Liquidity creation, investment, and growth

Abstract

Using panel analysis for a large cross-section of countries, we find that liquidity creation by banks is positively associated with economic growth at country and industry levels. Liquidity creation boosts tangible, but not intangible investment and does not contribute to growth in countries with a high share of industries reliant on intangible assets. These findings are consistent with a theoretical model in which liquidity creation fosters investment only if it is sufficiently tangible. Our results shed light on important heterogeneities in the role of banks in the economic development process and their limited role in countries’ transition to knowledge economies.

1 Introduction

A key function of banks in the economy is the provision of liquidity by funding illiquid assets with liquid liabilities (Diamond & Dybvig, 1983; Holmström & Tirole, 1998). Bank loans provide funding for long-term investments, while bank deposits serve as a safe and liquid transaction medium that forms the core of the payment infrastructure in modern economies. Prior work documents a positive relation between overall banking sector development and economic growth (Beck et al., 2000; Rajan & Zingales, 1998). However, there is little research focusing specifically on whether and how liquidity creation, as a key function of banks to foster long-term investments, contributes to growth. A few papers examine the relation between banks’ liquidity creation and growth, but only in a single-country setting (Berger & Sedunov, 2017; Fidrmuc et al., 2015).

Several studies indicate that the role of banks in fostering economic activity exhibits important heterogeneity. Across countries, Arcand et al. (2015) find that banking sector development stops contributing to growth beyond a certain threshold, while Čihák et al. (2012) show that, as economies develop, securities markets become more important for growth relative to banks. Across industries, Hsu et al. (2014) report that banking sector (equity market) development is negatively (positively) related to innovation in industries more dependent on external finance. Aghion et al. (2004) also show that R &D intensive firms are more likely to raise funds by issuing shares than through debt. Indeed, the role of banks in supporting innovation—an important channel through which finance can affect growth (Aghion et al., 2018)—is subject to debate. Dell’Ariccia et al. (2020) argue that banks have a comparative advantage in financing standardized and well-collateralized investment projects, as opposed to more innovative projects that rely on intangible assets such as R &D. None of these papers, however, focuses on the specific function of banks as liquidity creators to understand the interplay between banks, innovation, and growth.

This debate highlights the need to understand how liquidity created by the banking sector relates to overall but also industry-specific economic activity. This paper fills this gap by providing evidence that liquidity creation by banks is associated with higher economic growth at both country and industry levels. In particular, we find that liquidity creation boosts tangible, but not intangible investment both across countries and more so for industries more in need of debt financing. Our findings suggest an important non-linearity in the relation between liquidity creation and economic growth; liquidity creation does not contribute to growth in countries with a higher share of industries relying on intangible rather than tangible assets.

To examine the relation between banks’ liquidity creation and the real economy, we combine bank-, industry-, and country-level data. Specifically, we use unconsolidated data on 18,217 commercial, savings, and cooperative banks operating in 100 countries from 1987 to 2014 and build on the work of Berger and Bouwman (2009) to measure banks’ liquidity creation on and off the balance sheet. We find that global liquidity creation by banks has increased substantially since 1987, reaching $\$16.1$ trillion in 2014—of which $\$11.4$ trillion was created on their balance sheets and $\$4.6$ trillion off their balance sheets. Large banks (with assets exceeding $3 billion) consist of about $15\%$ of our sample, but are responsible for $72\%$ of global liquidity creation.

Next, we examine the relation between liquidity creation and economic growth.¹ We estimate dynamic panel models for GDP per capita including autoregressive dynamics as well as year and country fixed effects and time-varying controls, using both the standard within estimator and a Generalized Method of Moments (GMM) estimator. Accounting for dynamics in the GDP process allows us to distinguish short-run from long-run relations between liquidity creation and GDP. Our empirical strategy results in robust and precise estimates that indicate a $1.11\%$ increase in long-run GDP per capita following a permanent 10-percent increase in on-balance sheet liquidity creation per capita. We verify the robustness of our results to the inclusion of various time-varying controls that could confound the effect of liquidity creation on GDP (such as financial reforms and attributes of countries’ financial systems). We further find that the amount of liquidity created by banks off the balance sheet is consequential for growth. This effect is, however, quantitatively smaller. Our estimates indicate that a permanent 10-percent increase in off-balance sheet liquidity creation per capita is associated with a $0.36\%$ increase in long-run GDP per capita.

We address endogeneity concerns of these cross-country results by conducting additional tests that exploit industry heterogeneity. Although we control for time-varying country-characteristics and country fixed effects, other factors may still coincide with changes in liquidity creation, implying that we could incorrectly attribute the changes in GDP per capita to changes in liquidity creation. To address these concerns, we exploit industry variation in dependence on debt financing, thus extending the approach of Rajan and Zingales (1998) to the function of banks as liquidity creators. We find that liquidity creation has a systematically larger effect on output in industries more dependent on debt financing, consistent with our results on country-level growth.

Nonetheless, the functioning of the banking sector could merely respond to changing demands from the real economy. Although abundant research indicates that banking sector development leads to long-run growth (see our discussion of the literature below), reverse causality cannot be ruled out easily, because liquidity creation by banks can be a result of economic growth, rather than a cause thereof. To provide further support for a causal interpretation we explore empirically and theoretically the channel via which liquidity creation causes growth.

We use a variety of data sets on both country- and industry-level investment to derive three further sets of results. First, we find that liquidity creation is positively and significantly associated with net tangible (but not intangible) investment rates at the country level. This finding indicates that liquidity created by banks boosts investment in tangible, but not intangible assets. Second, we again exploit within-country variation across industries over time to sharpen the identification of the effect of liquidity creation on investment. Corroborating our country-level results, we find that liquidity creation increases net tangible investment rates in industries that are relatively more in need of debt financing, while we do not find evidence that liquidity creation affects net intangible investment rates in such industries. Third, we connect our industry-level results on tangible versus non-tangible investment with our country-level evidence on the relation between liquidity and economic growth. We show that liquidity creation has a weaker, if not insignificant, effect on growth in countries with a higher share of industries relying on intangible assets.

We rationalize our empirical findings in a model of liquidity creation based on Diamond and Dybvig (1983). While previous theories establish a positive effect of banks’ liquidity creation on investment and growth (Allen & Gale, 2000; Bencivenga & Smith, 1991; Levine, 1991; Wallace, 1996), the novel angle of our model is that it examines the role of asset tangibility in this process. To this end, we extend the baseline Diamond-Dybvig model by (i) adding liquidation costs differentiated according to whether the investment is tangible or intangible, and (ii) by explicitly modeling investors that are subject to a moral hazard problem when making the long-term investment, which is also affected by the investment’s (in)tangibility. In the model, banks can increase overall investment by providing liquidity through demand deposit contracts. However, this process is hampered by a moral hazard problem as investors may divert assets and default on bank loans—even though banks can seize the deposit claim of diverting investors. This moral hazard problem is particularly strong if asset tangibility is low, for two reasons. First, intangible investments may be more easily diverted as they are harder to assess by outsiders. Second, failing intangible investments leave the bank with relatively low collateral value, reducing the value of claims on the bank. This makes it attractive for investors with successful projects to divert even if the bank can seize their deposit claims. Overall, these theoretical results rationalize the empirical finding that liquidity creation by banks supports economic growth mainly through tangible investment.

Taken together, our analysis provides a unified framework that features liquidity creation by banks as a key mechanism to help understand a number of important findings in the finance and growth literature. We show that liquidity creation helps economies grow faster by fostering tangible investment. As economies rely more on intangible assets, the importance of liquidity creation decreases, where traditional bank lending hits its limits. Unlike prior studies, which use a general size-based indicator of banking sector development, we focus specifically on an empirical gauge of one of the critical functions of banking (liquidity creation), which captures the full spectrum of banking activities on the asset side and the liability side, as well as off-balance sheet activities. We also highlight tangible investment as a key channel through which banks support growth.

Our work contributes to several strands of the literature. It is most directly related to the finance and growth literature (see Levine, 2005; Popov, 2018, for surveys).² Until recently, the literature has used crude proxies focusing on the size of the banking sector (such as private credit to GDP) for lack of variables capturing the individual functions of the banking sector. Berger and Bouwman (2009) have made valuable progress by proposing a direct, bank-level measure of liquidity creation based on classifying all bank balance sheet items as liquidity creating or liquidity reducing.³ However, almost all studies using this measure focus on the US (Berger & Bouwman, 2017; Chatterjee, 2015; Chen et al., 2020; Jiang et al., 2019). This paper goes beyond these studies by measuring the amount of liquidity created by banks in 100 countries and uncovering first evidence on the relation between liquidity creation and economic growth at the country and industry levels.⁴

Our analysis also adds to our understanding of heterogeneity or non-linearities in the relation between banking sector development and economic activity, and of the role of banks in supporting innovation. Arcand et al. (2015) and Čihák et al. (2012) suggest that banks become less important in supporting growth for more developed countries. Aghion et al. (2005) show that banking sector development helps economies converge to the growth rate of the world frontier but does not help them grow beyond this frontier. The debate on the role of banks in supporting innovation (Aghion et al., 2004; Hsu et al., 2014) could shed light on these country-level results. Powerful banks can stymie innovation by extracting informational rents and protecting established firms (Hellwig, 1991; Rajan, 1992). Banks as debt issuers have an inherent bias toward conservative investments, so that bank-based systems might thwart innovation and growth (Weinstein & Yafeh, 1998; Morck & Nakamura, 1999). Related to our paper, Brown et al. (2013) find that better stock (credit) market access is associated with more R &D (fixed) investment. Other studies show that financial development has a causal role in the reduction of macroeconomic volatility, and more strongly so among industries more dependent on external finance (Braun & Larrain, 2005) and with higher liquidity needs (Raddatz, 2006).

In line with the channel behind our results, the corporate finance literature highlights that intangibles may reduce a firm’s debt capacity due to their low collateral value and because intangible assets are more easily diverted (e.g., Rampini & Viswanathan, 2010; Falato et al., 2020). Consistent with our findings, Dell’Ariccia et al. (2020) show that banks shift their lending away from corporate lending towards real estate when firms invest more in intangible assets, and Döttling and Ratnovski (2021) find that intangible investment responds less to the credit channel of monetary policy. We contribute to these lines of research by showing how liquidity creation affects economic activity in a non-linear way by fostering tangible but not intangible investment, and thereby failing to boost growth in more-developed countries with a greater reliance on intangibles. We thus also add new insights to the long-standing debate on the relative merits of bank-based versus market-based financial systems (Allen et al., 2018).

We also build on the theoretical banking literature that has shown several benefits of liquidity creation. Banks allow consumers (Bryant, 1980; Diamond & Dybvig, 1983) and producers (Holmström & Tirole, 1998) to share liquidity risk. Banks also help overcome adverse selection problems (Dang et al., 2017; Gorton & Pennacchi, 1990) and thereby produce safe claims that satisfy a demand for safety (Stein, 2012). These functions may be further supported by deposit insurance (Hanson et al., 2015), the ability of banks to monitor borrowers (Diamond, 1984; Holmstrom & Tirole, 1997), and store wealth (Donaldson et al., 2018). While this previous literature establishes a theoretical relation between liquidity creation, investment and growth, our model contributes by showing how the positive effect of liquidity creation on investment may be weakened by low asset tangibility, consistent with the empirical findings.

The paper is organized as follows. The next section describes the construction of our measures of liquidity creation, and provides data sources and summary statistics for our sample of 100 countries. Section 3 first discusses the relation between liquidity creation and GDP per capita based on correlations, then presents our panel model results at both country and industry levels and robustness checks. Section 4 shows our empirical results on the investment channel through which liquidity creation affects growth. Section 5 develops a model clarifying the investment channel. Section 6 concludes the paper.

2 Liquidity creation around the world

In this section, we present our measures of liquidity creation. We also compare our liquidity measures for US banks with those of Berger and Bouwman (2009).

2.1 Liquidity creation measures

To estimate how much liquidity banks provide to the economy, we combine detailed financial and demographic information from BvD/Fitch Bankscope. Our worldwide sample focuses on all commercial, savings, and cooperative banks that were in business during some period between 1987 and 2014. We discard data after 2014 because they have undergone important changes resulting from the termination of contract between BvD and Fitch. This choice is in line with recent studies using Bankscope data (Silva, 2019). Also following Silva (2019), we rely on information at the most disaggregated level and avoid double-counting within the same bank by discarding consolidated entries if banks report unconsolidated data.⁵

We then build on the work of Berger and Bouwman (2009) in creating our measures of liquidity creation, which incorporate the contributions of all bank assets, liabilities, equity, and off-balance sheet activities. As it is recognized that banks create liquidity when they engage in certain activities but reduce liquidity when they engage in other activities, their measure classifies and weights all bank activities based on the liquidity they create or destroy. The Berger–Bouwman liquidity creation measure delivers a cash-denominated amount of liquidity which is provided by a bank to the economy. Formally, the liquidity creation (LC) measure for bank b operating in country c at time t is defined as the liquidity-weighted sum of all balance sheet items:

$$\begin{aligned} LC_{bct}=\sum _k \omega _{A k} A_{bckt} + \sum _k \omega _{L k} L_{bckt}, \end{aligned}$$

where $\omega _{A k}$ and $\omega _{L k}$ are the weights for classes k of assets A and liabilities plus equity L, respectively. The liquidity weights are assigned based on the ease, cost, and time for customers to withdraw liquid funds from the bank, and for banks to dispose of their obligations to meet these liquidity demands. There are three liquidity weights: liquid, semiliquid, and illiquid. Since liquidity is created when illiquid assets are transformed into liquid liabilities, both illiquid assets and liquid liabilities are given a positive weight. Following a similar logic, a negative liquidity weight is given to liquid assets, illiquid liabilities, and equity since liquidity is destroyed when liquid assets are transformed into illiquid liabilities or equity. Because liquidity creation is only half determined by the source or use of funds alone, we assign weights of $+\frac{1}{2}$ and $-\frac{1}{2}$. The intuition is that liquidity creation equals $1 when a dollar of liquid liabilities (such as demand deposits) is used to finance a dollar of illiquid assets (such as commercial loans) ($\frac{1}{2} \times \$1 + \frac{1}{2} \times \$1$). However, liquidity creation equals -$\$1$ when a dollar of illiquid liabilities (such as long-term funding) or equity is used to finance a dollar of liquid assets (such as cash or trading securities) ($- \frac{1}{2} \times \$1 + -\frac{1}{2} \times \$1$). An intermediate weight of 0 is also applied to activities that fall halfway between liquid and illiquid activities, that is, both semiliquid assets (such as residential mortgage loans) and liabilities (such as term deposits). Subsequently, we add up all weighted items of both sides of the bank balance sheet to yield the total amount of liquidity created by a bank in a particular year. All balance sheet items are converted into $ millions. Besides measuring how much banks create liquidity on the balance sheet, we also assess how much liquidity they create off-balance sheet. We apply the same principles to off-balance sheet items (such as committed credit lines), which are classified and weighted consistently with those assigned to functionally similar on-balance sheet activities. Since the granularity of the data is different in Bankscope and the Call Reports used in Berger and Bouwman (2009), we accordingly adapt their classifications.⁶ Internet Appendix A presents a detailed overview of our classifications and weights.

The advantage of a liquidity creation measure over a banking sector size measure is that it captures the full spectrum of banking activities on both sides of the balance sheet, as well as off-balance sheet activities. On the asset side, bank loans provide funding for long-term investments, while credit lines and other forms of off-balance sheet commitments provide funding liquidity to firms. On the liability side, bank deposits serve as safe and liquid transaction medium that form the core of the payment infrastructure in modern economies. Therefore, we use a liquidity creation measure that combines asset and liability side activities. It is important to stress that the measure of liquidity creation looks beyond deposits and traditional loans to also include non-deposit borrowings by banks as well as securities and other forms of investment on the asset side; this in addition to off-balance sheet activities of banks.

The following two examples illustrate the difference between a broader size-based measure and the liquidity creation measure used here. First, consider a fully equity-funded bank that makes long-term loans. This bank would contribute to the overall banking sector size as measured by outstanding bank credit. However, it does not create liquidity and could be replaced by any other equity-financed intermediary. Second, consider a narrow bank that offers demand deposit contracts but only holds reserves. This bank can provide payment services and would contribute to banking sector size as measured by the amount of outstanding bank deposits. But it again does not create liquidity because the same liquidity service could also be provided by a central bank, issuing fiat or digital money. Therefore, a liquidity creation measure combining characteristics of both sides of banks’ balance sheets is better suited for capturing the impact of bank activities on economic growth.

2.2 Summary statistics

Table 1 shows summary statistics on our bank-level liquidity creation measures. Panel A shows summary statistics for the whole sample of banks in 100 countries, while Panel B splits the sample of banks by total assets (using $1 billion and $3 billion cutoffs to define medium and large banks, respectively). We note that for our regression analysis, we will aggregate these bank-level liquidity creation measures to the country level. Panel C shows the mean on- and off-balance sheet liquidity creation for six individual years throughout our sample period.

Table 1 Summary statistics on bank-level liquidity creation in 100 countries (1987-2014)

Our sample contains 18, 217 banks and 199, 812 bank-year observations (Panel A of Table 1). The mean value of liquidity creation (on- and off-balance sheet) of a bank in our sample is $\$985$ million per year. Total liquidity creation culminated to $\$16.1$ trillion globally in 2014, broken down as follows (untabulated): $\$11.4$ trillion on-balance sheet liquidity creation and $\$4.6$ trillion off-balance sheet liquidity creation. These numbers highlight the importance of considering off-balance sheet activities. Indeed, on average only $61.2\%$ of an individual bank’s liquidity creation is on-balance sheet (mean of $\$603$ million, see Panel A). Both the between and the within standard deviations suggest considerable heterogeneity in the degree of liquidity creation across individual banks and over time.

Panel B of Table 1 shows summary statistics of our liquidity creation measure by bank size. Large banks only consist of $15.3\%$ of our sample, but they create the vast majority ($71.7\%$) of total liquidity (mean of $\$4,630$ million). Medium banks comprise $10.0\%$ of our sample and create $8.9\%$ of total liquidity. Small banks account for the remaining $19.4\%$ of total liquidity creation, despite the fact that they constitute the bulk of our sample of banks ($75.1\%$). This pattern is consistent with Berger and Bouwman (2009).

Panel C of Table 1 reports the mean value of liquidity creation in various years. Liquidity creation (on- and off-balance sheet) steadily increases over our sample period, with a mean of $\$674$ million in 1989 and of $\$1,787$ million in 2014. This yearly comparison in Panel C has to be considered with caution as our sample is strongly unbalanced (limited country coverage in 1989) and our numbers are in current US$. Our conclusion that liquidity creation has grown rapidly over time is, however, an overall pattern observed within our sample countries. As an example, Fig. 1 exhibits the evolution of our liquidity creation measures for the US. Total liquidity created by the US banking sector has clearly increased over time—though not monotonically – and reached almost $\$4.8$ trillion in 2014.

Fig. 1

Liquidity creation in the US (1999-2014). This figure shows the amount (in $ billion) of liquidity created by virtually every bank in the US from 1999 to 2014. The solid line represents total liquidity creation, while the dot line is on-balance sheet liquidity creation and the dash-dot line is off-balance sheet liquidity creation. We refer to Internet Appendices A and B for details about the variables

Although it is useful to compare our estimates of the amount of liquidity created by the US banking sector with the annual US statistics produced by Berger and Bouwman (2009), the comparison is complicated by several data limitations. First, Berger and Bouwman rely on the Call Reports which, unlike our data, are on a consolidated basis. Second, there are differences between Bankscope and the Call Reports in the breakdowns of both loan and deposit categories as well as off-balance sheet items. Third, we note that our sample periods only overlap for a few years; Berger and Bouwman (2009) focus on 1993–2003, while our US data cover 1999–2014. Fourth, we also include savings and cooperative banks (besides commercial banks) to account for the specificity of the banking sector in some countries (such as in Germany). With these caveats in mind, we can compare our aggregate statistics. For the year 2003, for instance, Berger and Bouwman (2009) report $\$2.843$ trillion of total liquidity created by 6969 commercial banks (using the measure they label catfat), while we have $\$3.299$ trillion of liquidity created (on- and off-balance sheet) by 7539 banks. Our investigations reveal that our numbers tend to underestimate on-balance sheet liquidity creation somewhat and overestimate off-balance sheet liquidity creation relative to Berger and Bouwman (2009). However, our numbers remain broadly comparable.⁷

3 Liquidity creation and economic growth

In this section, we provide our empirical results on the relation between liquidity creation and economic growth at the country and industry levels.

3.1 Main variables and initial assessment

We first introduce our main variables and then present preliminary assessments of the relation between liquidity creation and growth. As our main dependent variable, we use the log GDP per capita in current US$, sourced from the World Bank Development Indicators. These data on GDP are available for our annual panel of 100 countries between 1987 and 2014, which make up our baseline sample. Additional dependent variables used include investment (fixed and inventory) and the total number of patents, also drawn from the World Bank Development Indicators, and the net (in)tangible investment rate, calculated from the KLEMS database. We aggregate the annual bank-level data to annual country-level data of liquidity creation and divide by population to obtain our main explanatory variables of interest: on- and off-balance sheet liquidity creation per capita at the country level. We follow Berger and Sedunov (2017) in using population as scaling factor for liquidity creation.⁸ All other variables are discussed when they are introduced in the analysis, and are summarized and defined in Internet Appendix B.

Table 2 contains summary statistics for the key variables. The median GDP per capita is $\$5983$, with a significant dispersion (standard deviation of $\$18,512$). The median country-level on-balance sheet liquidity creation per capita is $\$556$, with a high standard deviation of $\$5957$. Off-balance sheet liquidity creation per capita is smaller, with a median of $\$151$ (standard deviation of $\$4622$). Figure 2 presents a map of worldwide liquidity creation (on- and off-balance sheet) per capita. The correlation between the two liquidity creation variables is 72.4% (not tabulated). Both variables are also positively but considerably less than perfectly correlated with private credit to GDP, with correlations of 57.6% and 35.2% with on- and off-balance sheet liquidity creation per capita, respectively.

Table 2 Summary statistics for the main variables used in the country-level analysis

Fig. 2

Total liquidity creation around the world. This figure shows the worldwide distribution of total liquidity creation per capita. Total (i.e., on- and off-balance sheet) liquidity creation is averaged over the full sample period for each sample country. We refer to Internet Appendices A and B for details about the variable and Table C.1 for the list of countries included

For a preliminary look at the data, in Figure C.1 in the Internet Appendix, we plot the relation between on-balance sheet liquidity creation and GDP per capita for the 100 countries in our sample in four different years: 1999, 2004, 2009, and 2014. The graph shows a strong positive correlation between these two variables, regardless of the year (and phase of the business cycle) considered. If we take the whole sample (not restricted to a specific year), the correlations are also strong: The pairwise correlation coefficients (untabulated) between the two variables in levels and first differences are 0.82 and 0.22, respectively, and in each case statistically significant at the 1-percent level. The picture is fairly similar for off-balance sheet liquidity creation, as shown in Figure C.2 in the Internet Appendix.

These associations are consistent with our premise that liquidity creation (on- and off-balance sheet) represents an important factor fueling economic activity. Of course, these simple cross-country correlations are not evidence of a causal relation and may reflect other relevant country differences. Therefore, we now turn to a formal analysis aimed at identifying the growth effect of liquidity creation.

3.2 Panel estimates at the country level

To formally evaluate the effect of liquidity creation on growth, we estimate the following linear dynamic panel model (similar to, e.g., Acemoglu et al., 2019):

$$\begin{aligned} y_{ct}= \beta LC_{ct} + \sum ^{p}_{j=1} \gamma _j y_{ct-j} + \delta X^{'}_{ct} + \alpha _c + \alpha _t + \varepsilon _{ct}, \end{aligned}$$

(1)

where $y_{ct}$ is the log of GDP per capita in country c in year t and $LC_{ct}$ is the log of liquidity creation per capita (on- or off-balance sheet) in country c in year t. We include p lags of log GDP per capita to control for the dynamics of GDP. $X^{'}_{ct}$ is a vector of controls that prior studies show to be related to growth, including democracy (Acemoglu et al., 2019; Papaioannou & Siourounis, 2008), inflation (Barro, 1997), and private credit to GDP (Beck et al., 2000). To account for non-linearities in the relation between private credit and growth (Arcand et al., 2015), we also add the squared term of private credit. Furthermore, we include total bank assets per capita to capture the size of the banking sector. We select these controls to account for important time-varying determinants of growth, while aiming to preserve sample size. We also examine the robustness of our results (in Sect. 3.3) to a wide range of other determinants that could confound the effect of liquidity creation on GDP per capita. $\alpha _c$ and $\alpha _t$ represent a full set of country and year fixed effects, controlling for time-invariant country characteristics and global trends, respectively. The error term $\varepsilon _{ct}$ captures all other time-varying omitted influences. Throughout, we report standard errors clustered at the country level.

The coefficient of interest is $\beta$, which measures the short-run effect of liquidity created both on and off the balance sheet on GDP per capita. The long-run effect of liquidity creation on GDP per capita can be derived by dividing the parameter estimate ${\hat{\beta }}$ by $1-\sum _{j=1}^{p} {\hat{\gamma }}_j$, the estimates of the p lags of the dependent variable. Our empirical strategy thus differs from much of the finance and growth literature, which typically averages out data over a five-year or longer horizon. Smoothing out along the time-series dimension of our panel would remove useful variation from the data (on average, each country is observed 14.1 years), which helps to identify the parameters of interest with more precision while distinguishing long-run from short-run growth relations. However, below we still make sure our results are not purely driven by (short-run) business cycle effects by examining the relation between liquidity creation and GDP per capita using five-year overlapping growth spells.

Columns 1–4 of Table 3 report the within estimates of Eq. (1) for our whole sample of 100 countries (listed in Table C.1 in the Internet Appendix).⁹ Column 1 represents the most parsimonious specification, with as independent variables our measure of on-balance sheet liquidity creation, a single lag of GDP per capita, and the fixed effects. The within estimate of $\beta$ is 0.017 ($s.e.=0.005$), statistically significant at the 1-percent level. We also find a significant degree of persistence in GDP, with a coefficient on lagged GDP per capita of 0.836 ($s.e.=0.024$). The persistence of GDP is an overall pattern in all results that we present. Column 2 reports the same specification as in column 1 with the further addition of time-varying controls. The effect of the persistence of GDP is slightly lower, but still highly significant, with a coefficient of 0.805 ($s.e.=0.026$). The within estimate for the coefficient on liquidity creation is higher (0.023) than in column 1 and statistically significant at the 1-percent level ($s.e.=0.006$). The coefficients on all controls display the sign expected based on the prior literature. In particular, we find that the effect of democracy is positive and that the effect of inflation is negative and significant. The effect of private credit to GDP is positive and significant, while its square has a negative and significant coefficient. The finding that both liquidity creation and private credit are significant and positive suggests that they capture two different dimensions of banking sector development, where private credit to GDP captures the intermediation dimension (pooling of resources and channeling them to enterprises), while liquidity creation captures the liquidity transformation dimension.¹⁰ As the relation between liquidity creation per capita and GDP per capita might be driven by larger countries having larger banking sectors, we also control for total bank assets per capita, of which the effect appears positive though insignificant.

Table 3 Effect of liquidity creation on (log) GDP per capita

Column 3 of Table 3, which is our preferred specification, adds a second lag of GDP per capita. The implied dynamics are now richer, with the first lag larger and still positive and the second lag smaller and negative. The overall extent of persistence of GDP thus remains close to that found in previous columns.¹¹ The coefficient on on-balance sheet liquidity creation is again statistically and economically meaningful. The estimate of $\beta$ in Eq. (1), 0.024 ($s.e.=0.006$), implies that a 10-percent increase in on-balance sheet liquidity creation per capita increases GDP per capita by $0.23\%$ in the short run (recall that we have a log-log model). Our dynamic panel model also fully specifies how the effects of liquidity creation by banks unfold over time. From the estimates in column 3, we find that a permanent increase in on-balance sheet liquidity creation per capita of 10% increases GDP per capita by $1.11\%$ in the long run ($= 0.024 / (1 - 0.932 - (-0.149))$). In the remainder of the paper, we will focus on the specification with two lags of GDP per capita as our baseline model.

The amount of liquidity created by banks off the balance sheet is also potentially consequential for growth. In column 4, we use the same specification as in column 3 and add our measure on off-balance sheet liquidity creation. The effect of liquidity creation off the balance sheet is significant, though quantitatively smaller than the effect of liquidity creation on the balance sheet. The within estimate is 0.008 ($s.e.=0.004$), implying a $0.36\%$ increase in GDP per capita in the long run following a permanent 10-percent increase in off-balance sheet liquidity creation. The other parameter estimates in column 4, including the coefficient on on-balance sheet liquidity creation, are very similar to the ones reported in column 3.

The within estimates of the dynamic panel model in columns 1–4 suffer from an asymptotic bias of order 1/T, which is known as the Nickell (1981) bias. The Nickell bias results (by construction) from the correlation between the lagged dependent variables and the country fixed effects, making the within estimator inconsistent. However, the Nickell bias only vanishes as T tends to infinity. Arellano and Bond (1991) derive a consistent GMM estimator for the parameters of the dynamic panel model for finite T. The idea of GMM is to take the first difference of Eq. (1) to eliminate country fixed effects and time-invariant country characteristics. The orthogonal relation between the lagged values of the dependent variable and the new differenced error term then constitutes the moment conditions of the GMM procedure. This holds under the null hypothesis of serially uncorrelated $\varepsilon _{ct}$, which can be evaluated by testing for second-order serial correlation in the first-differenced residuals. In columns 5 and 6 of Table 3, we provide estimates from the specifications of columns 3–4, using this GMM procedure. The GMM estimates of on-balance sheet liquidity creation are larger than the within estimates. However, the GMM estimates uncover a smaller degree of persistence of the GDP process, which in turn results in somewhat smaller estimated long-run effects. In column 6, the GMM estimate of off-balance sheet liquidity creation is still positive but fails to be significant at conventional levels. The coefficients on the controls in columns 5–6 are overall similar to those in columns 1–4.¹²

In addition, the bottom of Table 3 shows the p-values of tests for serial correlation in the first-differenced residuals (recall that the first-differencing is because the Arellano and Bond (1991)’s estimator takes first differences). The first-differenced residuals exhibit an insignificant second-order serial correlation, lending confidence to our GMM estimates reported in columns 5 and 6.¹³

To further explore the long-run effect of liquidity creation on growth, we now exploit longer growth spells (in the same vein of, e.g., Bekaert et al., 2005). That is, we specify a version of Eq. (1) using as dependent variable the five-year average log GDP per capita over t to $t+4$ (${\bar{y}}_{ct,t+4}$), taking liquidity creation per capita at t ($LC_{ct}$), and controlling for the five-year average log GDP per capita over $t-5$ to $t-1$ (${\bar{y}}_{ct-5,t-1}$). The model is thus given by:

$$\begin{aligned} {\bar{y}}_{ct,t+4}= \beta LC_{ct} + \gamma {\bar{y}}_{ct-5,t-1} + \delta X^{'}_{ct} + \alpha _c + \alpha _t + \varepsilon _{ct}. \end{aligned}$$

(2)

The estimates reported in Table C.2 in the Internet Appendix show that liquidity creation has a positive effect on longer-horizon overlapping growth spells, which indicates that liquidity creation by banks is associated with long-run growth beyond just business cycle effects.

In sum, the results in this subsection indicate a statistically significant and economically sizable effect of both on- and off-balance sheet liquidity creation on GDP per capita. In the next subsection, we further examine the robustness of our baseline result.

3.3 Additional country-level controls

The validity of our estimates of the effect of liquidity creation on GDP per capita presented so far may be sensitive to the presence of time-varying determinants that simultaneously impact liquidity creation and GDP per capita. Table 4 shows the results from the same specification as in column 4 of Table 3 with the inclusion of further covariates. To save space, we only report the within estimates of the coefficients on the liquidity creation measures and the controls of interest. The corresponding Arellano and Bond’s GMM estimates produce very similar results (not tabulated).

Table 4 Effect of liquidity creation on (log) GDP per capita adding additional controls

In columns 1–3 of Table 4, we report results from specifications in which we include variables proxying for important regulatory changes that may affect growth directly or indirectly through their impact on financial development. We control for creditor rights in column 1 (as suggested by the findings of Levine, 1999; Djankov et al., 2007, among others), the use of macroprudendital policies in column 2 (Cerutti et al., 2017), and the equity market liberalization process in column 3 (Bekaert et al., 2005). Although these specifications tend to substantially reduce the number of observations due to missing data and although some of these additional controls enter the model with a significant coefficient, the liquidity creation effects continue to be statistically and economically significant.

Liquidity creation may also be driven by several attributes of countries’ financial system, not already absorbed by the country fixed effects. In particular, Cetorelli and Gambera (2001) provide evidence of a general depressing effect on growth associated with a concentrated banking sector. Langfield and Pagano (2016) find that an increase in the size of the banking sector relative to financial markets is associated with lower economic growth, especially during housing market busts. Another strand of the literature shows that stock markets are associated with greater economic growth (Levine & Zervos, 1998). In columns 4–7 of Table 4, we control for these attributes and find that they have a limited impact on our coefficients on liquidity creation—though they generally exhibit the expected impact on GDP per capita.

Furthermore, countries with good institutions, deepened economic and political integration, and a high human capital stock may be more likely to exhibit strong economic growth. Therefore, another concern is that the estimated effects of country-wide liquidity creation are a mere reflection of countries’ level of development. To assuage this concern, we add covariates that control for the potential effect of rule of law (in column 8), globalization (in column 9), and secondary school enrollment (in column 10). Reassuringly, controlling for these growth determinants yields qualitatively similar results to our baseline results.

Further tests (not tabulated to conserve space but available upon request) show that our results are also robust to using many other country determinants included in growth models, such as population growth, life expectancy, fertility rate, openness to trade, black market premium, government consumption (see Barro, 1997; Barro & Sala i Martin, 2003). By enhancing liquidity creation, bank capital is also another potential omitted factor in our growth model. Therefore, we also account for the direct effect of bank capital by aggregating annual bank-level data to annual country-level data of the equity-to-asset ratio. Controlling for bank capital (untabulated) hardly affects our baseline results in Table 3. Similarly, the cost of capital could drive both liquidity creation and growth. Our baseline results are also robust to controlling for different interest rate measures as proxies for the cost of capital (untabulated).

All in all, the robustness checks presented in this subsection bolster confidence that our results are not driven by the impact of confounding factors. The function of banks as liquidity creator seems to be more than just another aspect of more general financial or economic development, and thus worthy of more detailed study.

3.4 Panel estimates at the industry level

While the country-level results in the previous subsections suggest that liquidity creation helps to enhance economic growth, endogeneity concerns remain, related to both reverse causality and omitted variable bias. In this subsection, we therefore proceed to examine the differential impact of liquidity creation on the value added of industries that vary in their reliance on debt financing, in a “smoking gun” approach to identification.

We follow Rajan and Zingales (1998) to identify the effect of country differences in liquidity creation on growth by industry.¹⁴ The identifying assumption is that dependence on external financing differs across industries for structural reasons. We focus, however, on dependence on debt finance instead on external finance (debt plus equity) since our focus is on the banking sector, which provides debt rather than equity financing.¹⁵ Allen and Gale (1999) argue that equity financing is more adequate for projects with higher uncertainty, while debt financing more adequate for less risky projects. Several studies document that innovative investments such as R &D rely relatively more on equity financing and are not easily financed with debt (Brown et al., 2012, 2013; Hall & Lerner, 2010).

We estimate the equilibrium dependence on debt financing using Compustat data for listed US firms as in Rajan and Zingales (1998), measured as the ratio of net debt issuance to capital expenditures (see Internet Appendix B). We then aggregate the firm-level Compustat data by industry i in each year t, and take an average over time to obtain an industry-level measure of debt dependence $DD_i$. We use this US-based ratio as a proxy for the structural share of investment that is financed by debt in industries around the world. While a higher share of corporate debt is market-financed in the US than in other countries, we take debt financing by large US firms as benchmark for the inherent demand for debt finance across different industries, under the assumption that large firms in the US face a flat supply curve for corporate debt and industry differences are thus explained by inherent demand variation.

We obtain industry-level data on value added and gross output from the OECD (STAN database), which are available for a subset of 33 of the countries (mostly OECD member countries) in our sample. Table C.1 in the Internet Appendix lists the countries, which excludes the US as in Rajan and Zingales (1998), and Table C.3 reports the industries together with summary statistics on our measures of debt dependence and output.¹⁶ All regressions using debt dependence are carried out on this restricted sample.¹⁷

The model we estimate is given by:

$$\begin{aligned} Y_{cit} = \kappa LC_{ct} \times DD_{i} +\theta X^{'}_{ct} \times DD_{i} + \alpha _{ct} + \alpha _{it} + \varepsilon _{cit}, \end{aligned}$$

(3)

where $Y_{cit}$ denotes the log of value added or the log of gross output in country c in industry i at time t. As before, $LC_{ct}$ is one of the liquidity creation measures (in log per capita) in country c at time t. $DD_{i}$ is debt dependence of industry i. The specification contains the same vector of time-varying country controls as before ($X^{'}_{ct}$) interacted with debt dependence ($DD_{i}$). Furthermore, the specification includes country-year fixed effects ($\alpha _{ct}$) to control for time-varying country shocks and industry-year fixed effects ($\alpha _{it}$) to control for time-varying industry shocks. We note that, because LC varies at the country-year level and DD at the industry level, their individual effect is absorbed by the country-year fixed effects and industry-year fixed effects, respectively.¹⁸ In all specifications, we cluster standard errors at the country-year level to account for the within country-year correlation across industries.

The coefficient of interest is $\kappa$, which is identified from the within-country, cross-industry variation in debt dependence. It measures the effect of liquidity creation on output in industries with high debt dependence in a country (for a given amount of liquidity created) compared to industries with low debt dependence in the same country.

Table 5 presents the results from estimating Eq. (3).¹⁹ In column 1, we use value added as dependent variable, with the full set of fixed effects and interaction terms. The estimate on the interaction, $\kappa$, is 0.465 ($s.e.=0.063$), statistically significant at the 1-percent level. In column 2, we add off-balance sheet liquidity creation interacted with debt dependence. On-balance sheet liquidity creation continues to be positively associated with value added in industries more reliant on debt financing, with a coefficient estimate of 0.434 ($s.e.=0.075$). However, off-balance sheet liquidity creation does not enter significantly. This estimate implies that a 10-percent increase in liquidity creation is associated with 1.22% higher value added in the average industry with a debt dependence of 29.4% ($= \ln (1.1) \times 0.434 \times 0.294$), consistent with the results we found in Table 3. To illustrate the economic effect of the interaction of liquidity creation and debt dependence, consider two industries: one at the 75th percentile of debt dependence (Agriculture, forestry, and fishing) and one at the 25th percentile (Transportation and storage). The difference in debt dependence between the two industries is 0.156 ($=0.356-0.200$). A 10-percent increase in liquidity creation increases value added in the high-debt-dependence industry by 65 basis points more than in the low-debt-dependence industry ($= \ln (1.1) \times 0.434 \times 0.156$). In columns 3 and 4, we replicate these tests using gross output as dependent variable. The results are very similar, both statistically and economically.

Table 5 Effect of liquidity creation on output: STAN industry-level data

The results in this subsection indicate that on-balance sheet liquidity creation has a larger effect on output in industries that are relatively more in need of debt financing, corroborating our regression results at the country level. However, we do not find evidence of a significant impact of off-balance liquidity creation on industry-level economic activity. More importantly, identifying a differential relation between liquidity creation and growth across industries helps us mitigate endogeneity concerns that arise in the cross-country aggregate analysis.

4 Investment channel

So far, we have established that liquidity creation is positively and significantly associated with economic growth at both country and industry levels. Since liquidity creation adds economic value by converting illiquid long-term assets into short-term liquid liabilities, we hypothesize that the positive effect of liquidity creation on growth is driven by investment. In this section, we empirically evaluate this investment channel using data at both the country and industry levels. We distinguish between tangible and intangible investment, motivated by recent studies suggesting that banks are better at supporting tangible investment rather than intangible investment and innovation (e.g., Brown et al., 2013; Hsu et al., 2014; Dell’Ariccia et al., 2020). It is important to stress that measuring innovation and intangible investment is notoriously difficult. We therefore use an array of different measures and data sources to test the robustness of our findings. In Sect. 5, we provide further theoretical backing for our focus on investment by developing a model of liquidity creation, tangible and intangible investment.

4.1 Country-level evidence

We begin by investigating whether the investment component of GDP is affected by liquidity creation. We thus estimate similar models to Eq. (1), except that the dependent variable is one of the components of investment expenditures (i.e., fixed investment and inventory investment) and we control for lags of the corresponding dependent variable (instead of lags of GDP per capita as in Table 3) on the right-hand side. Consistent with our previous exercises, we scale each investment component of GDP by population and take the logarithm (see Internet Appendix B).

Table 6 shows the within estimation results for both fixed investment and for inventory investment. The results from Arellano and Bond’s GMM estimations are very similar (not tabulated to preserve space). We find that liquidity creation (on- and off-balance sheet) significantly affects investment through fixed investment expenditures (columns 1–2), consistent with our prediction that banks are more likely to finance fixed (tangible) assets.²⁰ The controls enter with the same signs as in Table 3. For inventory investment, we find a relation with off-balance sheet liquidity creation but not with on-balance sheet liquidity creation (columns 3–4), consistent with the notion that firms tend to use credit lines for working capital management.

Table 6 Effect of liquidity creation on (log) investment types

Columns 5-6 of Table 6 consider country-wide patent applications as a key outcome of intangible (and human capital) investment. The coefficients on on-balance sheet liquidity creation are positive but fail to be significant at conventional levels, suggesting that liquidity creation does not appear as a factor encouraging patenting. Although these results on patent applications are in line with our prediction, note that our patent variable is a rather coarse measurement of the extent of patenting in a country, and that there may be a considerable lag from the innovation stage to the patent application stage that our specifications may not adequately account for.

We therefore move on to assembling data that allow us to better distinguish tangible from intangible fixed investment. Following Döttling et al. (2017), we use two variables to measure the net investment rate, that is, the gross investment rate minus the depreciation rate. We define the gross investment rate as the ratio of gross fixed capital formation to lagged fixed assets and the depreciation rate as the ratio of gross fixed capital consumption to lagged fixed assets. We then use granular data on asset types (aggregated at the country level), sourced from the KLEMS database, to construct our measures of both tangible and intangible investment rate.²¹ KLEMS offers a great level of detail, but only covers a subset of 22 countries (listed in Table C.1 of the Internet Appendix) in our sample.

We estimate the effect of liquidity creation on the net investment rate using the following model:

$$\begin{aligned} I_{ct}= \beta LC_{ct} + \delta X^{'}_{ct} + \alpha _c + \alpha _t + \varepsilon _{ct}, \end{aligned}$$

(4)

where $I_{ct}$ represents our measures of the net investment rate in tangible and intangible assets and $LC_{ct}$ our measures of liquidity creation (in log per capita) in country c at time t. As in standard investment regressions (Carlin & Mayer, 2003), we do not assume persistence in the investment process (i.e., by lagging the dependent variable). All the other variables, parameters, and subscripts are defined as before. Standard errors are clustered at the country level.

Table 7 reports the results from estimating Eq. (4), using as dependent variables the net investment rate both in tangible assets (in columns 1–2) and in intangible assets (in columns 3–4). The number of observations is considerably lower in these specifications because of more limited data coverage. The table reveals that on-balance sheet liquidity creation is associated with an increase in the net tangible investment rate (columns 1–2), but not in the net intangible investment rate (columns 3–4). The effect of on-balance sheet liquidity creation on the net tangible investment rate is statistically significant at the 1-percent level and is also economically sizable. To be able to compare the magnitude of the effect of liquidity creation on tangible and intangible investment, Table 7 reports standardized coefficients. In columns 1–2, a one standard deviation increase in on-balance sheet liquidity creation is associated with a 0.876–0.949 standard deviations higher net tangible investment rate. In contrast, the effect on the net intangible investment rate is not statistically significant. We also do not find significant estimates for off-balance sheet liquidity creation in columns 2 and 4.

Table 7 Effect of liquidity creation on net investment rate: KLEMS country-level data

One explanation for the contrasting results on tangible versus intangible investment can be found in the law and finance literature that shows the importance of an efficient legal system for financial intermediation and hence liquidity creation (Claessens & Laeven, 2003). This expansive literature suggests that legal system efficiency is relatively more important for intangible than tangible assets, as the former are easier to divert and have lower liquidation value than the latter, thus providing less protection for creditors if used as collateral. This mechanism is reflected in our model presented in Sect. 5 below, in which liquidation value and diversion play a key role.

Here, we test this hypothesis empirically by interacting liquidity creation with two measures of legal system efficiency, corresponding to the ability of investors to protect their rights and recover value during a bankruptcy procedure. Figure 3 plots the marginal effect of liquidity creation on (in)tangible investment for different levels of country-level investor protection and the recovery rate.²² The figure illustrates that the relation between on-balance sheet liquidity creation and investment is stronger in countries with higher investor protection and recovery rates and may even turn positive and significant for intangible investment at high levels of investor protection and recovery rates. Interestingly, the slope of the curve is steeper for intangible than for tangible investment, indicating that improvements in investor protection and recovery rates matter more for intangible than tangible investment. While tentative, this finding is consistent with the law and finance literature predicting that more effective creditor protection enhances financial intermediation and more so for intangible assets and, as such, corroborates our results on (in)tangible investment.

Fig. 3

Marginal effects of liquidity creation on net (in)tangible investment rate. This figure shows the marginal effects of on-balance sheet liquidity creation on tangible investment rate (left graphs) or on intangible investment rate (right graphs), calculated from predictions of a fit model at fixed values of a proxy for two key parameters of the model and averaging the remaining covariates. The proxy variable for the diversion parameter $\gamma$ is Investor protection index (top graphs) and the proxy variable for the liquidation cost $\mu$ is Recovery rate (bottom graphs). The x-axis represents the values of one of these proxy variables (the simple slopes), and the y-axis represents the effect on linear prediction (based on a version of Eq. (4) as reported in even-numbered columns of Table 7 but with the further inclusion of an interaction between our liquidity creation measures and the corresponding proxy variable for the parameter of the model, $\gamma$ or $\mu$). The vertical lines are 95-percent confidence intervals. We refer to Internet Appendix B for a full description of the variables and their corresponding sources and Table C.1 for the list of countries included

4.2 Industry-level evidence

The country-level results in the previous subsection suggest that liquidity creation by banks helps to boost tangible rather than intangible investment. To confirm and further understand these results, we conduct additional tests that exploit industry heterogeneity within countries, as in Sect. 3.4, thus also addressing endogeneity concerns arising from the aggregate cross-country analysis. Although we control for country fixed effects in Table 7, it may be possible that other factors that vary across countries over time may coincide with changes in liquidity creation. In such case, we would incorrectly attribute the changes in the net investment rate to changes in liquidity creation. Exploiting within-industry variation within countries while accounting for debt dependence in the spirit of Rajan and Zingales (1998) is a way to address these concerns.

We construct our measures of the net investment rate in both tangible and intangible assets at the industry level in the same way as the country-level measures, using KLEMS industry data. Table C.1 in the Internet Appendix lists the 14 countries included in our analysis, which excludes the US as it is the benchmark. Table C.3 in the Internet Appendix reports the 25 industries and gives summary statistics on our measures of debt dependence and investment.

The model we estimate is the same as in Eq. (3), except that the dependent variable is the net tangible investment rates in country c in industry i at time t. Table 8 presents results in terms of standardized coefficients. In column 1, we use the net tangible investment rate as dependent variable and include the full set of fixed effects and interaction terms. The estimate on the interaction, $\kappa$, is 1.115 ($s.e.=0.389$), statistically significant at the 1-percent level. To gauge the economic effect, we compare the industries at the 25th and 75th percentile of debt dependence. A one standard deviation increase in liquidity creation is associated with a 0.174 standard deviations greater increase in tangible investment in a high-debt-dependence industry (Agriculture, forestry, and fishing), compared to a low-debt-dependence industry (Transportation and storage) ($= 1.115 \times 0.156$). This difference represents about 20% of the average effect of on-balance sheet liquidity creation on tangible investment estimated at the country level in Table 7. This result suggests that the banking sector fosters tangible investment by providing liquidity to industries that rely more heavily on debt financing. In column 2, we augment the previous specification with the off-balance sheet liquidity creation measure interacted with debt dependence. The effect of on-balance sheet liquidity creation is of similar economic magnitude but slightly weaker than in column 1. We also find that off-balance sheet liquidity creation affects the net tangible investment rate more in industries that rely more on debt financing, though this effect is difficult to interpret given there is no effect of off-balance sheet liquidity creation on tangible investment at the country level (see Table 7). In columns 3–4, we instead use a time-varying measure of debt dependence, which allows us to include industry-country fixed effects. The results are robust.

Table 8 Effect of liquidity creation on net tangible investment rate: KLEMS industry-level data

In Table C.6 in the Internet Appendix, we repeat these tests with the net intangible investment rate as dependent variable. We find no consistent statistically significant effect of on-balance sheet liquidity creation. This result expands on the findings in Table 7. Not only does on-balance sheet liquidity creation have no effect on overall intangible investment at the country level, but there also appears to be no differential effect between industries that rely more on debt financing.

For robustness purposes, we also run this industry-level analysis using alternative data sources. In particular, we look at fixed investment and R &D expenditures as measures of fixed tangible assets and intangible investment, respectively. These industry-level variables are sourced from the OECD (STAN and ANBERD databases). We estimate the same model as in Eq. (3), with either fixed investment (see Table C.7 in the Internet Appendix) or R &D expenditures (see Table C.8 in the Internet Appendix) as dependent variable (in logs). The results are consistent with our main findings presented so far: on-balance sheet liquidity creation positively and significantly affects fixed investment more in industries that rely more on debt financing, while we find no consistent (and in some specifications even negative) coefficient on R &D expenditures. It is reassuring that results are consistent using KLEMS or OECD investment data, given the difficulty of measuring intangible investment.²³

4.3 Asset intangibility and country-level growth

Our empirical results on the investment channel indicate that liquidity creation by banks fosters tangible investment, but not intangible investment. To close the loop, we relate these results back to the growth results from Sect. 3 to confirm that the investment channel indeed feeds through to aggregate growth. Specifically, we test whether liquidity creation has a weaker effect on growth in countries with a higher share of industries relying on intangible assets.

To measure a country’s intangible intensity, we start by measuring an industry’s intangible-to-total capital ratio in the US using KLEMS data. We then aggregate the industry-level intangible-to-total capital ratio for each country, weighting by the industry’s value added of the respective country. The idea is that, in the spirit of Rajan and Zingales (1998), the capital composition of US industries reflects the technological frontier. Thus, this measure is based on an industry’s technological reliance on intangible capital, and therefore unlikely to be correlated with other factors that may be correlated with growth. We refer to this measure as a country’s intangible ratio. The intangible ratio varies between 0.62% and 18.28%, with a mean of $8.58\%$ and a standard deviation of $2.61\%$. The average intangible ratio across countries increases over time, from around 5% in the early 1990s, up to 10% in 2013, in line with a shift to intangible assets due to technological advances during this time (Corrado & Hulten, 2010).

We use this measure to examine whether liquidity creation by banks has a differential effect on growth in countries with more intangible assets. To do so, we specify a version of Eq. (1) including an interaction between liquidity creation and intangible intensity:

$$\begin{aligned} y_{ct}= \beta _1 LC_{ct} + \beta _2 \ LC_{ct} \times IR_{ct} + \beta _3 IR_{ct} +\sum ^{p}_{j=1} \gamma _j y_{ct-j} + \delta X^{'}_{ct} + \alpha _c + \alpha _t + \varepsilon _{ct}. \end{aligned}$$

(5)

The dependent variable, $y_{ct}$, is the log of GDP per capita in country c at time t. $LC_{ct}$ is again one of our liquidity creation measures (in log per capita) in country c at time t, while $IR_{ct}$ is our measure of intangible ratio of country c at time t. All the other variables, parameters, and subscripts are defined as in Eq. (1), and standard errors are clustered at the country level.

The coefficient of interest, $\beta _2$, measures the short-run effect of liquidity creation on GDP per capita conditioned on the country’s intangible intensity. Table 9 reports the within estimates (in columns 1–2) as well as the GMM estimates (in columns 3–4) of Eq. (5). Across columns, the coefficients on all variables are comparable to Table 3. The estimate of the key coefficient of interest ($\beta _2$) is negative and highly significant across columns 1–4, indicating that on-balance sheet liquidity creation has a weaker impact on GDP per capita in countries with more intangible capital.²⁴ The effect is economically meaningful. A one standard deviation increase in the intangible ratio ($=0.027$) reduces the short-run effect of liquidity creation on GDP per capita by more than 23.6% ($= -0.402 \times 0.027 / 0.046$), and the effect turns negative when the intangible ratio exceeds 11.4% ($=0.046/0.402$), which corresponds to the 90th percentile of the distribution or the average intangible ratio of Germany. Figure 4 illustrates that the relation between on-balance sheet liquidity creation and GDP per capita turns insignificant at an intangible ratio of 9% and negative (but insignificant) at 11%. Specifically, Fig. 4 plots the slope for GDP per capita on on-balance sheet liquidity creation while holding the value of intangible ratio constant at values running from 0 to 22% (all observations are between these two values). It also appears that the slopes are significant for most values of intangible ratio. Off-balance sheet liquidity creation interacted with intangible intensity does not enter significantly, underscoring the weaker effect of off-balance sheet liquidity creation on growth.

Fig. 4

Marginal effects of liquidity creation on growth. This figure shows the marginal effects of on-balance sheet liquidity creation on GDP per capita, calculated from predictions of a fit model at fixed values of intangible ratio and averaging the remaining covariates. The x-axis represents the values of the variable Intangible ratio (the simple slopes), and the y-axis represents the effect on linear prediction (based on column 2 of Table 9). The vertical lines are 95-percent confidence intervals. We refer to Internet Appendix B for a full description of the variables and their corresponding sources and Table C.1 for the list of countries included

Table 9 Differential effect of liquidity creation on (log) GDP per capita across countries’ intangible intensity

Taken together, our country- and industry-level findings in Sects. 3 and 4 suggest that liquidity creation, especially on the balance sheet, drives economic growth by increasing tangible investment. In contrast, liquidity creation does not appear to affect investment in intangibles (unless at very high levels of creditor right protections and bankruptcy recovery rates), and the overall effect of liquidity creation on growth is larger in countries with industries that rely to a lesser degree on intangible assets.

These results speak to a number of non-linearities that prior research has uncovered in the relation between the banking sector and economic activity, such as the non-linear relation between banking sector development and growth (Arcand et al., 2015), the differential relation of bank financing with tangible versus intangible investment (Aghion et al., 2004; Dell’Ariccia et al., 2020; Hsu et al., 2014), and the decreasing role of banks in stimulating growth as economies develop (Čihák et al., 2012). More research is needed to fully understand these non-linearities, but our paper contributes to this debate by showing how one key function of banks—liquidity creation—contributes to economic activity through its impact on tangible but not intangible investment. To illuminate why the (tangible) investment channel may explain our findings on the relation between banks’ liquidity creation and economic growth at the country and industry levels, we develop a theoretical model describing this channel in the next section.

5 A model of liquidity creation and investment

To elucidate potential theoretical channels behind our empirical results, we build a theoretical framework based on the seminal Diamond and Dybvig (1983) model. While the baseline Diamond-Dybvig model is an ideal workhorse model to study liquidity creation, in the model banks are passive on the asset side and hence cannot affect overall investment. We depart from the baseline model in that we introduce two new elements: (i) liquidation costs differentiated according to whether the investment is tangible or intangible and (ii) credit risk and a moral hazard problem that may limit the extent to which banks can reduce liquidity risk. To be able to speak to the empirical finding that liquidity creation by banks boosts growth primarily via tangible rather than intangible investment we assume that both elements are a function of asset tangibility. As we elaborate below, intangible assets have lower liquidation values and are more prone to moral hazard problems because they are more easily diverted.

5.1 Model setup

Consider an economy with three dates, $t=0,1,2$, and a unit mass of investors. Investors are ex-ante identical but privately learn at $t=1$ whether they are early (probability $\lambda$ ) or late types. Early investors need to consume at $t=1$, while late investors can wait until $t=2$. Utility is given by:

$$\begin{aligned} U = {\left\{ \begin{array}{ll} u(c_1), &{}\text { if early}, \\ u(c_1 + c_2), &{}\text { if late}, \end{array}\right. } \end{aligned}$$

where $c_t$ denotes consumption at time t and investors have log-utility $u(c_t)=\log (c_t)$.

Investors have an endowment e at $t=0$ and access to a short-term storage technology that transfers resources one-for-one across time and resembles cash. They also have access to a long-term investment project. This project is more productive than storage, but it is risky and incurs losses if liquidated early. Investing I units in this project at $t=0$ yields $1 + r$ at $t=2$ with probability p and $1-\mu$ otherwise. If the technology is liquidated at $t=1$ it generates $1-\mu$ with certainty. Thus, $\mu \ge 0$ is the liquidation loss incurred if the project fails or is liquidated early. We assume that $p > \frac{\mu }{r + \mu }$ to ensure the long-term project has a higher expected return than storage.

At the end of $t=2$, after the long-term investment project pays off, investors may divert a fraction $\gamma$ of the investment return for their own consumption. As shown below, this moral hazard problem may limit the extent to which liquidity creation by banks can support investment as late investors may be tempted to divert when their projects are successful.

The long-term investment is characterized by its asset tangibility k and we assume that both the liquidation cost and diversion parameter are negatively related to k:

Assumption 1

The liquidation cost $\mu$ and diversion parameter $\gamma$ are decreasing functions of investment tangibility k:

• $\mu = \mu (k)$ with $\mu '(k) \le 0$ and $\mu (k) \in [0,1]$

• $\gamma = \gamma (k)$ with $\rho '(k) \le 0$ and $\gamma (k) \in [0,1]$

Realistically, intangible assets are harder to liquidate than physical assets such as machines and plants, reflected in a higher liquidation loss $\mu$. Moreover, the value of intangible investments may be hard to assess by outsiders, making it easier for insiders to divert spending to pet projects. This means that intangible investments plausibly have a larger value $\gamma$ so more can be diverted. Characterizing the long-term investment in terms of the technological parameter k allows us to later relate our results to the empirical findings. However, for brevity we omit the dependence of $\mu$ and $\gamma$ on k through most of this section.

Finally, we assume that the liquidation cost $\mu$ is not too small to ensure that banks can play a role in overcoming liquidity risk and thereby improve investment.

Assumption 2

$$\begin{aligned} \mu \ge \frac{(1-{\tilde{\lambda }})r}{{\tilde{\lambda }} + (1-\lambda )r} \equiv \underline{\mu }, \end{aligned}$$

with ${\tilde{\lambda }} \equiv [\lambda + (1-\lambda ) (1-p)]$.

5.2 The role of banks in supporting investment

This subsection compares the optimal investment under autarky, defined as a situation in which each investor invests in isolation, to the allocation that can be achieved if banks’ liquidity creation facilitates risk sharing. Detailed derivations are relegated to Appendix D.

Under autarky, an investor who invests I in the long-term project and $s = e - I$ into storage, needs to liquidate the entire investment if he or she is an early type. In contrast, late investors optimally do not liquidate any investment at $t=1$ and may enjoy a high level of consumption at $t=2$ if their project succeeds. Under autarky, investors are thus exposed to liquidity risk (and credit risk).

Banks can improve upon autarky because they can reduce liquidity risks by avoiding inefficient liquidation of the long-term investment. In fact, Diamond and Dybvig (1983) show that banks can implement the first-best allocation by offering demand deposits contracts.

In Appendix D, we solve the investor problem and derive the allocation under autarky and in the first-best allocation. These allocations can be solved in closed form and readily compared:

Autarky	First Best (Bank)
$I^{aut} = \frac{(1-{\tilde{\lambda }}) r - {\tilde{\lambda }} \mu }{\mu r} e$	$I^* = (1-\lambda ) e$
$c_E^{aut} = \frac{{\tilde{\lambda }} (r+\mu ) e}{r}$	$c_E^* = e$
$c_L^{aut} = \frac{(1 - {\tilde{\lambda }}) (r+\mu ) e}{\mu }$	$c_L^* = R e$

Here, $c_E$ and $c_L$ denote the consumption levels of early and late investors, respectively. It is easy to verify that under Assumption 2 investment is higher in the first-best allocation, i.e., $I^* \ge I^{aut}$.²⁵ This implies that, if a bank can implement the first best allocation through demand deposit contracts, the model predicts a positive link between liquidity creation and investment—in line with our empirical results. Intuitively, in the first-best allocation liquidation costs are reduced because agents can share liquidity risks and reduce liquidation costs. This makes it more attractive to invest in the long-term asset.

The next subsection clarifies how demand deposit contracts can implement the first-best allocation and maps banks’ liquidity creation in the model to our empirical measure. We then analyze under what conditions demand deposit contracts can be implemented depending on the long-term asset’s (in)tangibility.

5.3 Demand deposit contracts and liquidity creation

Consider a bank that offers the following demand deposit contract. Investors deposit their endowment with the bank at $t=0$ in exchange for a demandable claim. This claim allows investors to withdraw $c_E^*$ at $t=1$ or $c_L^*$ at $t=2$. The bank then invests $s^*$ in storage and makes a long-term loan of size $I^*$ back to investors. The loan matures at $t=2$ and has an interest rate of r. Investors use the loan to invest $I^*$ in the long-term investment technology.²⁶

Under this contract, the bank holds a mix of illiquid assets (loans) and liquid assets (cash), funded by short-term, liquid deposits. Investors are no longer exposed to liquidity risk because they finance the long-term investment with a long-term loan, while also holding demand deposits that allow them to consume early if desired without liquidating the long-term asset. That is, banks create liquidity.

To see more directly how the model maps to our empirical measure of liquidity creation, recall that deposits are classified as liquid liabilities and loans are classified as illiquid (or semiliquid in the case of mortgages; see Appendix A). Using the liquidity weights from our empirical measure, the bank creates liquidity of

$$\begin{aligned} LC^* = \underbrace{\frac{1}{2} e}_{\text {Liquid liabilities (deposits)}} + \underbrace{\frac{1}{2} I^*}_{\text {Illiquid assets (loans)}} \end{aligned}$$

In contrast, under autarky there is no liquidity creation because investors operate in isolation ($LC^{aut} = 0$). Therefore, comparing the allocation under autarky to the allocation that can be achieved with demand deposits is akin to comparing an economy with little liquidity creation to an economy with much liquidity creation, underlining that the contracting arrangement described above resembles banks’ liquidity creation while highlighting the relevance of both asset-side and liability-side activities.²⁷

5.4 Incentive compatibility and asset tangibility

Under what conditions can the bank implement the first-best allocation through demand deposit contracts? For demand deposits to be incentive-compatible, two conditions have to be satisfied. First, since an investor’s type is private information, late investors have to prefer not withdrawing early. At $t=1$ late investors prefer waiting until $t=2$ if

$$\begin{aligned} c_L^* \ge c_E^*. \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad ({IC_1}) \end{aligned}$$

This first incentive-compatibility condition is standard to Diamond-Dybvig type models and ensures that the payoff from waiting ($=c_L^*$) is larger than that from withdrawing early ($=c_E^*$).

Second, late investors may divert a fraction $\gamma$ of the long-term asset’s return instead of repaying their loan.²⁸ This temptation is particularly high if an investor’s asset succeeds, in which case they can divert $\gamma (1+r) I^*$. Upon observing diversion the bank can penalize investors by seizing their deposit as well as the remaining $(1-\gamma )$ of the project’s payoff. Thus, by diverting late investors give up the deposit claim worth $c_L^*$. Combining the two, a second incentive-compatibility condition requires that

$$\begin{aligned} c_L^* \ge \gamma (1+r) I^*. \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad ({IC_2}) \end{aligned}$$

This condition requires that the payoff from diverting ($= \gamma (1+r) I^*$) does not exceed the value of the deposit claim ($=c_L^*$). Thus, if (${IC_2}$) is satisfied, late investors prefer repaying the loan and holding on to their deposit claim over diverting.

It is straightforward to verify that the first incentive-compatibility constraint is satisfied, so that late investors wait and consume at $t=2$. However, the bank can only implement the first-best allocation if not only (${IC_1}$) but also (${IC_2}$) is satisfied.

Proposition 1

There exists a threshold ${\bar{k}}$ characterized by

$$\begin{aligned} \mu ({\bar{k}}) = 1 - \frac{(1+r)[(1-\lambda ) \gamma ({\bar{k}}) - p]}{1-p} , \end{aligned}$$

such that if and only if

$$\begin{aligned} k \ge {\bar{k}}, \end{aligned}$$

(6)

then (${IC_2}$) is satisfied and demand deposit contracts can implement the first-best allocation with a higher level of liquidity creation and investment than under autarky, $I^{*} \ge I^{aut}$.

Proposition 1 rationalizes the empirical finding in this paper that banks’ liquidity creation improves tangible but not intangible investment (see Tables 6 and 7). By transforming illiquid and risky assets into liquid claims, banks can reduce liquidity and credit risk and increase investment and consumption. Yet, Proposition 1 highlights that this positive effect of liquidity creation on investment relies on long-term investment that is not too intangible. If k is very small, then by Assumption 1$\mu$ and $\gamma$ are large. A high value of $\gamma$ means that investors can divert a larger fraction of investment returns. Similarly, if the liquidation loss $\mu$ is large, the return gap between successful and failing projects is large. This makes diversion attractive relative to a claim on the bank which earns the average return of failing and successful projects. Intuitively, intangible assets exacerbate moral hazard problems because they may be easier to divert and have lower liquidation value. This makes it harder for banks to make loans against intangible assets, explaining why liquidity creation supports tangible but not intangible investment.

Our model is also consistent with the results presented in Fig. 3, which show a stronger relation between liquidity creation and investment in countries with stronger investor protection and recovery rates, with the impact of legal system efficiency being stronger in the case of intangible than tangible assets. Specifically, if—in line with the literature—we expand our assumptions on $\gamma$ and $\mu$ to be a function not only of asset tangibility but also investor protection and recovery rate, respectively, our model explains why the effect of liquidity creation on investment is stronger in countries with higher investor protection and recovery rates. For example, comparing a “low institutional quality” configuration of the model with high $\gamma$ to a “high institutional quality” configuration with low $\gamma$, then (${IC_2}$) is more likely satisfied in the high institutional quality configuration, consistent with the results in Fig. 3 showing a greater effect of liquidity creation on investment in countries with stronger investor protection. Also in line with the model, the slope of the curve is steeper for intangible than for tangible investment, indicating that improvements in investor protection and recovery rates matter more for intangible than tangible investment.²⁹

If $k<{\bar{k}}$, investors fall back to the allocation under autarky. This case resembles an economy with mostly intangible investment opportunities, in which banks’ liquidity creation cannot support investment and investors who rely on direct equity financing are exposed to liquidity risk. Beyond the scope of the model, one might expect more specialized financial intermediaries such as venture capital funds to emerge in an intangible economy. Such intermediaries can provide funding and share risks while closely monitoring entrepreneurs to mitigate moral hazard problems associated with intangible investments.

While the model focuses on explaining our investment results, it also relates to the growth results because investment and consumption are parts of overall GDP. Therefore, the model also speaks to our result that banks’ liquidity creation supports growth more in economies that rely less on intangible capital (see Table 9).

6 Conclusion

In this paper, we present theoretical and empirical evidence that banks’ liquidity creation fosters economic growth, but with an important non-linearity; banks’ liquidity creation helps increase tangible but not intangible investment in an economy. Our findings thus stress the importance of liquidity creation by banks in economic development through overcoming market frictions in investment in tangible assets, but also the limitations of banks in supporting innovative industries that rely primarily on intangible assets. These findings also stress the importance of financial structures adjusting as economies transition from an economic structure where investment is mainly tangible to one where the intangible share is growing.

Given the increasing importance of intangible investment in the “knowledge economy” (Haskel & Westlake, 2017), these findings also shed light on the future of banking and finance. Consistent with other evidence that innovative industries rely mostly on non-bank forms of external finance, our findings suggest the importance of moving away from a bank bias in European finance to more diversified financial systems (Langfield & Pagano, 2016). Our findings of a reduced role of banks in funding intangible investment and thus fostering economic growth in high-income countries also weakens the “economic growth” justification for bank bailouts during episodes of systemic distress. While banks continue to have a central role in the payment system, our findings suggest that the expansion of the critical funding role for corporations can be viewed as an argument for expanding the regulatory perimeter beyond the banking system.”

Although our work helps understand how banking sector development facilitates economic growth, it excludes from its scope the potential role of liquidity creation by non-bank financial intermediaries. The mechanism of liquidity creation by banks we describe in this paper, including their preference for tangible assets, does not necessarily require banks. Other non-bank entities – sometimes called “shadow banks”—can also obtain an efficient allocation, potentially with other types of contracts than demand deposit contracts. Money market funds are one specific and potentially important example (Jacklin, 1987). As liquidity creation by non-banks continues to grow in most developed countries, other interesting questions that are outside the scope of this paper emerge: how much liquidity do non-bank financial intermediaries create relative to banks around the world? Does liquidity created by non-bank financial intermediaries substitute or complement liquidity created by the traditional banking sector? Does liquidity creation by non-bank financial intermediaries affect economic growth? And, if so, what is the mechanism?

Our findings also paint a positive role of liquidity creation in stimulating growth via investment. Yet, liquidity creation by banks can come at the cost of heightened fragility and failure risk (Chen et al., 2020). Indeed, banks may not always hold sufficient liquid assets to meet the immediate withdrawal demands by all depositors, which can be conducive to panic-based runs (Goldstein & Pauzner, 2005). Other questions thus naturally arise: does fragility induced by liquidity creation affect economic growth? Again, if so, how? Understanding and quantifying the interlinkages between liquidity creation (by banks and non-banks), financial fragility, and the real economy remains a fruitful area for future research.