An Examination of Country Member Bank Cash Balances of the 1930s: A Test of Alternative Explanations.

Wm. Stewart Mounts [+]

Clifford B. Sowell [+]

Atul K. Saxena [++]

Around March 1933, commercial banks began accumulating unprecedented amounts of cash. Uncertainty over deposits and loans, low interest rates relative to brokerage fees, and costly de novo capital have been used to explain this behavior. This paper employs aggregate call-date data for country member banks in the 12 Federal Reserve districts to formally investigate the role of these factors in the accumulation. The results indicate a minimal, if any, role for these factors. The findings suggest that it was the unintended consequence of unprecedented deposit growth in the face of large scale-related adjustment costs.

1. Introduction

Around the Emergency Banking Act of March 9, 1933 (the "bank holiday"), banks began accumulating unprecedented amounts of cash assets. [1] Figure 1 shows this for country member banks in a sample of Federal Reserve districts. [2] Table 1 documents this run-up relative to other assets at the country bank level for selected call dates. As shown, cash remained a relatively stable percentage of total assets through the December 1932 call date, whereas loans and securities changed over the business cycle. [3]

Explanations of this phenomenon may be found in the seminal work of Friedman and Schwartz (1963), Morrison (1966), Frost (1971), Ramos (1996), and Calomiris and Wilson (1996), with the macroeconomic framework of Bernanke (1983) being tangential. Each argues in some way that the accumulation of cash was desired and driven primarily by uncertainty over future bank runs, low (relative to brokerage fees) interest rates, uncertainty over loan payback, or by high costs of adding de novo bank capital.

The purpose of this paper is to present an alternative explanation for the cash accumulation within an empirical examination of the existing literature. Our principal point of differentiation turns on whether the cash accumulation was desired. Except for Bernanke (1983), the literature treats the cash buildup as desirable or something toward which bankers were indifferent. Whether it was accumulated to protect against runs or massive loan failures or to signal bank solvency or was the result of low net (of transactions costs) yields on securities, no clear distinction is made between actual cash holdings and target cash holdings.

Why does this distinction matter? Consider, for example, a passage from Ramos (1996):

During the 1930s, rather than redundant, banks' excess reserves were desired and were performing an economic function. Holding such amounts of cash was a profit maximizing decision. (p. 473)

As will be shown, the cash buildup was not due to an asset reallocation with a fixed deposit base but from a significant deposit inflow after the holiday. While loans and securities would have been acquired because of their return relative to non-interest-bearing cash and the need for cash flow, the recent work of Ramos (and Calomiris and Wilson 1996 for that matter) argues that the need to signal solvency was a significant component of bank decision making. Assuming that acquiring de novo capital (an alternative signal of solvency) was costly, the cash accumulation produced benefits unavailable elsewhere and, as such, was desired. To Friedman and Schwartz (and Morrison) it was desired because of the uncertainty over the Fed's willingness to be the lender of last resort and to Frost it was the desired (or natural) result from a simple asset substitution driven by net security yields of zero. [4]

However, what if solvency signals or uncertainty were not significant components in decision making? What if net security yields were not zero? With the absence of these factors, bankers would have sought to minimize cash given that it produced no significant benefits other than those offered under periods of normal banking operations.

Alternatively, low target levels would have been set, and security and loan assets would have been acquired. To do this, costs would have been incurred, especially with respect to loans. Two sources of costs need to be identified. First, there are those costs associated with managerial, administrative, and general day-to-day banking activities--costs directly related to the creation of interest-earning assets. We refer to these as "scale-related costs" in that it seems reasonable to expect these costs, like any cost of production, to be subject to diminishing returns because of the fixed nature, or scale, of the banking enterprise. Next, as found in Bernanke's (1983) "cost of credit intermediation" explanation of the Depression, there are also informational costs directly related to the inherent uncertainty of loan payback and collateral values. If this uncertainty over loans was not a significant factor in decision making, then the creation of loans and other assets would have been constrained largely by sc ale-related costs.

Given that the closings of the holiday had the effect of increasing market share for the remaining licensed banks, the large and possibly unanticipated deposit inflow would have encouraged banks, at some time, to expand. However, banking laws at the time limited expansion through unit banking laws and the like, restricting the production function and thereby asset creation. Thus, the impact of scale-related costs may have been significant in the buildup of cash assets.

The importance of scale-related costs is an empirical question, as is the importance of uncertainty, solvency, and so on. Our examination is framed around the model of cash management under uncertainty of Baltensperger and Milde (1976). Unlike the existing literature, a specific measure of uncertainty is included in the estimation. As will be developed, the absence of an explicit treatment of uncertainty may have kept the estimated role of interest rates from being appropriately measured. In addition, since the appearance of Frost (1971), Cecchetti (1988) has published a series of Treasury security yields that have been adjusted for an 'exchange privilege," a common feature of the securities of the time. It will be argued that the use of these yields introduces brokerage fees into the estimation, allowing for an accurate test of Frost's low-interest-rate argument.

The presentation is arranged as follows. The first part of section 2 reviews the entrenched and contemporary literature, while the second part critiques it within a historical setting. Several of the arguments found in the literature seem inconsistent with historical events and trends exhibited by the aggregate country member bank call-date data. Section 3 develops the model of Baltensperger and Milde (1976) as an appropriate framework for testing the alternatives. Also, the nature of the data is examined and the estimation procedure described in this section. Results are discussed in section 4. In general, contrary to the literature, uncertainty over deposits and loans appears to have played only a minor role in the accumulation of cash. Further, the demand for cash does not exhibit a kink, even after employing Treasury yields that have been adjusted for brokerage fees. In addition, the results indicate that cash was not a timely substitute to costly equity. What is consistent across the districts is (i) a g enerally slow but varied speed of adjustment in reaching target levels of cash and (ii) the failure of country member banks to accurately forecast cash inflows. Both imply that the accumulation at the country member bank level was the result of unprecedented deposit growth beginning with the bank holiday in the face of significant scale-related adjustment costs. These results are consistent with a partial or limited view of the "cost of credit intermediation" of Bernanke (1983). Section 5 offers some conclusions.

2. Previous Studies and Historical Considerations

Previous Studies

In their seminal work, Friedman and Schwartz (1963, pp. 449-62) argued that previous periods of bank runs and financial panics of the 1920s and, particularly, the early 1930s shocked bankers into expecting large deposit outflows. From this perspective, the accumulation of cash after the bank holiday served as a form of "protective liquidity" against anticipated runs and against uncertainty in general. The desire for liquidity was amplified further by the narrowing gap between market interest rates and the Fed's discount rate. In some months during the 1930s, the gap was even negative; a penalty on discount window access communicated an unwillingness by the Fed to be the lender of last resort.

In a related study, Morrison (1966) used a model of transitory deposit potential and found support for the Friedman and Schwartz protective liquidity hypothesis. Part of his analysis compared the structure and experiences of the banking system in the United States with that of Canada. The United States, with its dependency on unit banking, was more susceptible to periodic bank runs relative to the national branch banking of Canada. [5]

Elsewhere, Frost (1971) argued that there was a kink (as opposed to a shift, as argued by Friedman and Schwartz and by Morrison) in the demand for excess reserves. In this "interest rate/brokerage fee hypothesis," the observed accumulation of cash assets after the bank holiday was driven primarily by nominal interest rates being below some critical level. Given that brokerage fees had to be paid to acquire securities, it made sense to simply accumulate cash once interest rates fell below some level relative to the fees. [6] In such instances, the net effective yield on security assets would be zero. [7]

Bernanke (1983) indirectly addressed the accumulation of cash through loan uncertainty in his "cost of credit intermediation" macroeconomic analysis of the 1930s. Bernanke argued that during the Depression the channels of information used to determine loan payback probabilities and to forecast the future value of collateral became distorted and uncertain, causing the cost of credit intermediation to rise significantly. As such, bankers found it increasingly difficult to forecast the loan payback probabilities of borrowers, bankruptcies, and the value of collateral. These difficulties reduced their willingness to make loans. As such, cash may have been accumulated during periods of deposit growth as banks shifted away from loans toward more liquid assets of which cash was one.

Recently, Ramos (1996) and Calomiris and Wilson (1996) argued that commercial banks emerged from the late 1920s and early 1930s undercapitalized. Given (i) a desired default risk on deposits, (ii) the presence of asymmetric information, and (iii) costly de novo equity, the accumulation of (relatively inexpensive) cash lowered asset risk and signaled to depositors and security markets that bank solvency was not at risk and that runs were not warranted. Here, cash is seen as a form of self-insurance.

Historical Considerations

A review of the historical record questions the foundations of this literature. Several points should be developed.

First, changes in the institutional environment and in the soundness of the banking system need to be considered. The founding of the Federal Deposit Insurance Corporation (FDIC), as well as other components of the Banking Act of 1933, had the effect of stopping runs. [8] Bank suspensions numbered 506 in 1921, 975 in 1926, 1350 in 1930, 2293 in 1931, 1453 in 1932, and 4000 in 1933. By 1934, the number of suspensions had fallen to 61, reaching a post-holiday high of only 82 in 1937. [9] In addition, the large inflow of gold (initiated by the January 31, 1934, revaluation), the resulting growth in high-powered money, along with the return of currency from the pre-holiday drain to a smaller number of (post-holiday) banks, produced an expansion in average country bank deposits that is clearly seen in the panels of Figure 2. This unprecedented growth suggests a growing public confidence in the banking system, reducing the need for protective liquidity. However, as seen in Figure 1, the accumulation of cash did no t slow.

In support of the protective liquidity view, Friedman and Schwartz (1963, p. 458) argued that excess cash assets were held in Federal Reserve district banks. If public uncertainty over bank soundness was operative in forming expectations or, as Ramos (1996) and Calomiris and Wilson (1996) argued, bankers wanted to signal solvency to depositors and security markets, excess reserve deposits at Federal Reserve district banks would have clearly signaled soundness and safety by avoiding reserve pyramiding so common during the years of the National Banking System (1863--1913). Figure 3, however, shows that (average) country member bank cash assets were held primarily as deposits at other commercial banks (ABAL-) and not at the respective Fed district banks as excess reserve balances (AEXR-). The literature on financial crises indicates that bank panics and runs are initially local in character, suggesting that under heightened uncertainty reserves would not be held in other proximate commercial banks. Rather than worrying about the availability of reserve balances, the growth of these balances suggests that bankers were more likely exhibiting cost-minimizing behavior. [10]

Frost's (1971) interest rate/brokerage fee hypothesis implies that banks should have shown a decrease in the rate at which they accumulated securities (or a decline in their willingness to hold them) after the bank holiday. Falling nominal interest rates in the face of fixed brokerage fees would have made security accumulation less attractive. The panels of Figure 4 show no such pattern, however. Government securities (GS), shown as dashed lines, begin to grow in the late 1920s and continue to do so through the bank holiday until dips in 1937 and 1938, when reserve requirements were increased. Also, Table 1 shows that by the June 1936 call date, securities had reached a percentage of total assets that exceeded anything prior to the bank holiday. Furthermore, the Treasury subsidized security acquisition. Cecchetti (1988) has argued that the liquidity of Treasury securities was altered by an exchange privilege; at maturity, securities could be exchanged at par for a new issue. [11] Thus, banks could earn above -market returns by adding and then holding government security assets. Accordingly, the presence of this privilege reduced the liquidity of Treasury issues but raised the effective yield when held to maturity. The patterns in the panels of Figure 4 seem to support the idea that it paid to buy and then hold securities to maturity.

In another area, the closing of unsafe banks in March 1933 began contributing to the soundness of the banking system. One of the functions of banks is to screen and monitor loan applicants. It could be argued that most of the remaining (licensed) banks had been relatively more efficient (or luckier) in these activities than those closed by the authorities. Apparently, the remaining banks had better or, at least, more "acceptable" balance sheets, suggesting that they may have been better monitors of loan customers. Additionally, the remaining banks had a larger pool of loan applicants as market share grew with the suspensions. As a result, the average loan portfolio would have improved in average loan Figure 5 (average country member bank loans in four Federal Reserve districts) shows that loan activity started to increase by 1935 and, in most cases, continued to do so for the remainder of the period. [12]

Ramos (1996) and and Wilson (1996) argue that banks emerged from the Depression years undercapitalized. Given the high costs of raising new equity capital, the accumulation of cash offered banks a relatively low-cost substitute for signaling bank solvency. The data depicted in Figure 6 seem to question this view. As shown in the upper panel, the country member banks in these districts faced the bank holiday with relatively high, not low, capital/asset ratios. It is only after the bank holiday that this ratio falls. The lower panels show that, in most cases, actual average capital holdings either remained stable or increased slightly after the holiday.

3. Bank Management, Data, Methodology, and Strategy

A Model of Cash Management

Given the importance assigned to uncertainty in the literature, we draw on the work of Baltensperger and Milde (1976), which addresses the management of primary (cash) and secondary (securities) bank reserves under uncertainty coming from loans and deposits. [13] A unique feature of this model is that bankers can acquire information about depositors and borrowers (current and future) and reduce the uncertainty associated with forecasting deposit behavior and changes in loan portfolios. By holding cash, the probability of reserve deficiencies (failure to meet currency calls) is reduced, as are the costs of adjusting asset portfolios that such deficiencies might require. Alternatively, the banker can reduce cash reserves by acquiring information about borrowers and depositors. This information reduces the uncertainty associated with forecasting net reserve demand within a given period.

A demand for cash reserves (R) is derived principally from the uncertain callability of the implied contract between deposit holders and banks. In addition, the presence of brokerage fees, uncertainty over loan payback and future loan demand, as well as possible capital losses on the sale of securities contribute to the need for an inventory of cash balances.

The costs (C) associated with cash inventories include direct opportunity costs, the costs of adjusting the asset portfolio, and the costs of acquiring information. Again, following Baltensperger and Milde (1976), a banker attempts to minimize

C= rR + (r - i)S + m [[[integral of].sup.R+S].sub.R] (u - R)f(u, q) du

+ n [[[integral of].sup.[infty]].sub.R+S] [mS + n(u - R - S)]f(u, q) du + sq. (1)

The first term in (1) is the opportunity cost of holding cash (primary reserves) as opposed to holding loans, where r is the market interest rate on loans. [14] The second term in (1) is the forgone income from holding securities as opposed to loans. It is assumed that r [greater than] i, where i is the market interest rate on securities. While securities cannot serve as legal reserves, they are low-cost alternatives (secondary reserves) to cash. The third and fourth parts of (1) represent the costs of converting securities and loans into cash given that R (or R and S) is (are) insufficient to meet the realized net demand (u) for R. The per unit costs of these transactions are given by m and n, respectively, and q is the quantity of resources used to collect information. The final term is the total expenditures on producing information on depositors and present and future borrowers.

As shown in the density function f, portfolio adjustment costs are dependent, in part, on q. Increases in q allow the banker to assign smaller probabilities to extreme forms of the net demand for cash (u) like a bank run and panic or a total default of the loan portfolio, both measured as the standard deviation ([sigma]) of the density function. Increasing q lowers [sigma] (the fear of a cash deficiency) and reduces the amount of cash needed in the reserve inventory.

Several points found in the comparative statics must be stressed. [15] First, q (information) responds differently to changes in i (the rate on securities) than it does to changes in r (the rate on loans). Increases in r shift the asset portfolio toward loans as the opportunity cost of cash and security reserves rise. This induces a greater use of q. In contrast, an increase in i, ceteris paribus, makes securities more attractive. This raises reserves (albeit secondary), serving to reduce the need to use q to forecast u. However, the decline in the use of information increases the standard deviation of the density function.

A result of this is that a change in the interest rate on securities (i) has an ambiguous effect on cash reserves (R). This comes from two factors: (i) asset portfolio adjustments due to changes in relative rates of return (the price effect) and (ii) asset adjustments due to changes in the degree of uncertainty faced by the banker (the information effect). An increase in i reduces cash holdings through a simple opportunity cost argument (the price effect). But, from the previous arguments, an increase in i also lowers q. This leads to an increase in uncertainty ([sigma] rises) and then to an increase in cash reserves (the information effect). With respect to the period after the bank holiday, low rates on securities would have increased the cash component of reserves through the price effect. However, this would have also raised the use of q. As a result, [sigma] would have fallen, and a decline in cash holdings would have occurred. Thus, the relationship between i and R is an empirical question that depends on the relative strengths of the price and information effects. Alternatively, to measure the price effects of interest rates on cash (Frost), uncertainty must be modeled in some way. [16]

The Data

The data are taken from aggregate balance sheets for country member banks collected on the June and December call dates in the 12 Federal Reserve districts. [17] Country member banks are selected for analysis in order (i) to keep the meaning of balance sheet data tractable, especially with respect to "balances held at other commercial banks"; (ii) to have a wide geographically dispersed group; (iii) to have a class of banks that may have experienced a great deal of intertemporal uncertainty; and (iv) to have a set of banks that may have found access to equity markets costly because of size, location, and/or the nature of ownership. [18]

A time-series analysis of the cash/asset ratio indicates that a data-consistent model must accommodate trend stationarity and a moderate to slight persistence in transient shocks.[19] Apparently, shocks alter forecasts but only slightly over long time periods as agents seem to hold almost static expectations. [20]

Empirical Methodology and Estimation Strategy

Estimation involves two steps. In the first, two models--an ARCH model of reserve flows and an autoregressive forecasting equation of cash--are estimated recursively. In the second, the cash equation is estimated.[21] The dependent variable in the cash equation is the log of the cash/asset ratio.

The ARCH procedure provides estimates of the period-by-period variance of the residuals (the step-ahead error) that represents our proxy (SIGMA) for internal management uncertainty over cash flows conditioned on a set of regressors that vary across districts and that are consistent with Friedman and Schwartz, Morrison, and Bernanke. [22] These include the lag of the change in the log of deposits, the lag of the change in the log of loans, and the lag of the change in the log of the index of industrial production. [23] A separate ARCH model was selected for each district on the basis of goodness of fit and the included regressor's theoretical relationship to the cash decision.[24]

The autoregressive forecasting equation is estimated using the Kalman filter algorithm with an appropriate number of lags (generally two or three) of the log of the cash/asset ratio and a linear trend. This choice is based on the recognition of the mixed nature of the time series. The step-ahead errors from this model, SHOCK, are treated as forecast errors by bank management. [25]

The second step is the estimation of the cash equation. In addition to SIGMA from the ARCH model and SHOCK from the Kalman filter autoregressive forecasting model, other independent variables include the Treasury Bill rate adjusted for the exchange privilege (TBILL), the BAA rate, a one-period lag of the log of the capital/asset ratio ([CAP.sub.lag1]), and lags of the log of the cash/asset ratio (CASH), the dependent variable. [26] These variables are consistent with the optimization problem outlined in Baltensperger and Milde (1976), the pertinent literature, and scale-related adjustment costs. Also, two trends are included, one for the pre-holiday period (Trend-1) and one for the post-holiday period (Trend-2). [27]

The Treasury Bill rate, taken from Hamilton (1992), was derived from the work of Cecchetti (1988). Frost's argument centers around the return on securities relative to brokerage fees. A problem is that an appropriate data set on the bid/ask spread of bank-eligible bonds during the period is not available. However, we believe that yields based on Cecchetti's framework address the issue of brokerage fees in the following manner. Generally, brokerage fees are seen in the bid/ask spread on a security. Prior to 1929, it is probably reasonable to assume a symmetry between the bid/ask spread on a buy and a sell. However, beginning in 1929 the Treasury attached an exchange privilege to new issues; at maturity, old issues could be exchanged for new at par. As such, security holders could earn above-market rates, and the Treasury could be assured that a new issue would be fully subscribed. Cecchetti argues that the exchange privilege served the role of an underwriting fee. This privilege introduced an asymmetry in the bid/ask spread between a buy and a sell by making the effective brokerage fee on a buy equal to zero. Thus, his adjusted rates capture this subsidy and reflect the true opportunity cost of not holding securities.

The estimated signs of SIGMA and [CAP.sub.lag1] are of particular interest. A positive sign on SIGMA indicates that inherent managerial uncertainty in decision making raised the proportion of assets held as cash at the country bank level. A negative sign, however, would indicate that banks gathered information about depositors and borrowers in response to uncertainty as argued by Baltensperger and Milde (1976). Insignificance would indicate that uncertainty did not matter (or was only a minor influence) in the cash decision.

The lag of the capital/asset ratio is included because of the recent arguments of Ramos (1996) and Calomiris and Wilson (1996). [28] This literature leads one to expect a negative sign. Increases in capital would permit banks to increase their noncash asset portfolio risk (reduce cash balances) while maintaining a constant default risk on deposits.

Lastly, lagged cash is included in the estimation as a proxy for scale-related costs of adjustment associated with moving toward a preferred level of cash. Recall that brokerage fees are being held constant in the adjusted interest rates (Frost). Internal decision-making costs associated with loan (Bernanke) and deposit (Friedman and Schwartz) uncertainty are being held constant in SIGMA, whereas SHOCK measures managerial forecasting mistakes. As such, the remaining internal adjustment costs (those captured in lagged cash) in the banker's optimization problem reflect scale-related factors faced in moving toward cash targets. The latter influences are associated with limits imposed on the production of banking assets by unit banking laws as well as other regulations and legislation that constrain the expansion of the banking enterprise and the nature of assets offered. In each of the alternative arguments found in the literature, the accumulation of cash is implicitly treated as desirable. These views imply t hat internal and deliberate scale-related costs of adjustment should be small or insignificant, particularly using semiannual data and given the one-to-one relationship between deposit growth and the growth in cash.

4. Results

Table 2 lists the results from the previous procedure for nine Federal Reserve districts. [29] The cash equation was estimated and a Chow test was performed to check for coefficient stability with the December 1932 call date (just before the bank holiday) used as the breakpoint. The generally consistent results across the diverse nine districts listed (New York, Minneapolis, Chicago, Atlanta, St. Louis, Boston, Dallas, San Francisco, and Kansas City) could not reject both the null of coefficient stability and the null of normality (using the Jarque-Bera normality test) at the 10% level. In addition, the equations are stationary. [30]

As shown, the coefficient on the exchange privilege adjusted rate on Treasuries, TBILL, is insignificant in all but one district. With the inclusion of SIGMA in the estimation, TBILL coefficients represent the pure price effect from government securities in the Baltensperger-Milde model. Apparently, country banks were simply acquiring and holding securities as implied by Cecchetti (1988) and by the behavior exhibited in the panels of Figure 4. This would be expected given that above-market returns could be earned by holding securities to maturity.

As expected, the estimated coefficient on the BAA rate is negative in all but one district. However, the estimated coefficients are small. This and the insignificance of TBILL question the earlier findings of Frost (1971) that the demand for cash exhibited a kink at some low interest rate. [31] Further, at least on the country bank level, these results indicate that loans were more of an interest-sensitive decision variable with respect to cash management than were securities. [32]

The estimated coefficients on SIGMA are significant and negative in three of the nine districts. The six remaining coefficients are not significant. While the significance of the negative coefficients implies that bankers dealt with uncertainty by accumulating information on depositors and borrowers, consistent with the Baltensperger and Milde, the general conclusion to be drawn is that uncertainty appears to have played a small role in the accumulation of cash, contrary to the protective liquidity hypothesis of Friedman and Schwartz and Morrison and a portion of the "cost of credit intermediation" story of Bernanke.

The lag of the log of the capital/asset ratio, [CAP.sub.lag1], was included to examine the recent contributions of Calomiris and Wilson (1996) and Ramos (1996). The argument is that a high capital/asset ratio would allow banks to maintain a riskier asset portfolio if cash signaled safety to security markets and others. If this is the case, the percentage of total assets held as cash would be lower with prior higher capital/asset ratios. With a semiannual lag giving sufficient time for the requisite portfolio adjustments, an estimated negative sign or insignificance would be consistent with this view. [33] The reported results, however, are opposite of this expectation. [34] As indicated, prior high capital/asset ratios increased the proportion of assets held as cash.

These findings point toward a modified view to that of Ramos (1996) and Calomiris and Wilson (1996) that indicates a role for scale-related adjustment costs. In results not shown in Table 2, the estimated coefficient on the second lag of CAP is negative though not significant. Assume for the moment that a high capital/asset ratio does signal solvency and, as a result, that depositors are attracted to a bank. The initial effect of CAP is to increase cash. Next, bankers must convert these deposits into interest-earning assets. Given the use of semiannual data, the positive coefficient on the first lag and the negative coefficient on the second lag (although insignificant) indicate that country member banks found it difficult to make the necessary balance sheet adjustments suggested by the capital/asset view of the accumulation. Thus, scale-related costs of adjustment in the capital-to-cash arguments of Ramos and Calomiris and Wilson are probably small while the results point to them being large.

The estimated coefficients on lagged cash indicate the importance of these costs of adjustment in reaching a target level for cash, as do the step-ahead forecast errors in SHOCK. The coefficients on lagged cash and SHOCK are positive and significant in eight of the nine districts. The size of the lagged cash coefficients means that country member banks were facing large internal scale-related adjustment costs intrinsic to the banker's optimization problem, while the positive effect of forecast errors indicate that the cash inflow was largely underestimated. Except for the New York district, the importance of adjustment costs are also supported by the size of the mean lag (number of periods) for adjusting to target cash and by the size of the long-run multiplier. Even in the New York district, the positive sign on and the [CAP.sub.lag1] and ninemonth mean lag indicate the presence of meaningful adjustment costs.

Finally, it should be noted that variables for uncertainty over monetary policy and for policy in general were never significant at any stage of the estimation (including the ARCH estimations). This included variables for the discount rate, the gap between the discount rate and an open market rate, and a dummy variable indicating when the discount window was closed (a negative gap).

5. Conclusions

The 1930s mark a time of great institutional and structural change in commercial banking and financial markets. The Banking Act of 1933 (i) formalized deposit insurance with the founding of the EDIC, (ii) ended the practice of paying interest on demand deposits, and (iii) instituted Regulation Q; the Banking Act of 1935 restructured the Federal Reserve System into a central bank; and the Emergency Banking Act of March 9, 1933 (the bank holiday) suspended a sizable portion of "unsafe" financial institutions from the banking system.

Concurrent with these changes was an accumulation of non-interest-earning cash assets. While the literature addressing this accumulation stresses the role of uncertainty, low effective interest rates, and solvency signaling, the results presented here do not strongly support these alternatives, at least not on the country bank level.

An explanation consistent with the results points toward high scale-related adjustment costs in achieving desired cash targets. Deposit inputs to a bank, unlike inputs employed by other types of firms, are largely exogenous. This is especially true during the sample period and particularly after the Banking Act of 1933 banned the payment of interest on deposits and instituted interest rate ceilings (Regulation Q). Furthermore, the institutional events of 1933 and 1935 had the effect of reversing the deposit outflows of the 1920s and early 1930s, while the bank closings brought by the holiday had the effect of increasing the market share of the remaining banks. [35] However, the wide spread use of unit banking had the effect of limiting the expansion (or constraining the production function) of those that were licensed for operation after the bank holiday. One must also not forget the period's lack of banking technology, networking, and the absence (and lack) of easy access to extensive secondary markets. [36 ] In this setting, the accumulated cash at the country member bank level may have been like an unintended inventory for firms with expansion constraints (scale-related constraints) in an environment of exogenous inputs. Constraints on the size of the banking enterprise limited their ability to react to the large unanticipated inflow of deposit inputs. This interpretation is consistent with a limited version (with uncertainty held constant) of Bernanke's cost of credit intermediation. Scale-related costs limited the expansion of a major provider of credit to a segment of the economy unable to access money markets for their financial requirements.

Appendix

Call-date data were taken from the Board of Governors of the Federal Reserve System, Banking and Monetary Statistics (Washington, D.C., 1943). See "Principal Assets and Liabilities on Call Dates: 1920-1941" for the country member banks of the 12 districts: Atlanta, Table 242; Boston, Table 182; Chicago, Table 254; Cleveland, Table 218; Dallas, Table 302; Kansas City, Table 290; Minneapolis, Table 278; New York, Table 194; Philadelphia, Table 206; Richmond, Table 230; San Francisco, Table 314; and St. Louis, Table 266. Some excess reserve data were taken from Table 318, "Member Bank Reserve Balances, Excess Reserves, and Borrowings, by Class of Bank and by Federal Reserve District, Monthly, 1929-1941." Other data was taken from Hamilton (1992). The Fed classified member banks as central reserve city, reserve city, or country. Call reports were filed on call dates with the Comptroller of the Currency (by national banks) and with Federal Reserve District Banks (by state member banks). Reported call date data are for dates when both state-chartered member banks and national banks filed reports. June and December call data are used in this study to avoid the problems associated with irregular periodicity. Call data also contain the number of reporting banks.

(*.) Stetson School of Business and Economics, Mercer University, Macon, GA 31207, USA; E-mail mounts_ws@ mercer.edu.

(+.) Department of Economics, Berea College, Berea, KY 40404, USA.

(++.) Stetson School of Business and Economics, Mercer University, Macon, GA 31207, USA.

The authors wish to thank the participants in the Economic History II Session of the Southern Economic Association Meetings held in Washington, D.C., November 1996, especially Joseph R. Mason; the participants in the Research Seminar in Economics and Finance held at Georgia Southern University, October 1996; readers at the 1997 Financial Management Association; the participants of the Second Friday Research Colloquium of Mercer University held in December 1997; and Kent Kimbrough, Mark Toma, and two anonymous readers. However, any errors remain our own.

Received February 1999; accepted August 1999.

(1.) Cash is defined as the sum of vault cash, excess reserve balances held at the Federal Reserve district banks, balances held at other commercial banks, and float. Data sources are described in the Appendix.

(2.) The vertical line is for the June 1933 call date; the first one after the bank holiday. It should be noted that while the number of country banks varied widely prior to the bank holiday, the number within each Federal Reserve district changed very little afterward. Thus, the run-up of average cash seen in the figure is not simply from averaging a series. The figures presented throughout the paper represent only a selection of Federal Reserve districts. Other figures for other districts not pictured are available from the authors.

(3.) This is the last reported call date before the bank holiday. June 1933 is the next recorded date of call. Securities are defined as government securities.

(4.) Similar reasoning could be applied to Bernanke (1983). However, be does not directly address the cash buildup.

(5.) Ramos (1996) also discusses the differences in the two banking systems.

(6.) One could also argue that bankers felt that actual interest rates were below "normal" interest rates and, accordingly, expected interest rates to rise. Cash would be accumulated to avoid possible capital losses on securities in such a scenario. While this is Keynes's (1936) liquidity trap and theoretically different from Frost, empirically it is indistinguishable from the interest rate/brokerage fee hypothesis. As such, an empirical test of Frost implicitly tests Keynes.

(7.) Frost's theoretical analysis includes brokerage fees with fixed and variable components.

(8.) See Friedman and Schwartz (1963) and Wheelock (1993) and Berger, Herring, and Szego (1995). Whether or not the implied causality is true is a minor issue. As of these events, runs and panics stopped.

(9.) Bank suspension data are taken from Table 16 (Commercial Bank Suspensions, 1921-1960) in Friedman and Schwartz (1963), p. 438-9. Also, see Wheelock (1993, 1995) for an expanded analysis of the determinants of bank failures across states during this period. Calomiris and Mason (1997) examine the 1932 banking panic in Chicago.

(10.) Prior to the Banking Act of 1933, balances held at other commercial banks earned interest. This legislation banned this practice. After the act, banks received subsidizes services and lines of credit for maintaining balances in other commercial banks (see Gendreau 1983). At the country bank level, balances held in other commercial banks could be deducted directly from the base against which legal reserves were calculated (see Friedman and Schwartz 1963, p. 447).

(11.) This explains why recorded nominal interest rates were often negative during the period. See Cecchetti (1988) for a theoretical perspective.

(12.) In an analysis not reported here, the compound annual rate of growth in lending and the variance of average loans over two period--1922:2 to 1929:1 and 1936:1 until the end of the period--was calculated in each of the Federal Reserve districts. Except for the northeastern districts, the annual rate of growth in lending is larger in the second period. It is also the case that in 8 of the 12 districts, the variance in average loans is smaller in the second period. At least on the surface, this seems to question the importance of the uncertainty over loans as a reason to accumulate cash. This analysis is available from the authors.

(13.) Papers by Bernanke (1983), Bernanke and Gertler (1988), and Wheelock (1990) also contributed to the development of this section.

(14.) As in Baltensperger and Milde (1976), we will not directly incorporate legal reserves into our analysis.

(15.) See Baltensperger and Milde (1976) for the derivation of these conditions.

(16.) Frost does not explicitly model uncertainty in his analysis. Frost (1971, p. 819) notes that it is difficult to empirically differentiate his explanation from that of Morrison. As such, testing for the importance of uncertainty in general in the accumulation of cash may indirectly test Frost's hypothesis.

(17.) Data sources are described in the Appendix.

(18.) The sample in Ramos (1996) is all Federal Reserve member banks, while Calomiris and Wilson (1996) examined only New York City banks.

(19.) This analysis included Dickey--Fuller tests, KPSS tests (see Kwiatkowski et al. 1992), and the variance ratio test (See Kim, Nelson, and Startz 1991). A similar analysis was performed on other commonly used measures of excess reserves (aggregate monthly excess reserves, end-of-month, and weekly data). These tests are available from the authors.

(20.) It should be noted that conventional time-series models like Box-Jenkins or vector autoregressions overstate the longrun impact of shocks and as such will not yield data-consistent results. As an experiment, we estimated a set of loworder ARIMA models and used the Schwartz and Akaike information criteria to select alternative models for the dependent variable. The results reflect the difficulty that such methods have in simultaneously matching the short- and long-run dynamics in the data. These results are available from the authors.

(21.) Our choice of a single equation rather than a multiequation model of asset management is guided, in part, by sample size and data availability, However, as will be developed in the following, the single equation setup captures the role of other assets in the banker's cash decision.

(22.) A moving variance could also serve as a measure of uncertainty. However, sample size limits the use of this altemative.

(23.) The dependent variable is the change (log) in cash assets. Because of insufficient data, no district specific variables were used in any of the estimations.

(24.) Results for only nine districts are reported in the Results section. Reasons for this are presented there. The independent variables (all in logs) employed in each district's ARCH estimates are as follows: New York, Kansas City, Chicago, San Francisco, Minneapolis, and Dallas--one lag of the change in loans, the change in deposits, and the change in industrial production; Boston--one lag of the change in deposits, the change in industrial production, and contemporaneous change in loans; St. Louis--contemporaneous loans and one lag of the change in cash, the change in deposits, and the change in industrial production; Atlanta--one lag of the change in loans, the change in cash, the change in deposits, and the change in industrial production.

(25.) These must be white noise. Pagan (1984) discusses the use of regression-generated data.

(26.) F-tests of exclusion were used to determine the lag length of the cash/asset ratio.

(27.) The use of a split trend is discussed in Perron (1989).

(28.) In results not reported here, simple Granger tests of causality indicated that the capital/asset ratio causes the cash/asset ratio in 6 of the 12 districts, partially Consistent with this view. As such, this variable should be included in the cash equation. Results are available from the authors.

(29.) The results do not change appreciably with the log of average cash holdings as the dependent variable.

(30.) Results for the omitted Richmond, Cleveland, and Philadelphia districts were not stationary in that the sum of the coefficients on lagged cash exceeded 1. Given our interest in stationary and linear time-series processes, the techniques employed in this paper could not be applied to these districts. Please contact the authors to see the results of applying the data of these districts to the methodology described in the paper.

(31.) Dropping the BAA rate from the estimation does not make TBILL significant, while dropping TBILL does not change the coefficient or significance of BAA.

(32.) One must remember that the depth of the government securities market as well as the ease of access to it were products of the coming war years and the associated Treasury financing needs. Until then, the government securities market may have seemed too distant for regular day-to-day asset management decisions of country bank managers. This issue is difficult to quantify and is simply noted here.

(33.) The arguments of Ramos (1996) and Calomiris and Wilson (1996) suggest that CAP should be lagged. Given the balance sheet nature of the data, it is not clear whether contemporaneous data would indicate anything other than mechanical accounting relationships. Lagging the capital asset ratio may more accurately reflect the behavioral relationships presented in this literature.

(34.) Contemporaneous CAP was not significant.

(35.) Ramos (1996) and others stress that Canadian banks did not exhibit similar behavior with respect to cash. It is not noted, however, that Canadian banks did not experience the initial deposit outflows of the late 1920s and early 1930s, and therefore they could not experience the deposit inflows seen by U.S. banks in the post-holiday period. As such, Canadian banks would have had to engage in asset reallocation in order to accumulate cash. Banks in the United States, however, had to engage in earning-asset creation to not accumulate cash. The problems and situations seem to us to be very different.

(36.) This may be the reason for the creation of the "exchange privilege."

References

Baltensperger, Ernst, and Hellmuth Milde. 1976. Predictability of reserve demand, information costs, and portfolio behavior of commercial banks. Journal of Finance 31:835-43.

Berger, Allen N., Richard I. Herring, and Giorgio P. Szego. 1995. The role of capital in financial institutions. Journal of Banking and Finance 19:393-430.

Bernanke, Ben S. 1983. Nonmonetary effects of the financial crisis in the propagation of the great depression. American Economic Review 73:257-76.

Bemanke, Ben S., and Mark Gertler. 1988. Banking and macroeconomic equilibrium. In New approaches to monetary economics, edited by William A. Barnett and Kenneth J. Singleton. Cambridge, UK: Cambridge University Press, pp. 89-111.

Board of Governors of the Federal Reserve System. 1943. Banking and monetary statistics. Washington, DC: Board of Governors of the Federal Reserve System.

Calomiris, Charles W, and Joseph R. Mason. 1997. Contagion and bank failures during the great depression: The June 1932 Chicago banking panic. American Economic Review 87:863-83,

Calomiris, Charles W, and Berry Wilson. 1996. Bank capital and portfolio management: The 1930s capital crunch and scramble to shed risk. In Rethinking bank regulation: What should regulators do. 32nd Annual Conference on Bank Structure and Competition, Federal Reserve Bank of Chicago, pp. 515-30.

Ceechetti, Stephen G. 1988. The case of the negative nominal interest rates: New estimates of the term structure of interest rates during the great depression. Journal of Political Economy 96:1111-41.

Cecchetti, Stephen G. 1992. During the great depression: Was the deflation of 1930-1932 really anticipated? American Economic Review 82:141-57.

Friedman, Milton, and Anna J. Schwartz. 1963. A monetary history of the United States, 1867-1960. Princeton, NJ: Princeton University Press.

Frost, Peter A. 1971. Banks' demand for excess reserves. Journal of Political Economy 79:805-25.

Gendreau, Brian C. 1983. The implicit return on bankers' balances. Journal of Money, Credit, and Banking 15:411-24.

Hamilton, James D. 1992. Was the deflation during the great depression anticipated? Evidence from the commodity futures market. American Economic Review 82:157-78.

Keynes, John M. 1936. The general theory of employment, interest, and money. New York: Harcourt, Brace and Company.

Kim, M. J., C. R. Nelson, and R. Srartz. 1991. Mean reversion in stock prices: A reappraisal of the empirical evidence. Review of Economic Studies 58:515-28.

Kwiarkowski, D., P. C. B. Phillips, P. Schmidt, and Y. Shin. 1992. Testing the null hypothesis of stationarity against the alternative of a unit root: How sure are we that economic time-series have a unit root. Journal of Econometrics 54:159-78.

Morrison, George. 1966. Liquidity preference of commercial banks. Chicago: University of Chicago Press.

Pagan, Adrian. 1984. Econometric issues in the analysis of regressions with generated regressors. International Economic Review 25:221-47.

Perron, Pierre. 1989. The great crash, the oil price shock, and the unit root hypothesis. Econometrica 57:1361-401.

Ramos, Alberto M. 1996. Bank capital structures and the demand for liquid assets. Rethinking bank regulation: What should regulators do. 32nd Annual Conference on Bank Structure and Competition, Federal Reserve Bank of Chicago, pp. 473-501.

Wheelock, David C. 1990. Member bank borrowing and the Fed's contractionary monetary policy during the great depression. Journal of Money, Credit, and Banking 22:409-26.

Wheelock, David C. 1993. Government policy and banking market structure. Journal of Economic History 53:857-79.

Wheelock, David C. 1995. Regulation, market structure and the bank failure of the great depression. Review of the Federal Reserve Bank of St. Louis 77:27-38.

     Various Assets as Percentages of Total Assets: Country Banks [a]
Month/Year        Loans     Securities  Cash
June 1920        64.33%       12.53%   11.65%
December 1929    58.99%        9.45%   12.14%
December 1932    51.62%       13.90%   12.51%
              March 9, 1933
June 1936        32.06%       22.18%   26.17%
June 1941        35.23%       20.45%   31.22%
(a.)Each number is the percentage of total assets in each
category, averaged across the 12 Federal Reserve districts.