AbnormalReturns Documentation

This package provides functionality for getting slices of firm and market return data and running firm-specific regressions commonly used to calculate abnormal stock returns (actual stock return minus a benchmark). These are common in event studies in finance and economics and often require running a large number of regressions on different slices of firm and market data.

Most of the documentation is currently in the example.

Motivation

When estimating abnormal returns, it is common to estimate how the firm's return typically responds during an estimation window and use those predicted results in an event window:

The exact length of the estimation and event windows varies, but are typically about 150 and 3-5, respectively. The estimation is typically is a linear regression of firm specific return on market-wide factors.

The Problem

Estimating abnormal returns requires getting two separate slices of data (for the estimation window and event window) for each firm-event. This is relatively trivial for small datasets, but abnormal returns are often calculated for a large number of events. For example, there are over 600,000 firm earnings announcements since 1990.

Generally, creating the dataset is done through a range join (e.g., gather all firm data between the start and end of the estimation window), which is often time consuming and/or creates huge datasets.

This Package

This package uses a custom data structure to avoid repeating the data. The data structure is built on BusinessDays.jl, making it easy to get a slice of data between two dates. It also implements threaded solutions to make the regressions and aggregation as fast as possible.

In a benchmark on 1 million firm events, it runs all the regressions in under 3 seconds. In a larger benchmark with two different models (so 2 million regressions) and calculating abnormal returns for the events, along with other basic statistics, it takes less than 9 seconds on a Ryzen 5 3600.

Acknowledgements

This package would not be possible without BusinessDays.jl, which is used for all of the date operations in this package and StatsModels.jl, which provides an incredible @formula macro and the functionality that comes with that.