Uncertainty, Financial Markets, and Monetary Policy over the Last Century

3.1.1 Structural VAR Model

Our baseline VAR model includes five variables: the uncertainty index, the measure of financial conditions (the Baa–Aaa spread), the consumer price index (CPI), the industrial production index, and the federal funds rate. Our model is parsimonious but contains adequate information on output, prices, financial conditions, the monetary policy stance, and the measure of uncertainty over a long period, including 17 NBER recessions. The data spans from January 1919 to December 2017, the longest sample for which all five variables of interest are available. We impose structural assumptions on the variables, which are equivalent to the Cholesky identification arranging the variables in the order of the uncertainty index, credit spread, industrial production, CPI, and the federal funds rate.

This assumption implies that a variable is affected by the contemporaneous changes in the variables listed before it, while exogenous to the variables listed after it. For example, at time t, the unc _t , which is ordered first, does not respond to the concurrent innovations in the other variables. Still, it has contemporaneous impacts on the other four variables. Thus, the uncertainty variable is the most exogenous variable under this identifying assumption, which is in line with the practice of many VAR studies on uncertainty shocks (Basu and Bundick 2017; Bloom 2009; Caggiano, Castelnuovo, and Groshenny 2014; Fernández-Villaverde et al. 2015; Leduc and Liu, 2016).^[7] Although short-run restrictions impose stringent assumptions, especially when VARs include financial variables, the relatively high data frequency partly mitigates this concern. Moreover, placing the uncertainty index the last in Cholesky ordering, thereby treating the least exogenous, does not affect our conclusion.^[8]

To simplify the notation, let Y t = [ u n c t s p r e a d t i p t c p i t ff r t ] ′ be a 5 × 1 column vector of the five time-series of interest observed at time t. Each element indicates the uncertainty index, credit spread, the log of industrial production, the log of CPI, and the federal funds rate, respectively. We model the data in (log) levels to preserve the cointegrating relationships among the variables.^[9] Then the baseline VAR model can be represented in a matrix form as follows.

where, A ₀ is the lower triangular matrix that reflects the structural assumptions we imposed on the variables. p is the lag length, and six lags have been chosen in our baseline model. ɛ _t is a 5 × 1 column vector of structural shocks at time t, and each element in it is uncorrelated with each other.

3.1.2 Counterfactual Structural VAR Model

Our main focus is to examine how significant the role of financial conditions is in propagating the effect of uncertainty shocks on real activity.^[10] For this, it is conceptually necessary to shut down the dynamic response of financial conditions to the uncertainty shocks in the VAR and compare the results to those of the baseline model without any restrictions. In isolating and restricting the role of the financial conditions, we take a similar approach as that of Sims and Zha (2006), Kilian and Lewis (2011), and Bachmann and Sims (2012).

The resulting impulse response functions show the response of each variable to the uncertainty shock with the propagation mechanism by the financial conditions shut down. By comparing them to the impulse response functions of the baseline model without such constraints, we can identify the role of the financial conditions in transmitting uncertainty shocks to the other variables. To save space, we describe the procedure for constructing counterfactual impulse response functions in the online appendix B.

3.1.3 Baseline Results from a Hundred Years of Data

Figure 2A and B illustrate the responses of the four endogenous variables to a one standard deviation increase in economic policy uncertainty under our baseline VAR specification with sample period starting from January 1985 and January 1919, respectively. Although our baseline specification employs the historical sample from January 1919 to December 2017, as shown in Figure 2B throughout the paper, we also include Figure 2A to check whether our identification specification delivers findings consistent with those from previous studies using the EPU index dating from January 1985, such as Baker, Bloom, and Davis (2016).

Figure 2A:

Effect of uncertainty shocks: Baseline VAR model (1985M1–2017M12).

Note: This graph plots the effects of the one-standard-deviation uncertainty shock and their 68 and 90% confidence bands for the sample period between 1985M1 and 2017M12. The units of the horizontal axes are a month, and the units of the vertical axes are percentage points except for the federal funds rate (basis points).

Figure 2B:

Effect of uncertainty shocks: Baseline VAR model (1919M1–2017M12).

Note: This graph plots the effects of the one-standard-deviation uncertainty shock and their 68 and 90% confidence bands for the sample period between 1919M1 and 2017M12. The units of the horizontal axes are a month, and the units of the vertical axes are percentage points except for the federal funds rate (basis points).

Before discussing the role of financial conditions in propagating the effect of uncertainty shocks, a few novel observations from using historical data are worth to mention. While the responses in Figure 2A are fully consistent with Baker, Bloom, and Davis (2016), the response of CPI in Figure 2B changes its sign from Figure 2A. This finding is in sharp contrast to the recent studies by Leduc and Liu (2016) and Basu and Bundick (2017), who found a disinflationary effect of uncertainty shocks using a shorter span of the U.S. macroeconomic data and concluded that uncertainty shocks are a negative aggregate demand shock. However, the sign-varying response of CPI over time is consistent with the recent VAR evidence by Meinen and Roehe (2018) and Alessandri and Mumtaz (2019), who showed that the effect of uncertainty shocks on prices is indeterminate.

Theoretically, uncertainty shocks can also be inflationary through the so-called “inverse Oi-Hartman-Abel effect” (Born and Pfeifer 2014; Fernández-Villaverde et al. 2015). Since firms must satisfy demand given their prices and sticky prices create a convex marginal profit function, firms raise markups in response to heightened uncertainty to insure themselves against being stuck at too low a price – consequently, the aggregate output decreases and inflation increases. Using the U.S. data starting from the first quarter of 1950, Mumtaz and Theodoridis (2018) also found that uncertainty shocks are inflationary and suggested this pricing bias channel as a rationale. Fasani and Rossi (2018) considered the effect of uncertainty shocks under different Taylor rules and found that with a less active Taylor rule, uncertainty shocks look more like aggregate supply shocks rather than demand shocks.

Since the U.S. monetary policy was less active during the pre-1985 period than the post-1985 period analyzed in Figure 2A, the inflationary effect of uncertainty shocks is indeed consistent with Fasani and Rossi (2018). However, we pay further attention to the nature of uncertainty shocks as a complementary explanation for varying responses of inflation over time. Depending on its source, heightened uncertainty could have different effects on commodity prices, which pass through to CPI inflation. For example, if uncertainty shocks are originated from financial market disruptions, they would naturally reduce demand for commodities, including oil. On the other hand, if the increase in uncertainty is an outcome of the war or geopolitical tension, its effect on commodity prices, in general, would be the opposite. Depending on the relative occurrence between these two types of uncertainty shocks characterizing the sample period, inflation response to uncertainty shocks could be time-varying.

We test this hypothesis by including oil prices as an additional variable placed after the uncertainty variable.^[11] This ordering assumption appears natural given the (weak) exogeneity of global oil prices to U.S. economic activity. Figure A.1 presents the response of oil prices and CPI to the uncertainty shock for various subsamples. The results demonstrate the relevance of oil prices in characterizing the inflation response to the uncertainty shock. Although uncertainty shocks are always followed by a decline in output in any case, their effects on oil prices differ across subsamples, and such different effects translate into the different response of inflation. The decline in oil prices and CPI following the uncertainty shock is only observed in the recent sample since 1985, during which the sharp increases in uncertainty were often driven by financial market events, including Black Monday, Asian Financial Crisis, Dot-com bubble, global financial crisis, European sovereign debt crisis, and so on. On the other hand, in the earlier sample period, many heightened uncertainty incidences except for the Great Depression were an outcome of wars or political conflicts.

Our explanation also echoes with Carriero, Clark, and Marcellino (2018a), who found that financial uncertainty shocks have disinflationary effects, while macro uncertainty shocks have inflationary effects. Carriero, Clark, and Marcellino (2018a) estimated both financial and uncertainty using factor analysis of the U.S. data from 1959 to 2014. The fact that events with a large increase in their financial uncertainty measure are concentrated in the post-1985 period is consistent with our claim that a source of uncertainty matters in understanding the inflation effects of uncertainty shocks. Overall, our findings call for caution on the negative demand shock interpretation of uncertainty shocks by Leduc and Liu (2016) and Basu and Bundick (2017) using the recent sample only and suggest a careful investigation of the source of uncertainty to understand the effect on inflation.

The third panel depicts the response of output to the uncertainty shock. Figure 2B indicates a significant decline in output on impact, and the maximum decline occurs after six months, indicating that heightened economic policy uncertainty has a significant detrimental effect on real activity. A similar pattern is observed in Figure 2A, but the magnitude of the decline is smaller, which is consistent with Choi (2013), Choi (2017), and Mumtaz and Theodoridis (2018), who found that the real effect of uncertainty shocks has decreased over time. The last panel shows the response of the monetary policy stance to the uncertainty shock. The central bank takes an expansionary monetary policy stance by lowering the policy rate to counter the negative effect of uncertainty shock on real activity. However, the inflationary response to uncertainty shocks in the full sample implies a potential policy dilemma faced by the central bank due to the negative trade-off between output and inflation.

3.1.4 Counterfactual Exercises to Infer the Role of Financial Conditions

Figure 3 compares the impulse response functions of the baseline VAR model to those of the counterfactual exercise where propagation through the financial channel is shut down. The shaded area represents the 68% confidence bands for the impulse response functions, which is standard in the counterfactual VARs (e.g., Bachmann and Sims 2012; Kilian and Lewis 2011; Popp and Zhang 2016). However, we also plot the 90% confidence bands in dashed lines, which is more common in standard VARs. The decline in output is significantly smaller when the financial channel is shut down and no longer statistically significant.

Figure 3:

Effect of uncertainty shocks: Baseline vs. counterfactual.

Note: This graph plots the effects of the one-standard-deviation uncertainty shock and their 68 and 90% confidence bands for the sample period between 1919M1 and 2017M12 when the financial channel is shut down in the counterfactual exercise. The units of the horizontal axes are a month, and the units of the vertical axes are percentage points except for the federal funds rate (basis points).

While the response of CPI is more inflationary, the degree of monetary accommodation weakens slightly, probably because the monetary authorities do not need to lower the interest rate as much as in the baseline case. These findings corroborate the recent VAR studies employing a similar strategy but using much shorter periods (Bonciani and Van Roye 2016; Carrière-Swallow and Céspedes 2013; Popp and Zhang 2016).^[12] Thus, we conclude that the role of financial markets in transmitting and amplifying uncertainty shocks is robust and not restricted to the recent period.

3.1.5 Robustness Checks

We conduct several robustness tests to confirm the importance of the financial channel as a propagation mechanism of uncertainty shocks. To conserve space, we only discuss the results in the main text and present the relevant figures in the online appendix A. First, we run a parsimonious VAR model with only three variables (the uncertainty index, credit spread, and industrial production). Figure A.2 shows the responses of credit spread and output to the uncertainty shock using the baseline and counterfactual specifications in the trivariate VARs. The change in the response of output after shutting down the financial channel is not qualitatively different from the results from the 5-variable VARs.

We then repeat the same counterfactual exercise but limit our sample period to January 1919–December 2012, during which only the historical EPU index is used. This is to avoid any potential structural break in combining the historical EPU index with the current EPU index that incorporates the information beyond the newspaper coverage used in constructing the historical EPU index. Figure A.3 confirms that slightly different methodologies in constructing the historical and current EPU index do not affect our results. Additionally, given the unprecedented increase in both economic policy uncertainty and credit spread during the Great Depression and the Great Recession periods, we dummy out these periods using the NBER recession dates. Figure A.4 shows that our findings are robust even after these outlier events are dropped from the analysis.

We also replace the EPU index with a stock market volatility index, which is available from January 1928, to confirm whether the financial channel measured by credit spread still plays an important role in amplifying the so-called financial uncertainty shock.^[13] Figure A.5 compares the stock market volatility index constructed by Choi (2017) and the EPU index. Figure A.6 confirms that the adverse effect of uncertainty shocks is also substantially mitigated in this case, although the negative effect is still statistically significant in this case. Similar to the case of economic policy uncertainty, shutting down the financial channel makes the response more inflationary, although the baseline response is statistically insignificant.^[14] While the use of the EPU index demonstrates the importance of the financial channel more clearly, the qualitative results still hold when using the historical extension of the VIX.

We further test the robustness of our findings by replacing the Baa–Aaa spread with the financial stress index (FSI) constructed by Püttmann (2018). This index is constructed similarly to the EPU index by Baker, Bloom, and Davis (2016) but focused on the article pertaining to financial markets. The correlation between the Baa–Aaa spread and the FSI is 0.37, suggesting that they might capture different dimensions regarding financial market conditions. Figure A.7 compares both indices. Figure A.8 confirms that our main findings from Figure 3 still hold when using an alternative measure of financial market conditions.

Lastly, our benchmark measure of financial conditions mostly captures a firm’s borrowing costs (i.e., demand-side conditions) and does not necessarily capture the supply-side of credit. However, the supply of credit to the private sector via financial intermediaries has become an important source of business cycle fluctuations and a robust indicator of crises (e.g., Baron and Xiong 2017; Jordà, Schularick, and Alan 2013, 2017; Jordà et al. 2017; López-Salido, Stein, and Zakrajšek 2017). To understand the role of the supply of credit in amplifying uncertainty shocks, we replace the Baa–Aaa spread by the domestic bank credit to the private sector as a share of GDP.^[15] We use the cyclical position of the credit-to-GDP – via HP-filtering – instead of its level to isolate the role of financial development in driving the long-term trend in credit.^[16] Thus, the so-called “credit gap” measures the degree of relative abundance (or scarcity) of credit compared to its trend.

As shown in Figure A.9, uncertainty shocks tend to reduce the supply of credit, although the statistical significance is weaker compared to the response of the credit spread. This finding is not surprising given the persistence of the volume of credit compared to its price and echoes with the aggregate and bank-level evidence from Bordo, Duca, and Koch (2016). Importantly, once the response of the supply of credit held constant in the counterfactual exercise, the decline in output is substantially mitigated. While one standard deviation uncertainty shocks are followed by a 1.6% decline in output after three quarters in the baseline analysis, the counterfactual decline is only 0.8%. Overall, our findings suggest that the adverse effect of uncertainty shocks are amplified via both an increase in borrowing costs and a reduction in available credit, thereby further highlighting the role of financial markets.

3.1.6 Relative Importance of Uncertainty Shocks vs. Financial Shocks

So far, we have found robust evidence that a financial channel, proxied by changes in the Baa–Aaa spread or the financial stress index, plays a crucial role in amplifying the adverse effect of uncertainty shocks on the macroeconomy. Regarding the role of uncertainty shocks as a business cycle driver, Born and Pfeifer (2014) concluded that the so-called “pure uncertainty” effect of policy uncertainty is unlikely to play a major role in business cycle fluctuations. Similarly, Ludvigson, Ma, and Serena (forthcoming) found that higher macro uncertainty in recessions is often an endogenous response to output shocks, whereas uncertainty about financial markets is a likely source of output fluctuations.

Although the lack of a full-fledged general equilibrium model in our analysis cannot provide a definite answer to this question, we still gauge the relative importance of each channel in propagating the other shock using a parallel counterfactual exercise. We conduct a similar counterfactual approach but estimate the effect of financial shocks instead after shutting down the so-called uncertainty channel. Under the assumption that the two structural (uncertainty and financial) shocks are correctly identified, this exercise examines whether the uncertainty channel amplifies an adverse financial shock, thereby shedding light on the endogenous interaction between uncertainty and financial shocks. Figure 4 shows the results from the baseline and new counterfactual VARs.

Figure 4:

Effect of financial shocks: Baseline vs. counterfactual.

Note: This graph plots the effects of the one-standard-deviation financial shock and their 68 and 90% confidence bands for the sample period between 1919M1 and 2017M12 when the uncertainty channel is shut down in the counterfactual exercise. The units of the horizontal axes are a month, and the units of the vertical axes are percentage points except for the federal funds rate (basis points).

When compared to Figure 3, two findings stand out. First, unlike uncertainty shocks, financial shocks have a disinflationary effect even in the full sample, implying that higher uncertainty is not simply a reflection of financial market distress if we take a more historical perspective. Such qualitatively different responses call for caution when treating the two shocks interchangeably, which is a common practice in some empirical studies. Second, financial shocks have a significantly negative impact on output, which is much larger than that of uncertainty shocks, and shutting down the uncertainty channel does not mitigate the adverse impact at all.^[17]

Figure A.10 in the appendix shows the results using stock market volatility as a measure for the uncertainty channel. Although the importance of the uncertainty channel increases slightly, the strong asymmetry between the two channels still remains. In this sense, our finding is also related to Ludvigson, Ma, and Serena (forthcoming), who found that uncertainty from financial markets is a dominant source of U.S. business cycle fluctuations, whereas uncertainty about the rest of the economy is rather an endogenous response to other shocks causing business cycle fluctuations. Our finding in Figure A.6 that financial market-based uncertainty has a stronger real effect than policy-based uncertainty regardless of the transmission channel further supports this claim. This result is also consistent with Furlanetto, Ravazzolo, and Sarferaz (2019), who, using a sign-restriction VAR approach, find that shocks originating in the credit markets have larger effects on the U.S. economy than macro uncertainty shocks.

We further evaluate the role of uncertainty shocks in explaining the variation in output, the price level, and the policy interest rate when the financial channel is shut down. The top panel in Table 1 shows the share of variation in these variables explained by the uncertainty shock in different horizons when the financial channel is shut down. The bottom panel reports similar statistics corresponding to the exercise in Figure 4, where we shut down the uncertainty channel when evaluating the role of financial shocks.

Consistent with the evidence presented in the impulse response functions, a large share of macroeconomic variables, especially output, is still explained by the “pure” financial shocks, while only a very small share of these variables is explained by the “pure” uncertainty shocks. If anything, the role of pure uncertainty shocks in explaining output fluctuations diminishes, even more when we compare the historical sample to the recent sample. Again, the results from the forecast error variance decomposition highlight the importance of the financial channel in transmitting uncertainty shocks.^[18]

3.1.7 Constrained Monetary Policy as an Amplification Channel of Uncertainty Shocks

We have established a strong asymmetry between uncertainty and financial shocks in propagating the effect of each other. Now we evaluate the role of constrained monetary policy in amplifying uncertainty shocks. The recent literature has investigated whether the contractionary effects of uncertainty shocks are amplified under the binding ZLB constraint (Basu and Bundick 2017; Caggiano, Castelnuovo, and Nodari 2017; Caggiano, Castelnuovo, and Pellegrino 2017; Nakata 2017; Plante, Richter, and Throckmorton 2018). Leduc and Liu (2016) also argue that a much slower recovery from the Great Recession and the larger effect of uncertainty shocks on unemployment during this period is attributable to the binding ZLB constraint. While these studies often find that constrained monetary policy at the ZLB amplifies the adverse effect of uncertainty shocks on economic activity, the recent ZLB episode coincides with tightened financial market conditions, which are proven to be an independent driver of business cycles as well as a robust amplifying mechanism of uncertainty shocks.

Thus, identifying the role of constrained monetary policy in amplifying the adverse effect of uncertainty shocks crucially hinges on disentangling confounding factors, such as financial market conditions. To address this issue, we conduct a similar counterfactual exercise: we shut down the monetary policy channel through the federal funds rate but allowing the indirect effect of uncertainty shocks through credit spread. In other words, we construct a hypothetical sequence of the shocks that exactly offset the response of the federal funds rate to the uncertainty shock in each horizon, thereby nullifying the expansionary effect of monetary easing following the uncertainty shock.

Interestingly, Figure 5 shows that shutting down the monetary policy channel does not have any material impact on the effect of uncertainty shocks, which is in sharp contrast to Caggiano, Castelnuovo, and Nodari (2017), Caggiano, Castelnuovo, and Pellegrino (2017), who find the amplified effect of uncertainty shocks under the binding ZLB using interacted-VARs. It is true that constructing any counterfactual of the policy response is subject to the Lucas critique (Kilian and Lewis 2011; Sims and Zha 2006). While our empirical model cannot resolve this fundamental issue, we still provide the results from an alternative counterfactual exercise to enhance the credibility of our findings. As a robustness check, we shut down only the direct response to economic policy uncertainty while allowing for the Fed to react to fluctuations in other macroeconomic and financial variables as it normally would.^[19] This alternative counterfactual exercise was used by Kilian and Lewis (2011) in identifying the role of monetary policy in transmitting oil price shocks.

Figure 5:

Effect of uncertainty shocks: Baseline vs. fully constrained monetary policy.

Note: This graph plots the effects of the one-standard-deviation uncertainty shock and their 68 and 90% confidence bands for the sample period between 1919M1 and 2017M12 when the monetary policy channel is fully shut down in the counterfactual exercise. The units of the horizontal axes are a month, and the units of the vertical axes are percentage points except for the federal funds rate (basis points).

Figure 6 summarizes the results from the alternative exercise. Whereas the Fed still lowers the policy rate in response to the aggravating macroeconomy, the response is smaller than the unrestricted VARs. However, even in this case, the effect of uncertainty shocks on other variables remains virtually the same, suggesting that constrained monetary policy is unlikely a major propagation channel of uncertainty shocks. Figure A.11 in the appendix shows the same exercise using stock market volatility as an alternative measure of uncertainty, which confirms that shutting down monetary policy response to uncertainty shocks does not amplify their effect on the macroeconomy.

Figure 6:

Effect of uncertainty shocks: Baseline vs. partially constrained monetary policy.

Note: This graph plots the effects of the one-standard-deviation uncertainty shock and their 68 and 90% confidence bands for the sample period between 1919M1 and 2017M12 when only the direct monetary policy channel is shut down in the counterfactual exercise. The units of the horizontal axes are a month, and the units of the vertical axes are percentage points except for the federal funds rate (basis points).

While this finding seems puzzling at first glance, monetary policy ineffectiveness under heightened uncertainty might explain it. Bloom (2009) notes that uncertainty shocks induce a strong insensitivity to other economic stimuli via an increased real-option value of inaction, making monetary or fiscal policy particularly ineffective. Recently, Aastveit, Natvik, and Sola (2017) and Castelnuovo and Pellegrino (2018) indeed find that monetary policy becomes less effective (i.e., its effect on output and inflation becomes smaller) when uncertainty is high. If the monetary policy becomes ineffective following uncertainty shocks, shutting down this channel would not change the effect of uncertainty shocks on the macroeconomy anyway. To better understand this novel but seemingly puzzling finding, we will revisit this issue in the following section by further exploiting another historical incidence of effective lower bound in the 1930s.

3.2.1 Baseline Local Projections

Jordà’s method simply requires estimating a series of regressions for each horizon h for each variable. Following Auerbach and Gorodnichencko (2012) and Ramey and Zubairy (2018), we run a series of regressions for different horizons, h = 1, 2, …, H as follows:

where, y _t is the dependent variable of interest in time t, α _h is the vector of constants, unc _t is the identified uncertainty shock using the same Cholesky ordering from the structural VARs, and X _t is the set of control variables, including the lags of the dependent variable y _t , the uncertainty index, and other macroeconomic variables. To be consistent with the structural VAR analysis in the previous section, we include the same five variables in the baseline model. We embed the same kind of an identifying assumption here so that uncertainty affects other variables simultaneously, but these variables affect uncertainty only with a lag.^[21]

The coefficient β _h gives the response of y at time t + h to the shock at time t. Thus, we can construct the impulse responses as a sequence of the values of β _h estimated in a series of single regressions for each horizon. This method stands in contrast to the standard method of estimating the parameters of the VAR for horizon 0 and then using those parameters to iterate forward to construct the impulse response functions. The above equation is estimated for h = 0, 1, 2, …, 24, so that we can trace the dynamic effect of uncertainty shocks over two years. To construct confidence intervals, we use the bootstrap standard errors with 200 repetitions. In the baseline analysis, we use six lags of the control variables in X _t (i.e., n = 6) but the selection of the lag length beyond six does not affect our findings in a meaningful way.

Figure 7:

Effect of uncertainty shocks: Baseline local projection.

Note: This graph plots the effects of the one-standard-deviation uncertainty shock and their 68 and 90% confidence bands for the sample period between 1919M1 and 2017M12 using the local projection method. The units of the horizontal axes are a month, and the units of the vertical axes are percentage points except for the federal funds rate (basis points).

Figure 7 visualizes the dynamic responses of the four macroeconomic variables to the uncertainty shock using our baseline local projections. The responses of credit spread, industrial production, CPI, and the federal funds rate are consistent with our baseline 5-variable VAR model shown in Figure 2B, suggesting that the effect of uncertainty shocks does not depend on a particular estimation method.

Figure 8:

Financial states and the NBER recessions.

Note: The shaded region shows recessions as defined by the NBER. The solid red line shows the weight F(z) on the financially tightened state.

3.2.2 State-dependent Local Projections

The local projection method is easily adapted to estimating a state-dependent model (for example, how the effect of uncertainty shocks differs during expansions and recessions), as its application is much more straightforward than that of complex non-linear structural VAR models, such as the Markov-switching, or the interacted/smooth transition/threshold-VAR models used in the empirical literature on the macroeconomic effect of uncertainty shocks (Alessandri and Mumtaz 2019; Bijsterbosch and Guérin 2013; Castelnuovo and Groshenny 2014; Caggiano, Castelnuovo, and Nodari 2017; Caggiano, Casteln uovo, and Pellegrino 2017; Chatterjee 2019).

We estimate a set of regressions for each horizon h as follows to account for smooth transitions between financially tightened and financially relaxed states:

where F(z _t−1) can be interpreted as the measure of the probability of being in a financially tightened state at time t − 1 based on a measure of financial market conditions. Following Auerbach and Gorodnichenko (2012), we lag z _t by one period to minimize contemporaneous correlations between uncertainty shocks and macroeconomic variables.

We still use the spread between Moody’s Aaa and Baa corporate bond yields as the proxy for financial market conditions. To account for a long-run trend in credit spread, we control for both linear and quadratic trends. Moreover, we use the residuals from regressing credit spread on 12 lags of the monthly growth of industrial production to distinguish our financial states from the states based on the state of business cycles when studying the asymmetric effect of uncertainty shocks (Caggiano, Castelnuovo, and Groshenny 2014; Chatterjee 2019). As a result, Figure 8 demonstrates that our financial states are quite different from the NBER recessions and the states identified by Auerbach and Gorodnichenko (2012) or Caggiano, Castelnuovo, and Groshenny (2014). Since our local projection approach based on the financial state is conceptually very similar to the nonlinear-VAR model by Alessandri and Mumtaz (2019), we pay special attention to how our results compare with theirs.

We construct F z t = exp − θ z t / [ 1 + exp − θ z t ] and θ > 0 where z _t is normalized to have unit variance. Thus, F(z _t ) becomes the transition function that ranges between 0 (most relaxed) and 1 (most tightened). We choose θ = 1.5 following Auerbach and Gorodnichenko (2012) and calibrate the mean of z _t such that the economy spends about 20 percent of the time in the financially tightened state.^[22] The parameter θ > 0 governs the smoothness of transition from the financially relaxed state to a tightened state. As θ increases, the transition becomes more abrupt between the states, while setting θ = 0 is equivalent to the linear specification.

We allow all the coefficients in the model to vary according to the state of the economy when the shock occurs. The estimated parameters depend on the average behavior of the economy in the historical sample between t and t + h, given the shock, the initial state, and the control variables. The parameter estimates of the control variables explain the average tendency of the economy to evolve between states. Thus, the estimates incorporate both natural and endogenous transitions from one state to the other, on average, in the data. The coefficients β _G,h and β _B,h trace the dynamic response to uncertainty shocks when the economy is in the financially relaxed and tightened state, respectively.

Following Alessandri and Mumtaz (2019), Figure 9 compares the responses of macroeconomic variables to uncertainty shocks during the financially relaxed state to the responses during the tightened state. Despite the wider confidence intervals in the financially tightened state due to the effectively smaller sample size, the adverse effect on output during that state is much greater, especially in the short-run.^[23] Moreover, the response of prices to uncertainty shocks switches its sign from positive during the financially relaxed state to negative during the financially tightened state. Monetary accommodation is also stronger during the financially tightened state.

Figure 9:

Effect of uncertainty shocks: Financially relaxed vs. tightened state.

Note: This graph plots the effects of the one-standard-deviation uncertainty shock and their 68 and 90% confidence bands for the sample period between 1919M1 and 2017M12 using the local projection method. The black (red) solid line denotes the response to uncertainty shocks during the financially relaxed (tightened) state. The units of the horizontal axes are a month, and the units of the vertical axes are percentage points except for the federal funds rate (basis points).

Figure A.12 confirms that the financial state-dependent effect of uncertainty shocks still holds when using stock market volatility as an alternative measure of uncertainty. Indeed, the dramatic difference in the output effect between the two states is even more pronounced in this case. Figure A.13 repeats the exercise using the supply of credit as an underlying state variable in parallel with the counterfactual VAR exercise above. Consistent with the previous evidence, the adverse effect of uncertainty shocks on output is amplified when the supply of private credit is below the trend. Again, this finding confirms that financial market conditions – measured by both the price and quantity of credit – matter for the transmission of uncertainty shocks.

Overall, all of these findings are perfectly consistent with the findings of Alessandri and Mumtaz (2019), who applied the nonlinear structural VAR with stochastic volatility to the U.S. data from 1973. Despite the different estimation methods, the different measures of uncertainty, and the different sample periods from Alessandri and Mumtaz (2019), we confirm that a strong asymmetric effect of uncertainty shocks depending on the underlying financial conditions is a robust phenomenon.

3.2.3 Uncertainty Shocks and the Zero Lower Bound

In the previous section, we find that shutting down the endogenous response of the policy rate to the uncertainty shock does not alter their effect on other variables in the system, which seems to be a sharp contrast to the recent theoretical and empirical findings that the contractionary effects of uncertainty shocks are amplified when the ZLB constraint is binding (Basu and Bundick 2017; Caggiano, Castelnuovo, and Nodari 2017; Caggiano, Castelnuovo, and Pellegrino 2017; Nakata 2017; Plante, Richter, and Throckmorton 2018). To further elaborate on this puzzling finding, we use a different estimation approach using state-dependent local projections where the additional state is based on whether monetary policy is constrained or not.

As argued before, identifying the role of the binding ZLB constraint in amplifying the effect of uncertainty shocks crucially hinges on disentangling confounding factors, such as deteriorating financial market conditions. Indeed, Figure A.14 in the appendix highlights the fact that the effective ZLB episodes in U.S. history are also characterized by tightened financial conditions. Again, the major advantage of the local projection method over nonlinear VARs is its straightforward extension toward a multi-state situation.

Thus, the local projection method allows us to investigate the effect of uncertainty shocks depending on the joint state based on financial market conditions and whether the ZLB constraint binds. However, estimating the effect of uncertainty shocks based on the joint state requires more data points to draw reliable inferences. If credit spread remained high during most of the ZLB period, the state-dependent local projections must rely on only a handful of observations to disentangle the propagation mechanism of uncertainty shocks through the ZLB channel from that through the financial channel.

The use of nearly 100 years of data is particularly helpful in identifying the effect of uncertainty shocks during the ZLB period since the long historical data offers another interesting episode during which the U.S. economy was effectively in the liquidity trap (1934–1939). Of course, since the federal funds rate was not a monetary policy instrument at the time, we do not define the ZLB episode solely based on the interest rate behavior. Instead, we define the earlier ZLB episode based on the historical perspective. For example, according to Krugman (1998), the U.S. interest rates were “hard up against the zero constraints.” From 1934 through the outbreak of World War II in 1939, overnight rates were effectively zero, and Treasury bill rates were usually below 25 basis points (Hanes 2006).^[24]

We first analyze whether the adverse effect of uncertainty shocks are amplified under the binding ZLB constraint. In so doing, we replace F z t in Eq. (3) with a binary indicator I _t , taking a value equal to one when the ZLB constraint binds (March 1934 to August 1939; December 2008 to November 2015) and zero otherwise. Figure 10 shows that the short-term interest rate does not respond to the uncertainty shock during the binding-ZLB state, lending support to the correct identification of the ZLB episodes. Despite the wider confidence intervals due to the small sample size of the ZLB episodes, the adverse effect on output becomes much larger during the binding-ZLB state. Unlike the baseline results, the uncertainty shock is disinflationary during the binding-ZLB state, suggesting that they act as a negative aggregate demand shock when the ZLB constraint binds. Figure A.15 in the appendix also confirms the amplified real effect of uncertainty shocks when using stock market volatility as an alternative measure of uncertainty.

Figure 10:

Effect of uncertainty shocks: Baseline vs. binding ZLB constraint.

Note: This graph plots the effects of the one-standard-deviation uncertainty shock and their 68 and 90% confidence bands for the sample period between 1919M1 and 2017M12 using the local projection method. The solid black line denotes the response to uncertainty shocks in the baseline estimation, while the solid red line denotes the response during the binding ZLB constraint. The units of the horizontal axes are a month, and the units of the vertical axes are percentage points except for the federal funds rate (basis points).

While these findings corroborate the existing studies focusing on the recent ZLB episode, the macroeconomic variables’ documented responses resemble those during the financially tightened-state (Figure 9), casting doubt on the independent role of the binding ZLB constraint in amplifying the effect of uncertainty shocks. Thus, we test this possibility by considering the following multi-state local projection method, which extends Eq. (3) by augmenting a binary indicator variable I _t :

where the BZ, BN, GZ, and GN state denotes “the financially tightened and binding ZLB state,” “financially tightened and non-binding ZLB state,” “financially relaxed and binding ZLB state,” and “financially relaxed and non-binding ZLB state,” respectively.

To gauge the relative role between financial market conditions and the binding ZLB constraint in propagating uncertainty shocks, we compare the effect of uncertainty shocks during the BN state and the GZ state. The top panel in Figure 11 shows that the overall effect of uncertainty shocks during the BN state is similar to the financially constrained state: the adverse effect on output is still larger than the baseline, and the price response is significantly inflationary. However, the bottom panel shows that the effects during the GZ state differ sharply from those in the binding ZLB state shown in Figure 10. Uncertainty shocks are no longer contractionary, and their effect on prices becomes inflationary, which corroborates the evidence from the counterfactual VAR exercises. As long as financial conditions remain sound, the binding ZLB constraint alone fails to amplify the negative effect of uncertainty shocks. Figure A.13 in the appendix confirms that this finding still holds when using stock market volatility as an alternative measure of uncertainty.

Figure 11:

Effect of uncertainty shocks: Financial tightened/non-ZLB state (top panel) vs. financially relaxed/ZLB state (bottom panel).

Note: This graph plots the effects of the one-standard-deviation uncertainty shock and their 68 and 90% confidence bands for the sample period between 1919M1 and 2017M12 using the local projection method. The solid black line denotes the response to uncertainty shocks in the baseline estimation, while the solid red line denotes the response during the financial tightened/non-ZLB state (top panel) and the financially relaxed/ZLB state (bottom panel). The units of the horizontal axes are a month, and the units of the vertical axes are percentage points except for the federal funds rate (basis points).

Figure 11:

(Continued).

We conclude that the amplified adverse effect of uncertainty shocks during the binding ZLB constraint is largely masked by the simultaneous occurrence of tightened financial conditions during the same period. By no means are we trying to disregard the recent theoretical and empirical attempt to understand a propagation mechanism of uncertainty shocks using the ZLB episodes. The nature of our empirical analysis cannot rule out the endogenous interaction among uncertainty, financial conditions, and monetary policy such that the binding ZLB constraint tightens financial conditions, which eventually amplifies uncertainty shocks.

However, we also observe that unconventional monetary policy, such as quantitative easing, was at least partly successful in alleviating financial market distress (Baumeister and Benati 2013; Dell’ Ariccia, Rabanal, and Sandri 2018). Swanson (2018) also claims that the Federal Reserve has not been significantly constrained by the lower bound on nominal interest rates owing to its alternative policy tools. In this regard, the new evidence from the historical data using nonlinear methods calls for developing a structural model, which allows the identification of the role of the binding ZLB constraint in propagating uncertainty shocks separately from the direct influence of financial conditions.