Capturing international influences in U.S. monetary policy through a Natural Language Processing approach

Replication Files for ------------------------------------- -- Capturing international influences in U.S. monetary policy through a Natural Language Processing approach -- ------------------------------------- Laurent Ferrara and Nicolas de Roux ------------------------------ Contents and Replication Instructions This repository contains the replication code and dataset for the study. The current version reflects the code and data as of 2026-08-25. This replication package includes: (i) a detailed ReadMe file with step-by-step instructions for replication; (ii) a mapping of scripts to outputs, specifying which script generates each figure and table in the paper and its appendices; (iii) all datasets and source code needed to reproduce the analysis end to end, from the FOMC text corpora through sentence classification to the Taylor-rule estimations; and (iv) a test suite that checks every regenerated table cell by cell against the numbers the manuscript prints. The pipeline runs in five stages. Because the intermediate outputs are shipped - the processed corpora, the labelled sentences, the fitted classifiers and topic model, and the attention indices - the usual entry point is the last one: "uv run run_all.py --stage E" reproduces every table and figure in about five minutes. Earlier stages can be re-run individually; the ReadMe documents what each needs and how long it takes. ------------------------------ Code Annotations and System Requirements The code is extensively commented and needs no high-performance computing. The Python environment installs from a lock file with exact versions ("uv sync"); only the Taylor-rule script additionally requires R, reached through rpy2 and documented in the ReadMe. Replication runs on a standard laptop: the final stage takes about five minutes, and re-deriving everything from the raw text several hours, dominated by transcript parsing and classifier training. Shortcuts are provided so that no step is a prerequisite for those after it - cached feature matrices (1.8 MB) rebuild the classifier comparison without loading the 430 MB pretrained embedding, and cached GPT-3.5 labels remove any need for an API key. Two inputs are not redistributed. MSCI World index levels obtained through Bloomberg cannot be shared under the licence terms; a format stub and full retrieval instructions are included, and the three affected exhibits are identified in the ReadMe. The FOMC transcript PDFs (about 922 MB) are omitted for size only - the extracted text that the code actually consumes is included, and the ReadMe gives the command to re-download the PDFs.