VoxEU Column Monetary Policy

Identifying monetary policy shocks: A natural language approach

The recent surge in inflation has put central banks back into the spotlight. This column proposes a novel method to determine exogenous changes in monetary policy. Using information in the language of documents that economists at the Federal Reserve Board prepare for Federal Open Market Committee meetings, it predicts changes in the target interest rate and obtains a measure of monetary policy shocks as the residual. The dynamic responses of macroeconomic variables to the identified shock measure are consistent with the theoretical consensus, and the estimated shocks are not contaminated by the ‘Fed information effect’.

The recent surge in inflation has put the Federal Reserve back in the spotlight of the public conversation. How should monetary policy be conducted in the current economic environment? To answer this question, one needs to understand the causal effect of monetary policy on the economy. Since the Fed adjusts its policy based on changes in the economic outlook, this is not straightforward. One approach that macroeconomists have adopted involves looking at past data and isolating changes in interest rates that are not a response to economic conditions, but instead arguably occur exogenously. 

Romer and Romer (2004) suggest measuring exogenous movements in the federal funds rate as the difference between observed and intended changes in the rate. Intended changes are based on the economic outlook of policy makers at the time of their decisions. Romer and Romer (2004) use numerical forecasts of inflation, output, and unemployment contained in the ‘Greenbook’ documents prepared by Fed economists for Federal Open Market Committee (FOMC) meetings. This approach is applied in subsequent research, for example by Tenreyro and Thwaites (2013) and Coibion et al. (2014).

In a recent paper (Aruoba and Drechsel 2022), we propose a novel approach to identify monetary policy shocks. We follow the idea of exploiting the information in documents prepared by Fed economists for the FOMC. Our method, however, aims to capture the information contained in these documents more comprehensively, by including numerical forecasts as well as human language. We do so with natural language processing and machine learning methods.

A new method to identify monetary policy shocks with language

We estimate monetary policy shocks as the residuals from a prediction of changes in the federal funds rate using (i) numerical forecasts in the documents that Fed economists prepare for the FOMC, (ii) the verbal information in the documents, and (iii) nonlinearities in (i) and (ii). To obtain (ii), we first identify the most important economic terms in the documents. This results in a set of 296 single or multi-word expressions, such as ‘inflation’, ‘economic activity’, or ‘labour force participation’. Figure 1 illustrates this with a word cloud of the 75 most frequently mentioned economic concepts between 1982 and 2017. The size of each concept reflects the frequency across the documents.

Figure 1 Most frequently mentioned economic concepts, 1982-2017


We then construct sentiment indicators that capture the degree to which these economic concepts are associated with positive or negative human language. The basic idea is to count the number of positive or negative expressions, based on a pre-defined dictionary, that appear in proximity to each economic concept. Our collection of 296 sentiment time series paints a rich picture of the historical assessment of economic conditions by Fed economists. Figure 2 shows the sentiment surrounding ‘economic activity’ as an illustration. This time series reflects meaningful business cycle variables, contracting sharply in recessions.

Figure 2 Sentiment surrounding ‘economic activity’ concept


A regression with federal funds rate changes on the left-hand side and (i), (ii), and (iii) on the right-hand side is infeasible given that there are many more regressors than observations. To overcome this issue, we resort to machine learning techniques: we employ a ridge regression to predict intended changes in the federal funds rate using our large set of regressors. The idea of a ridge regression is to minimise the residual sum of squares and an additional term that penalises squared deviations of each regression coefficient from zero.

Systematic versus exogenous changes in interest rates

Most economists would argue that monetary policy is conducted in a highly systematic manner. An appealing feature of our procedure is that, in light with this perception, only a small fraction of interest rate changes is attributed to exogenous shocks. Our ridge regression implies that the systematic component of monetary policy explains 76% of the variation in the target interest rate, while 24% of the variation is attributed to shocks. Compared to existing applications of Romer and Romer’s idea, the systematic component is significantly more important in our approach. A larger set of forecasts, Fed economists’ sentiments, as well as nonlinearities all contribute to capturing the systematic component of monetary policy more comprehensively. 

In our paper, we also verify whether including additional information in our ridge regression alters our measure of shocks, by using information on the Fed transcripts and the personnel composition of the FOMC. We find that our measure of shocks is not explained by information beyond that made available to FOMC members by the Fed staff at the beginning of a meeting.

Inspecting the identified shocks

The dark blue line in Figure 3 plots the time series of monetary policy shocks estimated with our methodology. The figure compares this with the estimated residuals from Romer and Romer as the lighter orange line. Our measure of monetary policy shock displays a generally lower volatility and a lower degree of autocorrelation. They are not simply a scaled-down version of the shocks implied by the original Romer-Romer method. In many instances, the orange line implies a larger shock in absolute terms, while at other points in time larger shocks are visible for the blue line.

Figure 3 Time series of monetary policy shocks


For those episodes where monetary policy shocks are particularly large in magnitude, we closely inspect the discussion that took place in the FOMC. This sheds light on what estimated monetary policy shocks capture. The largest exogenous easing is estimated for the 7 November 1984 meeting of the FOMC. This is a period that has a mixed economic outlook: industrial production has declined for the first time in two years, yet investment and consumption show robust increases. The Fed staff concludes that the “slowdown may only be a pause in a recovery that has not run its full course.” When we read the transcript of the FOMC meeting, it becomes very clear that many participants find the staff forecast too optimistic. Their policy actions are consistent with a sizable easing of policy, providing a great example of a situation where the FOMC’s views about the economy are different from that of the staff economists. It is important to emphasise that this is an unusual situation. If the disagreement happened more often, then our procedure would have picked it up and the changes in policy would be predicted.

The largest exogenous tightening happened in the 15 November 1994 meeting. The Fed staff argue that the economy is above its full capacity with the inflationary consequences not yet realized. They propose two policy options: a no change option and one where the federal funds Rate increases by 50 basis points. During the FOMC meeting, Chairman Greenspan suggests that since the market already built in a significant rate hike “a mild surprise would be of significant value.” He proposes a rate increase of 75 basis points to get “ahead of general expectations.” Most of the participants agree with this proposal, with several participants emphasising credibility of keeping inflation under control. Once again this is a situation where the FOMC decided on an action not simply based on the current economic outlook but also other considerations. Our procedure therefore implies that this reflects a monetary policy shock.

The effects of monetary policy shocks on the economy

With our novel measure of monetary policy shocks at hand, we study impulse response functions (IRFs) of macroeconomic variables in a state-of-the-art Bayesian vector autoregression, estimated from October 1982 to 2016. The results for a monetary tightening are presented in Figure 4. The two panels of the figure show the IRFs of bond yields, stock prices, real GDP, the GDP deflator, and the excess bond premium, based on our shocks (blue) and shocks constructed using the original Romer-Romer methodology (orange).

Figure 4 Effects of monetary policy shocks on the economy



Using our monetary policy shocks, we find that a monetary tightening leads to a reduction in production activity and a fall in the price level, in line with what economic theory predicts. This contrasts with IRFs to the shocks constructed from the original Romer-Romer specification, where a monetary tightening appears to have no significant effect on economic activity. Earlier findings already suggest that more recent samples imply IRFs to monetary policy shocks at odds with theory, as discussed in Ramey (2016). One interpretation is that some systematic policy variation may still be present in shock measures constructed purely based on numerical forecasts. Figure 4 indicates that the novel method we develop overcomes this problem by including a larger set of information.

Finally, our shock measure does not appear to be subject to the ‘Fed information effect’ (Nakamura and Steinsson 2018). Jarocinski and Karadi (2018) argue that a monetary tightening should raise interest rates and reduce stock prices, while the confounding positive central bank information shock increases both. Figure 4 indicates that our shock measure leads to an interest rate increase and a fall in stock prices. We conclude that natural language processing and machine learning are useful to deliver a cleanly identified estimate of monetary policy shocks.


Aruoba, B and T Drechsel (2022), “Identifying Monetary Policy Shocks: A Natural Language Approach”, CEPR Discussion Paper No. 17133.

Coibion, O, Y Gorodnichenko, L Kueng and J Silvia (2014), “Innocent Bystanders? Monetary policy and inequality”,, 25 October.

Jarocinski, M and P Karadi (2018), “The transmission of policy and economic news in the announcements of the US Federal Reserve”,, 03 October.

Nakamura, E and J Steinsson (2018), “High-frequency identification of monetary non-neutrality: the information effect”, The Quarterly Journal of Economics 133: 1283–1330.

Ramey, V A (2016), “Macroeconomic shocks and their propagation”, Handbook of Macroeconomics 2: 71–162.

Romer, C D and D H Romer (2004), “A New Measure of Monetary Shocks: Derivation and Implications”, American Economic Review 94: 1055–1084.

Tenreyro, S and G Thwaites (2013), “Pushing on a string: US monetary policy is less powerful during recessions”,, 12 November.

3,150 Reads