finds.backtesting.eventstudy
Evaluate event studies
Event studies: cumulative abnormal returns
Copyright 2022, Terence Lim
MIT License
- class finds.backtesting.eventstudy.EventStudy(sql: SQL, bench: Benchmarks, stocks: Stocks, max_date: int, table: str = 'events')[source]
Bases:
StructuredClass to support statistical tests of event studies
- Parameters:
sql – connection to user database to store results
bench – Benchmark dataset of index returns for market adjustment
stocks – Stocks structured dataset containing stock returns
max_date – last date to run event study
table – physical name of table in user database to append summary results
- __call__(label: str, df: DataFrame, left: int, right: int, post: int, date_field: str) DataFrame[source]
Construct event window market-adjusted returns where valid/available
- Parameters:
label – Unique label (used for graph labels and table name)
df – Input DataFrame of stock identifiers and event dates
left – Start (inclusive) of announcement window around event date
right – End (inclusive) of announcement window around event date
post – End (inclusive) post-drift period
date_field – Name of date column in df
Notes: - populates ‘car’ attribute with cumulative arithmetic abnormal returns - populates ‘bhar’ attribute with abnormal cumulative product returns
- fit(model: str = 'scar', rows: List[int] = [], rho: float | None = None) Dict[str, Dict][source]
Compute car, or bhar, and summary statistics of selected rows
- Parameters:
model – name of predefined model to compute summary statistics
rows – Subset of rows to evaluate; empty list selects all rows
car – Whether to evaluate CAR (True) or BHAR (False)
rho – Average cross correlation of event returns. If None, then compute from max convolution of post-announcement returns
- Returns:
Dict of summary statistics of full and subsamples
’window’, ‘window-tvalue’ are CAR at end of event window
’post’, ‘post-tvalue’ are CAR from event end till post-drift end
’car’, ‘car-stderr’ are daily CAR from beginning of announcement
’rows’ is number of input rows
’days’ is number of unique dates (same announce dates are grouped)
’effective’ is the number of days after correlation effects
TODO names of models:
“car” or “bhar”
unstandardized or ‘s’tandardized or ‘adj-s’tandardized (simple rho) or ‘conv-s’tandardized (with average rho from convolution max)
Notes:
Kolari and Pynnonen (2010) eqn[3] cross-sectionally adjusted Patell or BMP with avg correlation: multiply variance by 1 + (p*(n-1))
Kolari, Pape, Pynnonen (2018) eqn[15] adjusted by average overlap (tau) and average covariance ratio (correlation rho)
- plot(model: str, drift: bool = False, ax: Any = None, loc: str = 'best', title: str = '', c: str = 'C0', vline: List[float] = [], fontsize: int = 10, hline: List[float] = [], width: float = 1.96)[source]
Plot cumulative abnormal returns, drift and confidence bands
- Parameters:
model – Name of model to plot computed results
drift – Whether to start confidence bands at post-event drift start
ax – Axes to plot
loc – Legend location
title – Main title
c – Color
vline – List of x-axis points to plot vertical line
hline – List of y-axis points to plot horizontal line
width – Number of std errs for confidence bands
fontsize – Base font size
- read(left: int, right: int, post: int) DataFrame[source]
Fetch daily cumulative abnormal returns series from user table
- Parameters:
label – name suffix of table to retrieve from
left – offset of left announcement window
right – offet of right announcement window
post – offset of right post-announcement window
- Returns:
DataFrame of car and bhar daily cumulative returns