Political Rhetoric & Market Impact

Notebook: notebooks/B3_trump_impact.ipynb

Abstract

Political rhetoric has measurable effects on financial markets, yet quantifying this relationship requires tracking how rhetorical themes evolve over time and aligning those trajectories with economic indicators. Traditional approaches reduce political text to sentiment scores, discarding the rich multidimensional structure of political discourse.

ChronosVector (CVX) tracks Donald Trump’s rhetorical trajectory through 28,000 tweets (2015-2021) embedded with MiniLM (D=384) into a continuous semantic space. Six semantic anchors — economy, trade war, immigration, media criticism, self-praise, and threat/warning — project each tweet into interpretable rhetorical dimensions via drift(). Daily aggregation produces rhetorical trajectories that are temporally aligned with five economic indicators: S&P 500, VIX, crude oil, USD index, and 10-year Treasury yields. CVX’s changepoint detection and velocity analysis identify rhetorical regime shifts and their temporal relationship to market movements, while path signatures capture the geometric shape of tweet-storm episodes.

Presidential Communication and Markets

Bianchi et al. (2019) demonstrated that Federal Reserve communication systematically affects asset prices, establishing that textual analysis of institutional discourse has predictive value for financial markets. Bollen et al. (2011) showed that aggregate Twitter mood (measured via OpinionFinder and GPOMS) is Granger-causal to Dow Jones movements with 87.6% directional accuracy.

Specific to presidential communication, Colonescu (2018) found that Trump’s tweets about specific companies produced statistically significant abnormal returns in the 30 minutes following the tweet. Ge et al. (2020) extended this to show that tweet sentiment toward trade policy predicted next-day VIX movements.

NLP for Political Discourse

Rhetorical analysis of political text has evolved from keyword counting to contextual embeddings. Card et al. (2015) introduced framing analysis using latent dimensions, while Gentzkow et al. (2019) used word embeddings to measure political polarization over a century of Congressional records. Transformer-based embeddings (BERT, sentence-transformers) now enable fine-grained tracking of semantic shifts in political language without manual coding.

Temporal Alignment of Heterogeneous Signals

Cross-correlation and Granger causality are standard tools for assessing lead-lag relationships between time series. Dynamic Time Warping (DTW) handles non-linear temporal alignment but assumes a single alignment path. CVX’s approach — computing simultaneous trajectories in their respective spaces and analyzing geometric correspondence — preserves the full temporal structure of both signals.

CVX’s Contribution. CVX provides a complete pipeline: embed political text, project onto interpretable anchors, construct temporal trajectories, align with economic indicators, and detect co-occurring regime shifts — all within a trajectory-native framework.

Methodology

Tweet Corpus Construction

Component	Detail
Source	Trump Twitter Archive (2015-2021)
Tweets	~28,000 original tweets (excluding retweets)
Embedding model	`all-MiniLM-L6-v2` (D=384)
Temporal resolution	Daily aggregation (mean embedding per day)
Active days	~2,100 days with at least one tweet

Semantic Anchor Definition

Six anchors are defined as mean embeddings of curated seed phrases:

Anchor	Seed Phrases (examples)	Dimension Captured
Economy	”economy growing”, “jobs report”, “stock market record”	Economic boasting/commentary
Trade War	”China tariffs”, “trade deal”, “unfair trade”	Trade policy rhetoric
Immigration	”border wall”, “illegal immigration”, “caravan”	Immigration framing
Media	”fake news”, “corrupt media”, “enemy of the people”	Press antagonism
Self-Praise	”I alone”, “nobody has done more”, “tremendous success”	Self-aggrandizement
Threat	”will pay a price”, “big consequences”, “not tolerate”	Warning/escalation language

Each tweet is projected onto all 6 anchors via drift(), producing a daily 6-dimensional rhetorical trajectory.

Economic Alignment

Five economic indicators are aligned to the rhetorical trajectory:

Indicator	Source	Frequency
S&P 500 (close)	Yahoo Finance	Daily
VIX (close)	CBOE	Daily
Crude Oil (WTI)	EIA	Daily
USD Index (DXY)	ICE	Daily
10Y Treasury Yield	FRED	Daily

Alignment uses same-day and lagged cross-correlation (lags of 1, 3, 5, 10 trading days) to identify lead-lag relationships between rhetorical shifts and market movements.

CVX Pipeline

Embed: MiniLM encodes each tweet to D=384.
Ingest: Daily mean embeddings ingested into CVX graph.
Anchor Projection: drift() to 6 semantic anchors produces a 6-D rhetorical trajectory.
Changepoint Detection: detect_changepoints() on each anchor dimension identifies rhetorical regime shifts.
Velocity Analysis: velocity() captures tweet-storm intensity — periods of rapid rhetorical acceleration.
Economic Alignment: Rhetorical changepoints temporally aligned with market indicator changepoints.
Signature Analysis: path_signature(depth=2) on rhetorical trajectories during key event windows.

CVX Functions Used

CVX Function	Purpose	Parameters
`cvx.ingest()`	Load daily mean embeddings	`dim=384, metric="cosine"`
`cvx.drift()`	Anchor-relative rhetorical distance	6 semantic anchors
`cvx.detect_changepoints()`	Rhetorical regime transitions	`min_segment=14`
`cvx.velocity()`	Tweet-storm intensity	Daily resolution
`cvx.hurst_exponent()`	Rhetorical persistence	`window=90`
`cvx.path_signature()`	Event-window fingerprints	`depth=2`
`cvx.trajectory()`	Full rhetorical path	Per-anchor dimension

Key Results

Rhetorical Dimensions

The 6-anchor projection captures distinct and interpretable rhetorical dimensions across the 2015-2021 period:

Anchor	Mean Drift	Std	Trend (2015-2021)
Economy	0.42	0.11	Increasing during presidency
Trade War	0.61	0.15	Peak 2018-2019, decline after Phase 1
Immigration	0.55	0.13	Peaks around caravan events
Media	0.38	0.09	Steadily increasing
Self-Praise	0.34	0.08	Relatively stable
Threat	0.58	0.14	Episodic spikes around geopolitical events

Anchor Trajectory Trends

Trade war rhetoric shows a clear inverted-U pattern: rising through 2018, peaking during tariff escalation (mid-2019), and declining after the Phase 1 deal (January 2020).
Media criticism exhibits a persistent upward trend (Hurst=0.81), consistent with escalating press antagonism throughout the presidency.
Economy anchor diverges sharply during COVID-19 — a period where economic boasting temporarily disappears from the rhetorical trajectory.

Changepoint Alignment

Rhetorical changepoints show temporal proximity to economic regime shifts:

Rhetorical Changepoint	Anchor	Market Event (nearest)	Lag (days)
2018-03-01	Trade War	Steel tariff announcement	0
2018-07-06	Trade War	China tariff escalation	-2
2020-03-11	Economy	COVID market crash	+1
2020-11-04	Self-Praise	Election uncertainty peak	0

Current Status

The 6 rhetorical dimensions are captured and the anchor trend analysis is complete for the full 2015-2021 window. Cross-correlation with economic indicators reveals several statistically suggestive lead-lag relationships, particularly for trade-war rhetoric and VIX. Formal Granger causality testing and event-study methodology are planned extensions.

Notebook Plots

The notebook produces the following interactive visualizations:

Rhetorical Radar: 6-anchor drift over time as an animated radar chart
Anchor Trajectories: Per-anchor drift time series with changepoint markers
Economic Alignment: Dual-axis plots pairing rhetorical dimensions with market indicators
Tweet Storm Velocity: Velocity spikes overlaid on daily tweet volume
Signature PCA: 3D projection of event-window signatures colored by rhetorical regime

Running the Notebook

# Install dependencies
pip install chronos-vector sentence-transformers yfinance plotly pandas

# Run analysis
cd notebooks && jupyter notebook B3_trump_impact.ipynb

Requirements: ~4 GB RAM for embeddings, ~20 min for full embedding + CVX ingestion pipeline. Pre-computed embeddings available in data/embeddings/.