 Let me introduce Anna Cieslak. So Anna, you are Associate Professor of Finance at Duke University, Fugar School of Business in the US. And we also share you and me, one thing in common is that I think you worked for the Bank of International Settlements in Basel as a research fellow in 2014 and 2019. And you conduct research in macrofinance and empirical asset pricing with emphasis on fixed income market. Your work has been published in leading academic journals, including journal finance, review of financial studies, and Journal of International Economics. So Anna, the floor is yours for 30 minutes and we'll be following up with 10 minutes of Q&A. All right, thank you very much for inviting us to present. So let me just start with a couple of quotes from former chairs of the Fed, who think about uncertainty as the pervasive feature or defining characteristics of the landscape and if they are operating and they are making decisions. And in fact, Philip, in his intro remarks, also mentioned the intrinsic uncertainty that is out there that policymakers are facing. So we ask a very simple question, how actually uncertainty affects policymaking? And there are few strict results that would allow to answer this question cleanly. On the theoretical front, we really have a multitude of predictions that depend on model specification. And from the empirical standpoint, both measurement of uncertainty is challenging as well as identification of the causal effects of uncertainty is difficult. So what we are going to do in this paper is to propose a measurement of policymakers' uncertainty using text. We'll refer to those measures as PMUs for policymakers' uncertainty and we'll rely on a detailed analysis of the FOMC transcripts. We'll distinguish between different sources of uncertainty that policymakers face, thinking about real and nominal variables, financial markets, and models. And finally, we will analyze the effect that PMU has on policymaking in terms of policymakers' preferences and obviously these preferences then propagate onto the policy surprises that Michael has been talking about. So what are the channels through which in general uncertainty could affect policymaking? So let's just think about the very standard reaction function where the Fed has some instruments and is responding with a set of phi coefficients to the state of the economy described by this omega. There are three broad types of reactions or effects that uncertainty could induce on monetary policy. So the first one is particularly uninteresting because in many standard models, in linear quadratic models, the certainty equivalence principle would imply that uncertainty has no effects on policymaking. More interestingly, there is obviously a large literature that thinks about uncertainty as being a demand shock and then the policy starts reacting to that shock to the extent that it affects the state of the economy that policymakers perceived, but not beyond that. And then finally, there is a large literature studying the uncertainty that policymakers face about parameters and models and or models. And here we really have quite different predictions depending on what source of uncertainty is. So in particular, when policymakers are unsure about policy multipliers, that is how they would affect the economy, this could lead to a more cautious response. And this is what is known as the Brunner's Conservatism principle. However, it turns out that as soon as we switch on uncertainty about the economic dynamics that policymakers face, for example, inflation persistence, or we allow policymakers to be unsure about the correct model specification, then this could induce a more aggressive response. So we really don't know, given the existing evidence, how policymakers behave in the face of uncertainty. So to make progress on this question, we start by proposing a measurement approach to understand first policymakers uncertainty and then how it affects the decision making. So we'll use the setting of the FOMC, the US Federal Reserve, where we have really a wealth of information that will allow us to construct proxies for policymakers beliefs, their uncertainty and their preferences in a way that is mutually consistent and such consistency is rarely feasible in other contexts. We'll observe speakers' statements at the sentence level in the FOMC transcripts over more than three decades of data. And we will also have a rich set of controls for the first moments that is beliefs and forecasts about the state of the economy. To achieve identification of uncertainty effects on policymaking and therefore on the surprises that we see in the markets, we'll exploit the regular structure of the FOMC meetings that has been there for the entire span of our sample periods. In particular, we'll use the discussions in the economy round to identify the different types of uncertainty that policymakers are faced with. And then we will ask how do these different types of uncertainty affect the policy preferences that FOMC members express in the policy round. And the assumption here is that the economy round discussions are free of expression of how policy itself could induce uncertainty in the economy. So it is purely uncertain to the policymakers face as they go into the meeting. The key assumption underlying our construction of uncertainty indices is that the PMUs correlate with the frequency and intensity with which policymakers express uncertainty in the meeting. And we'll use tools from computational linguistics to first construct measures of overall uncertainty in the economy round, and then to separate out different types of uncertainty that policymakers face. We also will obtain a set of textual controls for variation in what we call sentiment or directional language about the course of the economy. So we take a relatively broad view on what uncertainty means. And this is reflective of the challenges that policymakers themselves face when they think about uncertainty. In theory, we have a pretty clean sense and distinction between the notion of risk and the notion of the so-called nightian uncertainty. In practice, however, one is never quite sure what type of uncertainty one is dealing with. And so it is useful to think about that as a continuum. And so to reflect those practical difficulties that policymakers face will take a broad approach to measuring uncertainty. And we'll start with word embeddings focusing on two terms, terms related to risk and terms related to uncertainty. And we will estimate a model that tells us what are the close synonyms appearing in the economy around to risk and uncertainty. And here in this table, I'm showing you a set of estimates that come out from this exercise. Interestingly, risk embeddings are associated with quantifiable notions of risk, that is probability, likelihood, odds, whereas uncertainty embeddings have a relationship with anxiety, angst, interestingly, ambiguity, and so on. So there does seem to be a separation in language as to how policymakers talk about risk vis-à-vis uncertainty. But then given this broad approach that we take, we'll combine those two and only at certain point try to see whether there is any significant distinction in how policymakers think about the two dimensions of uncertainty. So we start by constructing an overall measure of policymakers uncertainty, the PMU index, simply as being the intensity in which the uncertainty risk terms appear in the policy round relative to all the words that are being expressed in that round. And here you can see the overall dynamics of our PMU index, which is expressed as fraction of uncertainty words in overall length of the economy round. So you see that they occupy probably 1.5% of all words in that round. And you can see that there are some usual suspects where uncertainty spikes, such as financial crisis, or wars, here we have Iraq war. In general, policy makers uncertainty seems to be higher in a more recent part of the sample and declines in the last part of the sample. The data is rich. And so we could disaggregate the data in a variety of ways. There is significant heterogeneity in expressed uncertainty across different members of the FOMC. And here you can see the, their own expression of uncertainty in the meeting compared to what is happening in the meetings that they are present on average. And again, there is substantial heterogeneity, the richness of which allows us to study also individual level effects of uncertainty on policy preferences. In this talk, however, I will focus more on the aggregate results. Now, obviously policy makers don't face just one source of uncertainty. And for our purposes, it turns out to be really difficult to be really important to separate the different types of uncertainty that policy makers are facing. We'll focus on four types of uncertainty that occupy the bulk of space in policy deliberations. The inflation uncertainty related to nominal variables, the real economy uncertainty related to the real economy, labor markets, productivity, et cetera. And financial markets uncertainty that policy makers also discuss frequently in the meeting can pay a lot of attention to. And the final category of explicit topics will be the model uncertainty, which by the way, Philip also mentioned talking about intrinsic uncertainty and structural change. So this is exactly what this category would be picking up. We have an unclassified category of uncertainty mentions. These are all the mentions that we fail to attribute to any of the four other categories. So let us take a look at the time series dynamics of these different uncertainty types or topic specific PMUs. And what you can see here is again, uncertainty indices expressed on the Y axis as a fraction of time that is being spent discussing a particular type of uncertainty. And so we are able to classify about 84% of all mentions of uncertainty into our four explicit topic categories. And so the bulk of dimensions will be associated with inflation, real economy and financial markets. Interestingly, the inflation and the economy uncertainty for example are only very weakly correlated with each other having a correlation of about 0.1. So now let me establish a few facts about the PMUs, policymaker uncertainty before we jump into discussing its implications for policymaking. So the fact number one that we establish is that PMU is not really clearly counter cyclical. And this contrasts with the perception of what typical proxies of public perception of uncertainty do over the business cycle. In particular, we find that policy makers become uncertain about the real economy. In fact, ahead of the financial crisis. Similarly, they become uncertain about financial markets about a year before the shock becomes visible in financial data. But then during the heights of the financial crisis their uncertainty actually goes down. Interestingly, we find that inflation uncertainty is highly pro cyclical that this policy makers become really uncertain about inflation when the economy is doing well. Again, this contrasts with the standard notion of how uncertainty evolves over the business cycle. The second set of results pertains to how uncertainty correlates with sentiment. So typically we think about uncertainty as reflecting the range of possible outcomes but not being really directional. What we find systematically across different measures of uncertainty that we construct is that uncertainty is in fact highly correlated with negative tone and negative sentiment in the meetings. So negative sentiment for inflation in the left panel means that policy makers at least during that whole sample period express negative views about inflation going up. And this tends to correlate very strongly with their expression of uncertainty as well. Even though in the construction of the two measures that is the sentiment and the PMU we use completely disjoint set of sentences. So the relationship is not mechanical. In particular for inflation uncertainty we show that neither the PMU nor the negative sentiment about inflation in the meeting has any predictive power to what actually happens to inflation subsequently. So we interpret this relationship between PMU and sentiment as a reflection of policy makers concern about rising inflation and that actually does not materialize during the sample we study. Of course we are constrained by the availability of the transcripts data. We don't have the last five years of data but that relationship holds pretty strongly during the years before 2015, up to 2015. What you can also see is that this negative sentiment and uncertainty are very strongly correlated in terms of policy makers thinking about financial markets developments. And as I mentioned, policy makers uncertainty about financial markets is sort of anticipating what happens in the public domains in that uncertainty in the meetings about the course of financial markets is high is strongly elevated about a year before the financial crisis shock actually materializes in the data. Now our next set of facts about uncertainty pertains to their relationship with forecasters. So what you would expect to happen is that policy makers become increasingly uncertain about the economy when their models start failing. And this is exactly what we find. In particular, we find that for inflation, policy PMU about inflation is actually strongly positively correlated with the magnitude of forecast errors about inflation that materializes in the Fed's Green Book forecast. So this suggests that the uncertainty is related to worry and concern about model mis-specification or models not delivering accurate forecasts. We don't find any relationship for the real uncertainty for the economic PMU and the magnitude of past forecast errors, although there is some weak relationship that is directional. However, for the real economy and the markets, past forecast errors are able to explain only about 4% of variation. The relationship is much stronger for inflation. So to summarize, it seems like policy makers become increasingly uncertain when they sense that the models are not accurately forecasting the economy. What about the relationship between perceived risk and uncertainty? So the distinction that I started off with based on Greenspan quote. So in this plot, you can see the topic-specific PMU indices based only on risk-related phrases. And we superimpose that with expression of uncertainty. Again, the set of terms entering the two is completely disjoint. And what you can see is that there is a very strong positive relationship between expression of quantifiable risk and expression of this more unquantifiable 19 types or type of uncertainty in the meeting. Consistent with Greenspan statement that the two notions are really hard to disentangle in practice. There is also significant relationship between our PMU index and measures of public perceptions of policy uncertainty. So here we use the Baker-Bloom and Davis index as well as the Husted Rogers and Sun indices that are aimed to reflect how uncertain the public is about the course of monetary policy. And so we do find positive relationship. However, the overall index that we used on the previous slide camouflages quite heterogeneous relationship between public perceptions of uncertainty and internal uncertainty regarding specific components of the economic environment. So in particular, we find that while the uncertainty about the real economy does correlate positively with the public perception of uncertainty about policy, it is the inflation uncertainty that policymaker face is in fact negatively correlated with both Baker-Bloom and Davis and the Husted Rogers and Sun measures. Well, again, consistent with the fact that policymakers become really uncertain about inflation in good times. So now our main set of results pertains to the question of how uncertainty affects policy preferences or whether it does affect policy preferences at all. And we start with a textual measures with a construction of textual measures of policy preferences trying to capture the nature of policy shocks that emanate from the discussions happening in the meetings. We develop a set of rules to classify sentences in the policy round of the meeting into a policy language. And we then focus on statement by FOMC members only to really zoom into the preferences of actual decision makers at the Fed. We separate their language into hawkish and dovish slant in terms of preferences by very precisely matching policy terms with directional language. And then at each meeting, we construct a policy preferences variable which is a balance of hawkishness and dovishness intensity expressed at each meeting. And this variable will be our key measure of policy preferences in the meeting that we'll try to match with proxies for uncertainty. Now, in terms of the properties of these hawk and dovish measures, they are in fact quite intuitive. Dove preferences spike in downturns as we would expect. hawkish preferences are elevated during expansions. And when you stare at the balance measure, it in fact shows that there is, that remains significant variation in policy preferences during the zero lower bound period. So one advantage of our approach based on text is that it allows us to study a long consistent sample period, including the zero lower bound period. And in fact, it allows us also to connect to a notion of a monetary policy surprise which admittedly is quite elusive in terms of available interpretations in the literature. So these textual measures of policy preferences turn out to be in fact very highly informative about what happens to actual policy outcomes in terms of the target changes or for example, Romer shocks. The signs of the loadings with which those variables load on hawk and dovish scores and then our balance variable, the HD variable are intuitive. These measures are highly significant and they are highly significant in presence of controls for green book expectations. So we interpret those as a measure of actual policy surprise based on that emerges from deliberations in the policy meeting. Quite interestingly, these pure text-based measures of policy preferences explain about a quarter of Romer and Romer shocks that are frequently used for policy analysis. Related to what Michael has been presenting, the hawk dovish scores are also highly predictive of the information that markets, the financial markets glean from the policy announcements. You see that across a range of measures, including GSS, Gorkaniak, Sakon, Swanson, Gertler, Karadi, Nakamura and Steinsen, you see a positive loading across these different measures. We show that the text of the transcripts in fact is highly predictive of the path of monetary policy, a number of quarter ahead, suggesting that it really has a lot of forward-looking content to it. So now our goal will be to think about how those policy preferences measured from the text are related to the uncertainty that policymakers express in the economy round of the meeting. So we start with a very simple specification where we regress the HD variable on overall policy index and then on its sub-components that we identify with our topic-specific approach. So overall, an increase in uncertainty in the economy round predicts a more dovish tone or a more dovish stance of preferences during the policy round. But that aggregate index, again, masks much more interesting effects that emerge when we disaggregate the uncertainty by topics. So looking first at the economy and markets PMU, we find negative signs that are significant, whether or not we control for topic-specific sentiments that is directional language in the meeting or green book forecast prepared by the staff. This negative coefficient is consistent with the broad demand channel of uncertainty whereby increased uncertainty would lead to weakening of the economy and therefore a policy response that is more dovish. However, the fact that these effects are persistent even if we control for green book forecasts suggests that those green book forecasts may not entirely take the effect of uncertainty through the demand channel on board. Interestingly, and in contrast to the coefficients on the real economy PMU and markets, for inflation PMU, we find a strongly positive response. And this is parallel with the uncertainty about model specification also inducing a positive response of policy preferences. Positive coefficient means that increased uncertainty of about either inflation on models would lead to more hawkish preferences. Finally, the unclassified types of uncertainty are not contributing to predictive power for the policy preferences in a significant way. Now, this was a simple linear specification, but the models of optimal monetary policy choice consider frequently a more complicated set of effects that are of multiplicative nature. That is, uncertainty alters the strength of policy response to the economy. Similar to what Michael has been alluding to that the markets may not anticipate correctly the strength of the policy response. So one factor that could be driving the strength of response could be uncertainty as predicted by a variety of models. The theoretical predictions, however, how uncertainty would affect policy are really highly model dependent. So the question is in the data, how does uncertainty actually affect policy makers response? Does it lead to strengthening or weakening of that response? And do these effects differ across the different state variables that policy makers care about? So what we will do is to estimate a bunch of text-based policy rules in which we'll allow interactions. We'll essentially allow the coefficients of the policy reaction function to depend on the extent of uncertainty that policy makers face. So for example, uncertainty about economy, about inflation could alter the strength of the response that policy makers have to fluctuations in inflation and similarly for real GDP growth. It could be that higher uncertainty weakens the response or strengthens the response to expected growth. So let's estimate these regressions and let's see what we are finding when allowing for interactions. The main effect that we find across different specifications is that uncertainty about inflation would amplify the policy makers response to inflation. And this is probably the most robust feature that we find across different specifications. We do find some effect of economic uncertainty that is the uncertainty about the real economy also leading to an amplification of growth response but this is much more dependent on the different types of controls we include in the specification. So let us take a look at how these effects play out in a graph. So what these graphs are showing you is the effect of inflation on policy preferences as a function of the amount of uncertainty that policy makers are facing. First uncertainty about inflation. Second uncertainty about the real economy. And the effects we are looking at is those of expected inflation on policy preferences and expected real GDP growth on policy preferences. In particular for inflation, what we find is that moving from an environment in which uncertainty about inflation is slow to where uncertainty about inflation is high leads to quite dramatic strengthening of the policy response to inflation. And this is the main effect that we have found before. We do find some strengthening, some amplification of policy response in reaction to growth given different levels of uncertainty about the real economy. However, this effect is much more sensitive to controls for the first moments. Now, overall, the finding of the amplification effects of uncertainty on policy reaction is inconsistent with the frequently quoted Brynham's conservative principle, conservatism principle, whereby a policy makers that is uncertain about what the effects of their actions would be on the economy might want to behave more cautious. In contrast, at least the qualitative effects that we find are consistent with two types of channels. One is policy maker displaying preference for robustness and or policy makers being uncertain about the dynamics of the economy, in particular about the persistence of the inflation process and acting more aggressively in order to prevent costly outcomes. So does this have any reflection in how policy makers talk about uncertainty that they face and then what type of reactions this would induce? And here's a quote for Peter Pratt who in fact acknowledges that exact amplifying effect that we find empirically that policy makers may want to be aggressive when they face a risk of inflation becoming ingrained in the economy, the risk of inflation becoming this anchored and the aggressive action should then serve as a signal of policy makers commitment to its objectives and as a notch for expectations to become more anchored. So let me conclude what we do in this paper, we proposed a bunch of new text based measures of policy makers uncertainties of different kinds, not just one, which we call PMU and a new measure of policy preferences that gets us to the question, where do policy shocks actually come from? And we for identification of effects of uncertainty on policy preferences, we exploit the sequential nature of deliberations in FOMC meetings that structure the economy around deliberations before any policy deliberations take place. And our main finding pertains to the fact that inflation uncertainty just as we face it today will lead to amplification of policy response to fluctuations in inflation. This effect is inconsistent with Brynard's conservatism, but may suggest policy makers concern about model mis-specification structural change in the economy preference for robustness. And we also established this interesting asymmetric relationship between PMU and policy makers concern at least in our historical data up to 2015 policy makers concern about rising inflation. Also the independent additional effect of uncertainty on policy preferences is actually suggesting that for a long period there might have been deviations from symmetric that this quadratic preferences in policy making. There seems to be additional demand shock channel of uncertainty that is at work. And that demand shock channel seems to go beyond what is taken into account in green book forecast. So thank you very much. And I look forward to your questions.