04
From release architecture to public understanding.
Worcester College
Communication as Text Data.
Measurement, public attention, strategic disclosure, climate translation, and understandability.
From Release to Reception
- Measurement: make the missing middle observable.
- Architecture: ask how bundling changes attention.
- Unexpected content: isolate the first public reading.
- Public engagement: distinguish posting from exposure.
Communication becomes data only after the economic object, source, and empirical unit are fixed.
day4_00_opening.tex · line 7 · D4-F002
01
Central Banking · Lecture 4 · Worcester College
Part 01
Measure communication.
Text and digital traces reveal the missing middle
Measure the objectIdentify the mechanismRead the evidence
Standard Data Miss the Middle
\[\underbrace{i_t,\ \pi_t,\ y_t,\ r_t,\ q_t}
_{\text{standard macro-finance data}}
\qquad\Longrightarrow\qquad
\underbrace{\text{beliefs, expectations, prices}}
_{\text{final outcomes}}.\]
The empirical gap lies between the released decision and the observed outcome.
day4_01_measurement.tex · line 10 · D4-F004
Mechanism · read left to right
Transmission Runs Through People
\[\text{communication}
\longrightarrow
\text{attention}
\longrightarrow
\text{interpretation}
\longrightarrow
\text{belief updating}.\]
- A message cannot update a belief if it is not noticed.
- Exposure does not guarantee a shared interpretation.
- Text and digital traces make parts of this information environment observable.
Communication data measure the mechanism that endpoint data leave hidden.
day4_01_measurement.tex · line 30 · D4-F004
Conceptual illustration · not empirical evidence
Attention and Interpretation Fill the Gap
Teaching point
Reach is not impact: attention and interpretation sit between publication and outcomes.
Use the next slides for the documentary evidence.
day4_01_measurement.tex · line 54 · D4-F004 — pedagogical visual, no new empirical claim
Communication Expands the Information Set
\[\mathcal I_t
=
\mathcal I_t^{macro}
\cup\mathcal I_t^{market}
\cup\mathcal I_t^{communication}
\cup\mathcal I_t^{public}.\]
\[\mathcal I_t^{macro},\qquad \mathcal I_t^{market}.\]
\[\mathcal I_t^{communication},\qquad \mathcal I_t^{public}.\]
A policy message matters only after it enters an audience's information set.
day4_01_measurement.tex · line 65 · D4-F005
Economic object · define before estimating
Text Extends Policy Measurement
| Object | Standard data | Text or digital data |
|---|---|---|
| Policy decision | yes | yes |
| Public attention | rarely | yes |
| Public interpretation | rarely | yes |
| Disclosure choices | rarely | yes |
| Institutional admissibility | rarely | yes |
| Asset-price response | yes | yes |
The added value is access to reception, selection, and admissibility—not another policy-rate series.
day4_01_measurement.tex · line 90 · D4-F006
Match the Corpus to the Behaviour
| Source type | Examples | Economic object |
|---|---|---|
| Policy and official text | statements, minutes, speeches, transcripts, fiscal reports | priorities, guidance, disclosure |
| Firm disclosures | earnings calls, annual reports, filings, presentations | expectations, risk, investment plans |
| News and media | newspapers, newswires, analyst reports, blogs | information flow, sentiment, uncertainty |
| Social and public text | Twitter/X, Reddit, survey comments, reviews | interpretation, disagreement, beliefs |
| Attention traces | Google Trends, Wikipedia views, web traffic | search and reference behaviour |
Availability is not validity: source choice follows the economic object.
day4_01_measurement.tex · line 111 · D4-F007
Economic object · define before estimating
Rules Turn Sources into Variables
\[\text{raw source }s_i
\xrightarrow{\ \widehat\phi\ }
\widehat x_i
\xrightarrow{\ \text{econometric design}\ }
Y_i.\]
Unstructured material becomes evidence only through an explicit measurement rule.
day4_01_measurement.tex · line 136 · D4-F008
Interpretive boundary
Sources Reveal Different Stages
| Source | Measured object |
|---|---|
| Search or page views | attention and information acquisition |
| Official posts | engagement with public communication |
| Public text | interpretation, beliefs, disagreement |
| Institutional text | disclosure, omission, policy priorities |
No single source measures attention, interpretation, and disclosure at once.
day4_01_measurement.tex · line 158 · D4-F008
Economic object · define before estimating
One Document, Many Measurements
| Measurement | Variable | Question |
|---|---|---|
| Tone | hawkish/dovish score | how restrictive is the communication? |
| Topic | inflation, labour, risk | what is the discussion about? |
| Surprise | residualised tone | what was not predictable before release? |
| Coverage | statement–transcript gap | what was omitted from the focal release? |
| Linkage | mandate category | how does an issue become policy-relevant? |
The same minutes become a different variable when the research question changes.
day4_01_measurement.tex · line 177 · D4-F009
Economic object · define before estimating
The document is not the variable
\[\text{document}\neq\text{variable}.\]
Start with the quantity to be measured; only then select the text method.
day4_01_measurement.tex · line 197 · D4-F009
Economic object · define before estimating
Auditable Measurement Has a Workflow
\[\begin{aligned}
\text{source}
&\longrightarrow\text{cleaned corpus}
\longrightarrow\text{measurement rule}\\
&\longrightarrow\text{structured variable}
\longrightarrow\text{econometric design}.
\end{aligned}\]
Measurement is not preprocessing; it is an explicit empirical stage.
day4_01_measurement.tex · line 213 · D4-F010
Interpretive boundary
Each Stage Has Its Own Failure
| Stage | What happens |
|---|---|
| Source | collect posts, speeches, minutes, transcripts, searches, or views |
| Cleaning | remove retweets, bots, false matches, duplicates, and missing metadata |
| Measurement | classify tone, topic, interpretation, linkage, or omitted content |
| Aggregation | move to event, meeting, institution-year, or release-window units |
| Validation | compare human labels, test placebos, and propagate uncertainty |
A credible variable records how raw material changed at every stage.
day4_01_measurement.tex · line 233 · D4-F010
Three questions organise the empirical cases
\[\text{communication}
\longrightarrow
\text{measured variable}
\longrightarrow
\text{economic mechanism}.\]
| Research question | Observed source | Economic object |
|---|---|---|
| Who pays attention? | searches, page views, public and official posts | attention and engagement |
| How is the message interpreted? | public text around announcements | direction, disagreement, narrative |
| What becomes part of the record? | statements, minutes, transcripts, speeches | disclosure and admissibility |
The lecture moves from reception to interpretation and then to institutional selection.
day4_01_measurement.tex · line 253 · D4-F011
02
Central Banking · Lecture 4 · Worcester College
Part 02
Architecture shapes attention.
One richer event can still create fewer opportunities to notice
Measure the objectIdentify the mechanismRead the evidence
Institutional change · August 2015
Warsh Replaced Sequence with Bundle
Before
Several release dates
Decision → forecast report + press conference → minutes + vote
Several smaller attention waves
After
One full-release bundle
Decision + minutes + vote + forecast + press conference
One larger focal occasion
The reform changed release architecture, not the existence of policy information.
day4_02_architecture.tex · line 10 · D4-F013
Economic mechanism · two margins move at once
Bundling Moves Two Attention Margins
Intensive margin
A richer event
More content arrives together, raising salience on the bundled day.
Attention per event ↑
Extensive margin
Fewer occasions
One forecast-round opportunity disappears after bundling.
Attention occasions ↓
A larger event can coexist with less attention over the whole forecast round.
day4_02_architecture.tex · line 38 · D4-F014
Conceptual illustration · not empirical evidence
A Bigger Release Can Mean Less Attention
Teaching point
Bundling raises attention per event but removes an attention occasion.
Use the next slides for the documentary evidence.
day4_02_architecture.tex · line 63 · D4-F014 — pedagogical visual, no new empirical claim
Interactive question · which attention margin dominates?
Net Attention: Gain Minus Lost Occasion
\[\Delta A_q=\text{salience gain}-\text{lost occasion}\]
Net attention · hypothetical units
+25
The actual sign must come from the event and forecast-round evidence.
Communication architecture determines both the size and number of attention triggers.
day4_02_architecture.tex · line 75 · D4-F014
Interpretive boundary
Three Traces, Three Attention Behaviours
| Trace | Series | Behaviour measured |
|---|---|---|
| Twitter/X | log unique users posting original BoE/MPC tweets | real-time public discussion |
| Google Trends | UK index for BoE, interest rates, and inflation | active information acquisition |
| Wikipedia | log views for BoE and monetary-policy pages | reference and background checking |
The outcome is revealed attention: discussion, search, and reference behaviour.
day4_02_architecture.tex · line 96 · D4-F015
Abnormal Attention Uses a Baseline
\[\underbrace{A_{e,p}}_{\text{abnormal attention}}
=
\underbrace{Y_{t(e),p}}_{\text{event-day activity}}
-
\underbrace{\bar Y^{\,base}_{e,p}}_{\text{matched baseline}}.\]
The design explains an event spike, not raw platform volume.
day4_02_architecture.tex · line 122 · D4-F016
Match Weekdays; Exclude Nearby Events
| Object | Construction |
|---|---|
| Event activity \(Y_{t(e),p}\) | log users, search index, or log page views on the event date |
| Baseline \(\bar Y^{base}_{e,p}\) | same-platform mean on matching non-event weekdays in the prior 30 days |
| Exclusion | remove days near other MPC events from the baseline |
| Outcome | event-specific abnormal attention \(A_{e,p}\) |
Matching the attention environment separates an announcement spike from ordinary weekly rhythms.
day4_02_architecture.tex · line 144 · D4-F016
Interpretive boundary
Event and Round Analyses Differ
| Unit | Pre-reform | Post-reform |
|---|---|---|
| Event | decision day or Inflation Report day | full-release day or standard meeting |
| Forecast round | decision day plus Inflation Report day | one full-release day |
- Event level: is the bundled day more salient?
- Forecast round: does one bundled day replace attention from two occasions?
The unit of analysis determines whether the intensive or extensive margin is visible.
day4_02_architecture.tex · line 166 · D4-F017
Quarterly Attention Sums Event Spikes
\[CumLogAtt_q
=
\sum_{e\in\mathcal E_q}A_{e,TW}.\]
Digital traces let the researcher count attention by event before aggregating it.
day4_02_architecture.tex · line 190 · D4-F017
Documentary evidence · source-preserving figure
Twitter Peaks on Decision Day
Across \(h=-3,\ldots,+3\), event-day log unique users spike and attention decays sharply after the decision.
day4_02_architecture.tex · line 207 · D4-F018
Documentary evidence · source-preserving figure
Google Search Peaks on Decision Day
The event-day distribution shifts upward relative to adjacent days, consistent with active information acquisition around MPC announcements.
day4_02_architecture.tex · line 219 · D4-F019
Documentary evidence · source-preserving figure
Wikipedia Views Peak on Decision Day
Reference and background-checking activity rises on the MPC event day and recedes afterward.
day4_02_architecture.tex · line 231 · D4-F020
Documentary evidence · source-preserving figure
Quarterly Twitter Attention Shifts Left
Pre: \(N=18\), mean \(2.966\), s.d. \(0.631\). Post: \(N=25\), mean \(1.742\), s.d. \(0.466\); the post mean is below the pre minimum \(1.945\).
day4_02_architecture.tex · line 243 · D4-F021