Policy uncertainty and loan loss provisions in the banking industry

Stated differently, policy uncertainty is a significant predictor of future negative economic conditions, incremental to other existing predictors. The Federal Open Market Committee (2009) and the International Monetary Fund (2012, 2013) also suggest that uncertainty about US and European economic policies contributed to the steep depression in 2008–2009 and the slow recovery afterward.

The incurred loss model or SEC Staff Accounting Bulletin: No. 102 does not necessarily require banks to recognize provisions anticipating future downturns. Nevertheless, banks can exercise discretion over loan loss provisions to engage in information signaling or earnings management (Beatty and Liao 2014).

Many studies either argue or provide evidence that managers can have incentives (e.g., career concerns) to withhold or delay the disclosure of bad news (e.g., Graham et al. 2005; Kothari et al. 2009). As an illustration within the context of our study, while policy uncertainty indicates that a bank will have expected loan losses of $100 million, the bank might record only $80 million of loan loss provisions.

This index consists of four components. The first and most important of these captures policy uncertainty using a newspaper-based approach, based on the frequency of articles in 10 leading US newspapers that contain the following trio of terms: “economic” or “economy”; “uncertain” or “uncertainty”; and one or more of “Congress,” “deficit,” “Federal Reserve,” “legislation,” “regulation,” or “White House.” The other three components capture tax code expirations, disagreement over CPI forecasts, and disagreement over government purchases forecasts.

In the sample, about 5.60% of the observations have nonpositive earnings before loan loss provisions.

It is very challenging to conclude an earnings bath occurred based simply on this analysis. For example, an alternative interpretation of this result is that banks suffering from losses before loan loss provisions might have a weaker loan portfolio that is more harmed by policy uncertainty. Hence one takeaway from this analysis is that banks with profits (or higher profits) before loan loss provisions are less likely to be subject to an earnings bath.

Timeliness in loan loss recognition is an important topic, due to the recent financial crisis, literature showing the positive consequences of such timeliness (Beatty and Liao 2011; Bushman and Williams 2012; Akins et al. 2017), and the coming shift from the incurred loss model to the expected loss model due to Financial Accounting Standards Board’s (2016) ASC 326 Financial Instruments – Credit Losses. In general, under the incurred loss model, a trigger event (e.g., bankruptcy filing) for a loan is required before a bank can accrue losses for it.

This approach is consistent with more recent loan loss provisioning models that include indicators of expectations of future loan losses, such as the change in nonperforming loans at t + 1 (Beatty and Liao 2011; Bushman and Williams 2012).

For example, using Chinese commercial banks as a sample, Chi and Li (2017) document a positive relation between policy uncertainty and banks’ credit risk.

The news article can be found online: http://​www.​reuters.​com/​article/​us-platinum-week-anglo-platinum/​more-south-african-mining-jobs-at-stake-on-policy-uncertainty-amplats-ceo-idUSKCN18C1YA.

We might also find no relation between loan loss provisions and policy uncertainty, because accounting standards do not explicitly require banks to record higher loan loss provisions when there is higher uncertainty, whether it be economic or another type. In fact, standard setters have made a recent push toward more forward-looking provisioning for losses, as evidenced by the introduction of a new accounting standard, Financial Accounting Standards Board’s (2016) ASU No. 2016-13, which will be mandatory for fiscal years beginning after Dec. 15, 2019.

International studies on timely loan loss recognition also document that such recognition leads to greater bank prudence. For example, there is evidence that timely loan loss recognition enhances bank risk-taking discipline (Bushman and Williams 2012) and lending corruption (Akins et al. 2017).

Data for all US commercial banks are publicly available online at the Federal Reserve Bank of Chicago: https://​www.​chicagofed.​org/​banking/​financial-institution-reports/​commercial-bank-data; we alternatively use the bank holding companies sample obtained from the Compustat Bank database and obtain empirical results that are qualitatively the same (untabulated).

As quarterly data are available for all US banks, most banking research uses bank-quarter observations, e.g., Beatty and Liao (2011), Bushman and Williams (2012, 2015), and Goetz et al. (2016).

Alternatively, we follow Beck and Narayanamoorthy (2013) in starting our sample period with 1992 (when Basel took effect) and then constructing a capital ratio proxy as the percentage of total equity to total assets. We get qualitatively similar results.

The final sample period starts with the second quarter of 1996, because we have to control for the lagged risk-based capital ratio in our baseline model.

Because the measure of policy uncertainty is a macro-level variable with only time-series variation, the regression model cannot include quarter fixed effects.

In untabulated graphs, we also observe a very clear pattern of the aggregate level of loan loss provisions being associated with some other macro-level indicators; e.g., loan loss provisions increase with macroeconomic uncertainty and the unemployment rate, and they decrease with GDP growth, the Case Shiller home price index and the managerial sentiment index. Those univariate relationships are consistent with the economic explanation.

The reported percentage is calculated as (regression coefficient × standard deviation of PU_t) / mean value of LLP_i,t.

While the negative coefficient on EBP_i,t is inconsistent with earlier studies, such as by Beatty et al. (1995), Collins et al. (1995) and Liu and Ryan (2006), it is consistent with recent studies, such as by Beatty and Liao (2011) and Hribar et al. (2017).

The sum of the sample size in Table 2, columns (2)–(4), is slightly smaller than in column (1), because we include bank fixed effects and the singletons in each subperiod are automatically dropped.

The managerial sentiment measure is derived from the Duke University/CFO Magazine Business Outlook survey, which is only available after 2002. We do not control for managerial sentiment in our main specification, because doing so would significantly shorten our sample period. In untabulated results (available upon request), we find that the empirical tests of all our hypotheses are robust to the inclusion of managerial sentiment as an additional control variable.

In the latter part of this paper, we provide more evidence using future loan charge-offs and loan growth to support the notion that managers do indeed communicate their expectations.

The sample size of this analysis is slightly smaller because the political polarization data ends in 2014.

We may have more than one neighboring state for each election state, and each neighboring state with no election can also be matched with more than one election state. To deal with these duplicate observations, for each nonelection state with multiple matches, we only use the nonelection state as the benchmark for the election state with the closest GDP. Finally, we have to drop election states if they lack benchmarking states. To this end, we make sure that our final sample consists of unique bank-quarter observations and that neighboring states can serve as a benchmark for each gubernatorial election. The final sample size for the gubernatorial election analysis is 245,241 bank-quarter observations, which is much smaller than our primary sample size in the baseline analysis. There are two reasons for the smaller sample size. First, here we only focus on single-state banks. Second, here we only focus on the year preceding a gubernatorial election and the election cycle is two or four years.

The sum of the sample size in Table 4, columns (1) and (2), is slightly smaller than our full sample size, because we include bank fixed effects and the singletons in each subsample are automatically dropped.

In our sample, 406 observations have zero earnings before loan loss provisions (EBP_i,t = 0). Dropping those observations or including them into either of the subsamples will not change our inference. We also obtain consistent results via an interaction approach, i.e., to create a dummy variable of negative EBP_i,t to interact with PU_t and find a significantly positive coefficient on the interaction term (untabulated).

Another possible interpretation is that there is nonlinearity in the economic relation between policy uncertainty and loan loss provisions to the extent that banks in worse shape expect more loan losses than those in better shape.

We thank the referee for suggesting this approach.

This finding is consistent with that documented by Bordo et al. (2016).

However, this percentage only accounts for the past level of nonperforming loans. Akins et al. (2017) modify this measure to include the current changes in nonperforming loans; i.e., they compute the loan loss reserves at quarter t-1 as a percentage of the nonperforming loans at quarter t. We obtain qualitatively similar results using this modified measure.

The only exception is that, when we use the tier 1 capital ratio to measure bank prudence and run the regression without the interaction term, the main effect on bank prudence is positive but insignificant (Coeff. = 0.0014 and t-value = 0.99). This finding is more or less consistent with the shift in the relation between regulatory capital and loan loss provisions from negative in the pre-Basel period to positive post Basel (Beatty and Liao 2014).

One might suspect that our results are driven by larger banks, especially those with total assets over $500 million, as they are subject to a mandatory independent audit. To rule out this concern, we repeat the analysis on smaller banks. Because such banks’ total assets are less than $500 million, their financial statement audits are purely voluntary. We get qualitatively similar results (untabulated).

In addition to the newspaper-based index comprised of the 10 leading US newspapers, Baker et al. (2016) provide an alternative policy uncertainty measure derived from the Access World News database, which includes over 2000 US newspapers. We find similar results using this alternative measure.

In particular, it is unclear which type of uncertainty would have a larger effect on the kind of adverse future economic conditions that would affect borrowers’ ability to repay their loans.

The sum of the sample size of two subsample periods, in Table 8, panel B, is slightly smaller than our full sample size, because we include bank fixed effects and the singletons in each subsample are automatically dropped.