The politics of inflation diagnosis at the Fed
London School of Economics
Université Paris I
Hertie School
Sciences Po
PhD on Green central banking
Postdoc on Central bank politics
Not a typical paper in history of economic thought
We look at practitioners:
How central bankers read the macroeconomic situation, which theories they mobilize (often implicitly), and what else shapes that reading.
Paper 1 — Interests and prices
Paper 2 — Fear of full employment
Inflation is (relatively) easy to capture, but its causes lead to heated debates based on competing explanations
Controversies, even decades later!
Yet this question cannot be avoided by central bankers
No matter the actual drivers of inflation, their reading by central bankers will have material consequences!
Research questions:
Two long-term trends may push central bankers closer to a technocratic ideal:
Yet de jure independence is not de facto insulation:
Independence does not suppress politician pressures, nor removes politics entirely
Externally: reputation management
Internally: individual preferences
The way central bankers read and narrate the causes of inflation may therefore not only be driven by economic fundamentals, but also by external and internal politics
Public speeches (1970–2023)
Private deliberations (1976–2019)
Our goal: detect, categorize, and track what factors central bankers blame for inflation in their private deliberations and public speeches.
However, central bank talk creates to identification challenges:
“On the inflation number […] one of the major factors is the effect of the structural deficit plus the projected decline of the dollar.”
Robert Forrestal, FOMC meeting, 1984
“Demand for workers far exceeds the supply of available workers, and nominal wages have been growing at a pace well above what would be consistent with 2 percent inflation.”
Jerome Powell, speech, 2022
Legacy options fall short
We opt for a Large Language Models (LLMs) classification pipeline:
Not a magical solution:
Codebook
| Driver | Captures |
|---|---|
| Labor | wage-price spirals, employment, unions |
| Market power | pricing, competition, profits |
| Fiscal | spending, deficits |
| Energy | oil, gas, electricity |
| Commodities | food, raw materials |
| Exchange rate | dollar strength / weakness |
Inter-coder agreement
12 LLMs vs the human gold standard
12 LLMs vs the human gold standard
Over the whole corpus:
Yearly shares among the six drivers (points), loess smoothing with 90% confidence bands. FOMC transcripts 1976–2019, speeches 1970–2023.
Public speeches and private deliberations systematically favor different explanations:
This is consistent with a reputation management strategy (Diessner, 2023; Moschella, 2024)
PhD economists vs others
PhD economist: terminal degree is a PhD in economics or finance. Shares among the six drivers.
Does the PhD school matter?
PhD economists only.
First, it is the most prominent driver of inflation in Fed discourse
It clashes with the Fed’s dual mandate: full employment becomes a threat to price stability
It rests on central but contested theories: the Phillips curve and the NAIRU
→ Given these features, can we explain the variation in Labor blaming?
First, let’s look at economic fundamentals:
Phillips curve / NAIRU inspired thinking sticks despite growing criticism (Arbogast et al., 2023; Tarullo, 2017)
This “fear of full employment” is baseless, since the Phillips curve is dead (Galbraith et al., 2007; Ratner & Sim, 2022)
Formally (negative binomial, quarterly)
\[ \begin{aligned} \log E[\text{Labor}_t] = {\color{#8f8880}{\beta_0}} &+ {\color{#7b3fa0}{\beta_1 U_{t-1} + \beta_2 U^2_{t-1}}} + {\color{#00a88f}{\Omega W_{t-1}}} \\ &+ {\color{#8f8880}{\Gamma Z_{t-1}}} + {\color{#8f8880}{\beta_3 N_t + \alpha_c}} \end{aligned} \]
| (1) | (2) | (3) | (4) | (5) | |
|---|---|---|---|---|---|
| Unemployment |
−0.200*** (0.042) |
−0.200*** (0.045) |
−0.183*** (0.046) |
−0.192*** (0.043) |
−0.183*** (0.040) |
| Unemployment² |
0.060*** (0.014) |
0.060*** (0.014) |
0.059*** (0.014) |
0.058*** (0.015) |
0.055*** (0.014) |
| Nominal wage growth |
0.034 (0.038) |
||||
| Real wage growth |
0.021 (0.035) |
0.027 (0.034) |
|||
| Inflation |
0.006 (0.023) |
0.024 (0.031) |
0.037 (0.032) |
||
| Real GDP growth |
0.032 (0.024) |
||||
| Word count (log) |
0.542** (0.180) |
0.546** (0.182) |
0.545** (0.182) |
0.543** (0.185) |
0.556** (0.176) |
| Observations | 167 | 167 | 167 | 167 | 167 |
| Chair dummies | ✓ | ✓ | ✓ | ✓ | ✓ |
Negative binomial, Newey-West standard errors. * p<0.05, ** p<0.01, *** p<0.001. Robust to Tealbook forecasts, the CBO unemployment gap, Poisson / quasi-Poisson models.
| (1) | (2) | (3) | (4) | (5) | |
|---|---|---|---|---|---|
| Unemployment |
−0.200*** (0.042) |
−0.200*** (0.045) |
−0.183*** (0.046) |
−0.192*** (0.043) |
−0.183*** (0.040) |
| Unemployment² |
0.060*** (0.014) |
0.060*** (0.014) |
0.059*** (0.014) |
0.058*** (0.015) |
0.055*** (0.014) |
| Nominal wage growth |
0.034 (0.038) |
||||
| Real wage growth |
0.021 (0.035) |
0.027 (0.034) |
|||
| Inflation |
0.006 (0.023) |
0.024 (0.031) |
0.037 (0.032) |
||
| Real GDP growth |
0.032 (0.024) |
||||
| Word count (log) |
0.542** (0.180) |
0.546** (0.182) |
0.545** (0.182) |
0.543** (0.185) |
0.556** (0.176) |
| Observations | 167 | 167 | 167 | 167 | 167 |
| Chair dummies | ✓ | ✓ | ✓ | ✓ | ✓ |
Wage growth: Average Hourly Earnings (AHE, BLS). Negative binomial, Newey-West standard errors. Same with median weekly earnings → appendix
Then, let’s compare public and private settings:
Blaming workers sits badly with the maximum employment mandate: a reputational cost in public (Best, 2020; Moschella, 2024)
We expect it reactions to unemployment to be stronger in private FOMC deliberations than in public speeches
\[ \begin{aligned} \log E[\text{Labor}_{ft}] = {\color{#8f8880}{\beta_0}} &+ {\color{#8f8880}{\beta_1 U_{t-1} + \beta_2 U^2_{t-1} + \beta_3 FOMC_f}} \\ &+ {\color{#9e2f27}{\beta_4 (U_{t-1} \times FOMC_f)}} \\ &+ {\color{#9e2f27}{\beta_5 (U^2_{t-1} \times FOMC_f)}} \\ &+ {\color{#8f8880}{\Omega W_{t-1} + \Gamma Z_{t-1} + \beta_6 N_{ft} + \alpha_c}} \end{aligned} \]
| (1) | (2) | (3) | (4) | (5) | |
|---|---|---|---|---|---|
| Unemployment |
−0.124** (0.048) |
−0.124* (0.049) |
−0.110* (0.052) |
−0.119* (0.049) |
−0.107* (0.047) |
| Unemployment² |
0.059** (0.019) |
0.060** (0.019) |
0.059** (0.019) |
0.058** (0.019) |
0.055** (0.018) |
| FOMC |
0.193 (0.187) |
0.195 (0.187) |
0.204 (0.183) |
0.197 (0.185) |
0.178 (0.190) |
| Unemp × FOMC |
−0.162** (0.057) |
−0.162** (0.057) |
−0.159** (0.055) |
−0.161** (0.056) |
−0.167** (0.057) |
| Unemp² × FOMC |
−0.001 (0.032) |
−0.001 (0.032) |
−0.002 (0.032) |
−0.001 (0.032) |
−0.001 (0.032) |
| Nominal wage growth |
0.032 (0.033) |
||||
| Real wage growth |
0.015 (0.031) |
0.021 (0.030) |
|||
| Inflation |
0.005 (0.021) |
0.018 (0.027) |
0.029 (0.027) |
||
| Real GDP growth |
0.032 (0.022) |
||||
| Observations | 334 | 334 | 334 | 334 | 334 |
| Chair dummies, log words | ✓ | ✓ | ✓ | ✓ | ✓ |
Negative binomial panel, Driscoll-Kraay standard errors. Robust to speaker composition (Board vs regional presidents) and to the 1993 transparency shock.
Finally, do electoral politics matter?
Controlling for economic fundamentals, is labor blaming higher under Democrats?
\[ \begin{aligned} &\log E[\text{Labor}_{ft}] = \\ &\quad {\color{#8f8880}{\beta_0}} + {\color{#8f8880}{\beta_1 U_{t-1} + \beta_2 U^2_{t-1}}} \\ &\quad + {\color{#8f8880}{\beta_3 FOMC_f + \beta_4 (U_{t-1} \times FOMC_f)}} \\ &\quad + {\color{#0b3d91}{\gamma_1 Pres_t}} + {\color{#2b6bd6}{\gamma_2 (Pres_t \times FOMC_f)}} \\ &\quad + {\color{#6a9be8}{\gamma_3 (Pres_t \times U_{t-1}) + \gamma_4 (Pres_t \times U_{t-1} \times FOMC_f)}} \\ &\quad + {\color{#8f8880}{\Omega W_{t-1} + \Gamma Z_{t-1} + \beta_6 N_{ft} + \alpha_c}} \end{aligned} \]
| (1) | (2) | (3) | |
|---|---|---|---|
| Unemployment |
−0.137** (0.049) |
−0.135** (0.045) |
−0.049 (0.098) |
| Unemployment² |
0.054** (0.021) |
0.053** (0.018) |
0.061* (0.026) |
| FOMC |
0.139 (0.186) |
−0.111 (0.207) |
−0.087 (0.216) |
| Unemp × FOMC |
−0.173** (0.057) |
−0.198*** (0.057) |
−0.311** (0.095) |
| President (Dem) |
0.407*** (0.117) |
0.160 (0.114) |
0.224 (0.165) |
| Pres × FOMC |
0.495** (0.190) |
0.424 (0.270) |
|
| Pres × Unemp |
−0.157 (0.120) |
||
| Pres × Unemp × FOMC |
0.229* (0.116) |
||
| Observations | 334 | 334 | 334 |
| Macro controls, Chair dummies, log words | ✓ | ✓ | ✓ |
Negative binomial panel, Driscoll-Kraay standard errors. Macro controls in all models: real wage growth, inflation, real GDP growth. Partisan terms jointly significant (Wald, p < 0.001 in model 2). Not explained by labor policy, Congress or the fiscal stance → appendix
The Phillips curve and the NAIRU are still there: labor blame rises as unemployment falls, and accelerates near the NAIRU, over 1978–2019
Likely deeply embedded: the fear persists without any wage growth, and labor claims do not anticipate future wage growth either → forward-looking wages
“Many good monetary policymakers […] have an almost instinctual attachment to some of those problematic concepts.” — Tarullo (2017)
But hidden from the public: strongest in FOMC deliberations, muted in speeches
And politically biased: stronger under Democratic presidents, not explained by labor laws → labor policy
Interests and prices
What’s next? - Exploring individual variation?
Fear of full employment
What’s next? - Digesting AJPS rejection and resubmission
| inflation | cpi | pce |
| price increas | price stab | overall price |
| price level | consumer price | producer price |
| price index | price pressure | price shock |
| higher price | raise price | raising price |
| rising price |
Legacy triplets
Our method
Develop a codebook
Validation sample
Test LLMs
“In recent months, turmoil in North Africa and the Middle East has reduced the global supply of oil and likely added a substantial risk premium to the price of a barrel of crude as well. What do these fast-rising commodity prices mean for inflation for the rest of the year and beyond? I believe that the inflation rate will reach a peak around the middle of this year and then edge back downward. […] First, commodity prices are not likely to keep increasing indefinitely at a rapid rate.” (Williams, speech, 4 May 2011)
Code(s): Energy, Commodities
“As I look at these projections, Mr. Chairman, I think that the risk is probably on the up side in terms of higher inflation and growth than projected. […] The only point of departure I have is on the inflation number, which I think will be higher. Some of the reasons have already been stated, so I’ll skip over those but one of the major factors is the effect of the structural deficit plus the projected decline of the dollar.” (Forrestal, FOMC, 30 January 1984)
Code(s): Fiscal, Exchange rate
“Many manufacturers are continuing to implement staff reduction plans, and I just don’t sense any fundamental improvement in the underlying employment conditions. Inflation continues to look constructive. Intensive competitive conditions seem to be an absolute lid on significant price increases. The people I talk to say the pricing environment for all products is really very tough.” (Keehn, FOMC, 30 June 1992)
Code(s): Market Power
“[…] Despite moderate decreases in PPI and CPI inflation so far in 1997, there is a heightened concern about an inflationary impulse emanating from observed price increases, a tight labor market, and a continuing easy credit market. In addition, we have seen anecdotal evidence […] with respect to the apparently shifting pattern of health care cost increases from a declining to a rising trend that will no longer tend to offset at least a portion of the wage increases.” (Melzer, FOMC, 25 March 1997)
Code(s): Labor, Other (credit, health care)
“I see job creation eventually falling to about 100,000 a month by the end of 2019. Just as I think we need to put monthly fluctuations in inflation data in perspective, I think we need to put job growth in context. In the very near future, we’re not going to need to create jobs at the intense pace we’ve been experiencing over the past few years.” (Harker, speech, 1 June 2017)
Code(s): None
We are exploring central bank communication by looking at central banker policy deliberations. We want to understand how they discuss the causes of inflation. Please read the excerpt, and classify it using the following labels:
You can give excerpts multiple labels (for example, “Market power, Fiscal”). IF the excerpt has the label “None”, it should have no other labels.
Reply ONLY with the assigned label(s) (for example, “Labor, Other”). DO NOT EXPLAIN YOUR ANSWER. This is the excerpt:
Develop a codebook
Validation sample
| (1) | (2) | (3) | |
|---|---|---|---|
| Unemployment |
−0.206*** (0.043) |
−0.206*** (0.044) |
−0.197*** (0.041) |
| Unemployment² |
0.061*** (0.014) |
0.062*** (0.014) |
0.059*** (0.013) |
| Nominal wage growth (MWE) |
−0.007 (0.010) |
||
| Real wage growth (MWE) |
−0.009 (0.009) |
−0.005 (0.010) |
|
| Inflation |
0.001 (0.026) |
0.012 (0.025) |
|
| Real GDP growth |
0.037 (0.029) |
||
| Word count (log) |
0.511** (0.185) |
0.519** (0.184) |
0.520** (0.173) |
| Observations | 162 | 162 | 162 |
| Chair dummies | ✓ | ✓ | ✓ |
With Median Weekly Earnings, wage growth is again never significant, and even slightly negative
MWE (BLS) available from 1979. Negative binomial, Newey-West standard errors.
| (1) | (2) | (3) | (4) | (5) | (6) | |
|---|---|---|---|---|---|---|
| Unemp × FOMC |
−0.190** (0.059) |
−0.194** (0.060) |
−0.193** (0.060) |
−0.195** (0.061) |
−0.318*** (0.095) |
−0.326*** (0.098) |
| President (Dem) |
0.108 (0.111) |
0.085 (0.101) |
0.205 (0.112) |
0.164 (0.105) |
0.120 (0.155) |
0.136 (0.149) |
| Pres × FOMC |
0.604*** (0.171) |
0.611*** (0.176) |
0.609*** (0.174) |
0.612*** (0.177) |
0.639** (0.224) |
0.658** (0.232) |
| Pres × Unemp |
−0.172 (0.121) |
−0.255* (0.121) |
||||
| Pres × Unemp × FOMC |
0.276* (0.114) |
0.283* (0.118) |
||||
| Quarters since min. wage hike |
0.009 (0.005) |
0.007 (0.005) |
0.009 (0.006) |
|||
| Δ Workers per WHD investigator |
0.001 (0.007) |
−0.001 (0.007) |
−0.002 (0.008) |
|||
| Observations | 320 | 320 | 320 | 320 | 320 | 320 |
| Macro controls, Chair dummies, log words | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
Labor policy proxies (federal minimum wage, Wage and Hour Division enforcement) are never significant, and the partisan terms keep their size and significance
| (t+1) | (t+2) | (t+3) | (t+4) | |
|---|---|---|---|---|
| Labor talk (log) |
0.096 (0.131) |
0.186 (0.142) |
0.198 (0.193) |
0.112 (0.199) |
| Unemployment |
−0.132** (0.041) |
−0.152** (0.048) |
−0.170** (0.064) |
−0.237** (0.089) |
| Inflation |
0.161** (0.054) |
0.172** (0.060) |
0.177 (0.110) |
0.178 (0.129) |
| Wage growth (t) |
0.585*** (0.087) |
0.540*** (0.069) |
0.477*** (0.083) |
0.359*** (0.085) |
| R² | 0.796 | 0.789 | 0.743 | 0.680 |
| Observations | 166 | 165 | 164 | 163 |
| Chair dummies, log words | ✓ | ✓ | ✓ | ✓ |
If the Fed were acting prospectively, labor claims today should predict wage growth tomorrow: they do not, at any horizon up to a year
Estimation
Sample and context
H2M Seminar | CES | Jérôme Deyris