Replication Code: Do CEO Voices Move Markets?

Olivia Yang — Princeton Senior Thesis (2026)
Advisor: Daniel Rigobon

Note: Portions of the pipeline code were debugged with assistance from Claude AI (Anthropic). Core statistical design and all empirical choices are my own.

Overview

Complete replication pipeline for "Do CEO Voices Move Markets? Vocal Stress and Analyst Tone in Earnings Calls as Predictors of High-Frequency Order Flow."

The pipeline integrates five data sources: NYSE TAQ (millisecond trade/quote data), Refinitiv StreetEvents (MP3 audio + transcripts), WRDS Compustat (ROA, B/M), CRSP (log market equity), and I/B/E/S (analyst consensus). The European ADR extension adds 23 EU firms across 11 countries using the same TAQ infrastructure.

---

Requirements

• Python 3.11
• WRDS database access (Princeton subscription)
• OpenSMILE with eGeMAPS v02 feature set
• HuggingFace transformers (ProsusAI/finbert)
• Packages: wrds, pandas, numpy, scikit-learn, statsmodels, scipy, torch, opensmile, pytz, matplotlib

See requirements.txt for full package list with version constraints.

---

Execution Order

US Pipeline

Step	Script	Purpose	Key Output
1	01_select_sample.py	Stratified CRSP sampling (40 firms, 5 volatility quintiles)	selected_sample_40_FINAL.csv
2	14_extract_call_times.py	Parse call dates/times from Refinitiv transcript headers	call_times_extracted.csv
3	09_download_taq_data.py	NYSE TAQ millisecond download (6h window, UTC→ET corrected)	taq/{TICKER}_{QTR}_taq.csv.gz
4	35_rename_2023.py	Audio/transcript standardisation to TICKER_Qn_YYYY	audio/processed/, transcripts/processed/
5	26_extract_all_audio.py	eGeMAPS 88-feature extraction via OpenSMILE	audio_features_all308.csv
6	28_run_finbert.py	FinBERT analyst tone scoring (Q&A turns only)	finbert_tone_results_all.csv
7	27_sync_all.py	Lee-Ready OI computation with DST-correct ET timestamps	full_synchronized.csv
8	29_financial_controls.py	Compustat/I/B/E/S financial controls per firm-quarter	financial_controls_all.csv
9	rebuild_master_v2.py	US master dataset assembly (310 × 156)	analysis_dataset_MASTER.parquet
10	summary_statistics.py	Table 3.4: Summary statistics (US, EU, pooled)	summary_statistics.csv
11	run_regressions.py	M1–M5, regime split, Chow test, BH scan	regression_results.csv, regime_results.csv, bh_scan_results.csv
12	run_validation.py	Placebo, permutation, bootstrap, wrong-quarter	validation_results.csv
13	robustness_tests.py	SE specifications, autocorrelation diagnostics (Table 5.9)	stdout
14	subgroup_analysis.py	Subgroup analysis (Table 5.3) + SVB experiment (Table 5.8)	subgroup_results.csv, svb_experiment.csv
15	economic_magnitude.py	Incremental R² decomposition (Fig 5.4) + economic magnitude (§5.10)	incremental_r2.csv, economic_magnitude.csv
16	run_vix_interaction.py	Continuous VIX interaction robustness (Table 5.10)	vix_interaction_results.csv, vix_interaction_table.tex
17	make_figures.py	All thesis figures (Figs 5.1–5.7)	figures/fig5_*.pdf

European ADR Extension

Step	Script	Purpose	Key Output
1	01_eu_adr_sample_v2.py	EU ADR sample selection (23 firms, 11 countries)	european_adr_sample.csv
2	14_extract_call_times_eu.py	Parse EU call dates from Refinitiv filenames and headers	call_times_eu_extracted.csv
3	09_download_taq_eu_adr_v2.py	NYSE TAQ download for EU firms (open-reaction window)	taq/eu/{TICKER}_{QTR}_taq.parquet
4	rename_eu_files_v2.py	Refinitiv→pipeline filename convention; dup resolution	audio/europe/processed/, transcripts/europe/processed/
5	26_extract_eu_audio.py	eGeMAPS extraction for EU (auto-discovers, skip-existing)	audio_features/europe/eu_audio_features_all.csv
6	28_run_finbert_eu.py	FinBERT analyst tone for EU (skip-existing guard)	finbert/finbert_tone_eu.csv
7	38_sync_eu_adr.py	Lee-Ready OI with open-reaction window for EU ADRs	synchronized/eu_adr_synchronized.csv
8	rebuild_master_global_v2.py	Global dataset assembly (US + EU)	analysis_dataset_GLOBAL.parquet
9	run_regressions_global.py	G1–G6 cross-market specifications (Tables 5.6–5.7)	regression_results_global.csv

---

Thesis Tables and Figures — Script Mapping

Thesis item	Script
Table 3.4: Summary Statistics	summary_statistics.py
Table 5.1: M1–M5 Main Results	run_regressions.py
Table 5.2: Regime Split / Chow	run_regressions.py
Table 5.3: Subgroup Analysis	subgroup_analysis.py
Table 5.4: BH Scan Top-8	run_regressions.py
Table 5.5: Validation Tests	run_validation.py
Table 5.6: Cross-Market G1–G6	run_regressions_global.py
Table 5.7: EU Regime Split	run_regressions_global.py
Table 5.8: SVB Natural Experiment	subgroup_analysis.py
Table 5.9: Autocorrelation Diagnostics	robustness_tests.py
Table 5.10: VIX Interaction	run_vix_interaction.py
Figure 5.1: M1–M5 Coefficient Plot	make_figures.py
Figure 5.2: Regime Split + Quarterly	make_figures.py
Figure 5.3: OI Distributions by Year	make_figures.py
Figure 5.4: Incremental R²	make_figures.py + economic_magnitude.py
Figure 5.5: Permutation Null Distributions	make_figures.py
Figure 5.6: Bootstrap Confidence Intervals	make_figures.py
Figure 5.7: SE Specifications	make_figures.py

---

Key Outputs

File	Dimensions	Description
analysis_dataset_MASTER.parquet	310 × 156	US sample; 281–282 complete cases
analysis_dataset_GLOBAL.parquet	490 × 257	US + EU pooled; 438 complete (reduced controls)
analysis_dataset_EU_only.parquet	196 × 257	EU subset; 164 tone+OI complete cases

---

Critical Notes

Timezone Correction (2023 TAQ data)

2023 TAQ timestamps were stored in UTC in the source data but were incorrectly applied as Eastern Time, shifting all event windows by +4 hours. 14_extract_call_times.py uses pytz for DST-aware UTC→ET conversion and stores the result in call_datetime_et. 27_sync_all.py reads this column directly, correctly handling both EDT (UTC−4, summer) and EST (UTC−5, winter). Pre-fix OI standard deviation ≈ 0.50; post-fix ≈ 0.07–0.10.

EU Open-Reaction Window

All European ADR calls occur 07:00–09:00 ET (pre-market). NYSE TAQ is illiquid pre-open, so EU OI is measured using a fixed open-reaction window: pre-window 09:00–09:30 ET, event window 09:30–10:30 ET. The is_eu indicator in pooled regressions absorbs any level difference from this window choice.

Semi-Annual Reporter Exclusions

CFR (Richemont) and RIO (Rio Tinto) are retained through data collection but excluded at the regression stage. DGE, REL, BATS, and ULVR are excluded during file renaming.

Refinitiv → NYSE Ticker Mapping

Key mappings used in rename_eu_files_v2.py: ABBN→ABB, INGA→ING, NOKIA→NOK, TTEF→TTE, PHIA→PHG, STMPA→STM, HSBA→HSBC.

VIX Data

run_vix_interaction.py requires data/vix_daily.csv, downloaded manually from FRED: Series VIXCLS — https://fred.stlouisfed.org/series/VIXCLS Save with columns: date, vix.

---

Data Sources

All data accessed via Princeton WRDS subscription:

• NYSE TAQ: taqmsec.ctm_YYYYMMDD (trades), taqmsec.cqm_YYYYMMDD (quotes)
• CRSP: crsp.msf (monthly security file)
• Compustat: comp.fundq (quarterly fundamentals)
• I/B/E/S: analyst consensus data
• Refinitiv StreetEvents: MP3 audio and matched transcripts (accessed via Princeton library subscription)

See data/README.md for the full expected directory structure.

---

HPC / SLURM

For long-running steps on Princeton Adroit, use the provided SLURM scripts in slurm/:

Script	Runs
slurm_audio_features.sh	26_extract_all_audio.py
slurm_finbert.sh	28_run_finbert.py
slurm_sync_all.sh	27_sync_all.py
slurm_financial_controls.sh	29_financial_controls.py

Submit with: sbatch slurm/slurm_audio_features.sh

---

Replication Environment

All scripts executed on Princeton's Adroit HPC cluster.
Working directory: /scratch/network/oy3009/thesis_week1/
Conda environment: thesis_env_scratch
Activate with:

cd /scratch/network/oy3009/thesis_week1/code
source /scratch/network/oy3009/thesis_env_scratch/bin/activate