, the slope remains positive and relatively close to one and the interbank market as a whole does not collapse but functions under distress. Finally, note that β₃ indicates that the correlation is largely reduced for failing banks during the run, echoing the findings in Figure 8. That is the interbank market collapses mostly for failing banks.

As indicated above our sample does not cover the entire banking system but the subset of commercial banks. Eventual concerns that we only have data for a subset of the entire banking system are addressed by the fact that the interbank market Germany was segmented. In one part of the interbank market which we only have partially covered in our data, Landesbanken and Girozentralen interact with their associated local savings banks. In the second part, which our data covers entirely, commercial banks such as the Berlin banks and the regional banks would borrow and lend. Thus, it is reassuring that we can show that our results are robust to including bank-type-time fixed effect or analyzing only Berlin banks and regional banks (see columns 2 and 3 Table 5).

Table A.4 in the Internet Appendix shows results when estimating a Probit model that uses an indicator whether a bank increases its interbank funding from t—1 to t as the dependent variable. Table A.5 shows results when using the change in interbank deposits as opposed to interbank exposure as the outcome variable. The findings confirm that banks tend to increase their interbank borrowing whenever faced with deposit outflows in normal times. This relationship continues to hold throughout the run for surviving banks but stops to hold for failing banks.

Taken together, our evidence on the dynamics of the interbank market stand in contrast to theories of liquidity hoarding by banks in times of distress (Allen, Carletti, and Gale 2009; Caballero and Krishnamurthy 2008) and interbank market contagion (Allen and Gale 2000; Liu 2016) as well as empirical findings that suggest that interbank market necessarily amplify financial distress (see, e.g., Mitchener and Richardson 2019). However, our empirical findings complement and support the findings of Afonso, Kovner, and Schoar (2011) who show that the interbank market in the United States became more risk sensitive during the GFC but continued to function. We show that for the run on the German banks in 1931, the interbank market continued to provide liquidity to surviving banks, but not to failing banks for whom the interbank market collapses.

5.1 Discussion

Altogether, we argue that our findings suggest that banks have the most information about the state of the banking system and are in effect the most sophisticated type of depositor. Banks seem to have information about which of their competitors will likely fail in light of an aggregate shock. The pattern that emerges from the data is very close to the mechanism suggested by Calomiris and Kahn (1991) in which the most informed depositors are rewarded for being informed since they can withdraw from failing banks before other depositors do. Note, however, that this interpretation needs to be taken with some caution. While we believe that our evidence supports the interpretation that banks are most informed, we cannot entirely rule out alternative explanations that can be true at the same time.

Importantly, the fact that banks are seemingly well informed about their competitors financial fortunes in the light of the run in turn allows them to provide liquidity to surviving banks via the interbank market: to the extent that banks receive deposit inflows during the run, they continue to intermediate these funds to banks that are subject to deposit outflows. However, banks discriminate between weaker banks that end up failing and stronger banks that end up surviving and only provide liquidity to the latter.

What information do banks have that regular depositors do not have? Unfortunately, we cannot identify what exact information banks are acting on. Our findings allow for different possibilities: For instance, banks can have information about a specific bank’s solvency, or banks have information about which banks are more likely to fail when other depositors withdraw funds. In either case, however, banks can distinguish failing banks from surviving banks while regular depositors cannot. Moreover, being able to observe interbank deposit flows allows to predict with a high degree of precision which banks will fail.

6 Additional Results

6.1 The role of foreign-currency-denominated deposits

The literature on the German Crisis of 1931 has typically stressed the role of foreign-denominated deposits in the run (see, e.g., Ferguson and Temin 2003). Next, we analyze the deposit flow across banks with and without historical reliance on deposits denominated in foreign currency in more detail.

Importantly, the exposure to foreign deposits was not publicly available. Banks only reported their exposure infrequently in confidential filings with the Reichsbank. Figure A.18 in the Internet Appendix shows an example of one such filing. Of course, depositors could have had private information about which banks used foreign-currency-denominated deposits. We can thus ask: are domestic interbank deposits more likely to flow out of banks that rely on foreign deposit funding? If that were the case, to the extent that depositors with foreign currency denominated deposits had stronger incentives to withdraw, it would be another smoking gun indicating that banks are informed about which banks are more likely to be troubled during the run.

In columns 4 of Table 2, we learn that interbank deposits are indeed more likely to flow out of banks that rely on deposit funding. We find that banks with reliance on deposits denominated in a foreign currency see a roughly 64% higher contraction in domestic interbank deposits. Column 10 suggests that these banks also lose around 11% more total deposit funding. Further note that Figure 9 confirms this pattern. Banks that rely on more foreign funding lose interbank funding over time. Note, however, that this is while controlling for whether a bank fails.

Fig. 9

Deposit dynamics for banks with foreign deposits

The figure displays the sequence of coefficients |$ \{\beta_{s} \} $| that results from estimating the model:

$$\displaystyle y_{b t} = \gamma_{b} + \gamma_{\theta t} + \sum \limits_{s \neq \text{April} \,31} \beta_{s} \times \mathbb{I} \left[ s = t \right] \times \text{Foreign} + \sum \limits_{s \neq \text{April} \,31} \mu_{s} \times \mathbb{I} \left[ s = t \right] \times X_{b} + \epsilon_{b t}.$$

where y_bt is the logarithm of one plus the type of deposit indicated in the figure legend for bank b in month t and |$ \text{Foreign}_{b} $| is an indicator whether a bank relies on foreign deposit funding. We multiply y_bt with 100 to convert the coefficients into percentage points. X_b is a set of bank-level control variable. We include a bank’s ratio of total liabilities (total assets net of equity) to equity, liquid assets (securities and interbank claims) to total deposits, interbank funding to total deposits, indicators of the size quartile based on total assets, an indicator for use of foreign-currency denominated deposits, and an indicator for whether a bank was connected to the nonfinancial firm “Nordwolle” that declared bankruptcy in June 1931. All control variables are calculated by averaging at the bank level from February through April 1931. γ_b are bank and |$ \gamma_{\theta t} $| are bank-type*time fixed effects. The model is estimated using balance sheets reported from February through July 1931 in the Deutscher Staats- und Preussischer Reichsanzeiger, dropping banks once they have failed. The first vertical line, on May 11, 1931, marks the date of the failure of the Austrian Creditanstalt. The second vertical line, on July 13, 1931, represents the failure of Danatbank and the start of the banking holiday.

Open in new tabDownload slide

The above findings are further a reassuring robustness test as they reveal that all our main findings hold when controlling for the exposure to foreign-currency-denominated deposits. They also support our main finding on the interbank market having private information on other banks’ risks. The exposure to foreign-denominated deposits was not publicly known, but our findings suggest that the interbank market is very well informed about which banks may be subject to withdrawals because they rely on foreign-currency-denominated deposits.

6.2 Does the stock market identify failing banks?

A natural additional test is whether the stock market identifies failing banks. If stock price dynamics reflect the chance of bank failure, the findings that regular deposits are not able to distinguish between failing and surviving banks would of course be even more striking as stock prices are publicly observable and easily available via widely circulated newspapers. Similarly, it is of interest to look at the extent to which stock prices are following or leading the dynamics in the interbank market.

We first study the difference in stock price between failing and surviving banks across time. Note that the sample we can use here is much smaller than our original sample as we only have data for 32 banks in the Monatskursblatt of which 28 report balance sheet information and of which only five become distressed and of which only two fail.³⁵ Hence, the small sample size restricts the ability to make inference. Further, after the breakdown of the banking system on July 13, 1931, the stock exchange was closed and only reopened in September 1931.

Figure 10 shows the results when estimating Equation (3) for daily stock market data. It reveals that failing and surviving banks’ stock were following a quite similar trajectory before the run started. Note that at the first vertical lines, right at the failure of the Creditanstalt, stock prices for failing banks drop by around 5% compared to those that survive. This is an indication that stock market participants realize the importance of the event for the stability of German banks, especially for those banks that are weaker and ultimately fail. However, the difference in the level only turns statistically significant a few days later.

Fig. 10

Stock price dynamics for failed banks

This figure plots the sequence of coefficients |$ \{\beta_{s} \} $| from estimating a regression of the form

$$\displaystyle y_{b t} = \gamma_{b} + \gamma_{\theta t} + \sum \limits_{s \neq \text{April} \,31} \beta_{s} \times \mathbb{I} \left[ s = t \right] \times \text{Failed} + \epsilon_{b t},$$

where y_bt is the natural logarithm of the stock price of bank b on day t. γ_b are bank and |$ \gamma_{\mathit{thetat}} $| are bank-type* time fixed effects. 95% confidence intervals have been applied. The regressions have been normalized to t=May 1st, 1931.

Open in new tabDownload slide

Interestingly, the stock prices for failing banks are already substantially lower (by around 10%) by June 6, when the German government announced the end of reparations. Thus, similarly to the aggregate interbank flows, stock prices start to fall for failing banks early on.

By July 13 when the banking system breaks down entirely and the stock market closes, banks that end up failing have lost around 25% more of their stock market value than those that survive the run. This findings emphasizes how striking our original finding on the behavior of regular depositors is. Regular depositors are seemingly unable or unwilling to incorporate the information contained in stock prices—which are, as mentioned above, publicly available via newspapers—into their withdrawal decision.

Table 6 reports results. We find no general relationship between past or contemporaneous interbank flows and stock prices, see column 1 in both panels A and B. Further, studying the effect month by month, we find that outside of the run, interbank flows have no predictive power for stock market prices. For instance, interbank flows in March or April 1931 have no effect on stock price in April or May, respectively (see columns 2 and 3). However, we find that during the first phase of the month, May 1931, banks that lose less interbank funding have abnormally high stock market returns in June 1931 (see column 4 in panel A). However, the same is not true for contemporaneous interbank flows (see column 4 of Panel B). Thus, the interbank market seems to have more information about which banks may fail than the stock market, especially early in the crisis. In the final month of the run, the stock market seems to have incorporated all information and there is again no relation between interbank flows and stock prices (see column 5).

7 Concluding Remarks

In this paper, we exploit the unique historical incident of a run on the entire German banking system during the summer of 1931. Having granular balance-sheet data for commercial banks as well as the central bank, we provide a comprehensive empirical description of the dynamics of the run and establish which types of depositors can discriminate between failing and surviving banks in a bank run.

We find that all banks lose around 20% of their overall deposit funding before the height of the crisis and there is an equal outflow of retail and wholesale deposits from both ex post failing and surviving banks. Regular depositors are thus unable to identify failing banks. In contrast, the interbank market anticipates which banks will fail. The interbank market collapses for failing banks entirely but it continues to function for surviving banks, which can continue to borrow from other banks in response to deposit outflows (Afonso, Kovner, and Schoar 2011).

Given that both failing and surviving banks lose the same amount of deposits in the run, it is thus unlikely that the interbank market run causes bank failures (Perignon, Thesmar, and Vuillemey 2018). However, we cannot tell what banks are informed about. Our findings allow for two possibilities: banks having information about a specific bank’s solvency or banks having information about which banks are more likely to fail when other depositors withdraw from them.

Our paper contributes to the broader understanding of the role of short-term debt for financial intermediaries. Our findings highlight the different roles of short-term debt. We argue that some depositors are uninformed and hold short-term debt to obtain liquidity services (Diamond and Dybvig 1983; Gorton and Pennachi 1990), while others are informed and able to discipline banks (Calomiris and Kahn 1991; Diamond and Rajan 2000, 2001). Specifically, our evidence indicates that interbank depositors are the most informed and are rewarded for being informed since they are the first depositors that withdraw from failing banks, in line with the mechanism in Calomiris and Kahn (1991). However, it is important to highlight that we are not testing the effectiveness of depositor discipline itself. In fact, it is plausible that, rather than disciplining motive, coordination problems may be first order during a financial crisis. Our evidence muted about whether depositors provide discipline in equilibrium.