Event Study Analysis in R

A comprehensive, modular R package for financial event study analysis. Implements the classical methodology (MacKinlay 1997) and extends it with modern multi-factor models, long-horizon methods, panel (DiD) event studies, and more.

Ecosystem

EventStudy is part of the eventstudy.de toolkit -- "Event Study Analysis Made Simple":

• Google Sheets Template -- quick, no-code event study for small samples
• R Package (this package) -- full programmatic pipeline with 15+ models, 12 test statistics, AI advisor, and publication-ready export
• WebAssembly App -- browser-based analysis with no installation required

Installation

GitHub

install.packages("devtools")
devtools::install_github("sipemu/eventstudy")

🤖 Claude Code Skills (AI pair-programming)

Two Claude Code Agent Skills ship with this repo as an installable plugin, giving your AI assistant grounded knowledge of the package API:

• es-capabilities -- discover the full API surface: return models, test statistics, task types, function signatures, and method-for-task routing.
• es-advisor -- run the grounded load -> run -> diagnose -> advise -> re-run -> compare loop (offline degrade path, never fabricates a number).

Install via the built-in plugin marketplace:

/plugin marketplace add sipemu/eventstudy
/plugin install eventstudy@eventstudy

The skills are then auto-invoked from their descriptions, or callable as /eventstudy:es-capabilities and /eventstudy:es-advisor.

Features

• AI Advisor: Grounded, LLM-agnostic advice via es_advise() plus offline es_diagnostics(). Interprets only package-computed numbers and never fabricates results. Works fully offline with no API key required; optional LLM providers (Anthropic, OpenAI-compatible, custom) can be supplied via provider().

• One-Call Reporting: Generate a complete, multi-format event study report in a single call: es_report(task) -- deep-clones the task, harvests diagnostics, and renders HTML/PDF/Word/Markdown with a grounded narrative (offline by default). Also available as run_event_study(..., report = TRUE).

• 15+ Return Models: Market Model, Market Adjusted, Mean Adjusted, Fama-French 3- and 5-factor, Carhart 4-factor, GARCH(1,1), DCC-GARCH, Rolling-Window, Buy-and-Hold Abnormal Returns (BHAR), Volume, and Volatility models.

• 12 Test Statistics: Parametric (AR T, CAR T, BHAR T, Cross-Sectional T, Patell Z, BMP, Kolari-Pynnonen) and non-parametric (Sign, Generalized Sign, Rank, Calendar-Time Portfolio).

• Diagnostics: Shapiro-Wilk normality, Durbin-Watson, Ljung-Box autocorrelation, pre-trend testing.

• Visualization: Event study plots (AR, CAR, AAR, CAAR) with confidence bands, diagnostic plots, CAR distribution histograms, panel event study plots.

• Result Export: CSV, Excel (multi-sheet), and LaTeX (publication-ready tables). Broom-compatible tidy() method for tidyverse workflows.

• Cross-Sectional Analysis: Regress CARs on firm characteristics with HC-robust standard errors. Group comparisons, quantiles, and distribution plots.

• Intraday Event Studies: POSIXct timestamp support with minute/second-level windows.

• Panel Event Studies (DiD): Two-way fixed effects (static and dynamic), Sun & Abraham (2021) interaction-weighted estimator for staggered treatment.

• Extensible Design: Add custom models (inherit ModelBase) and test statistics (inherit TestStatisticBase).

Quick Start

library(EventStudy)

# Create task from your data
task <- EventStudyTask$new(firm_tbl, index_tbl, request_tbl)

# Run with default settings (Market Model, simple returns, all test statistics)
task <- run_event_study(task)

# Inspect results
print(task)
summary(task)

AI Advisor

New in 0.60.0. Get grounded advice about your fitted study — fully offline, no API key, and it never invents numbers (it interprets only what the package computed):

diag   <- es_diagnostics(task)
advice <- es_advise(diag, task_type = "recommend_stat")
print(advice)

Runs offline out of the box. To use an LLM for prose interpretation, pass an optional provider() (Anthropic, OpenAI-compatible, or custom).

The best entry point for understanding the advisor is the "The Grounded AI Advisor" vignette (vignette("ai-advisor")), which walks through the bundled dieselgate dataset (data(dieselgate) — Volkswagen vs. the DAX around the 2015 EPA disclosure), runs the deterministic advisor layer live, and shows the LLM layer as a captured static example.

Example: Dieselgate

Analyzing the 2015 Volkswagen emissions scandal and its impact on automotive stock prices.

以上為 README 的開頭部分。其餘內容請閱讀原文。 檢視完整 README ↗