which_extreme: Calculate a _Boolean_ vector that identifies extreme tail...

In algoquant/HighFreq: High Frequency Time Series Management

Calculate a Boolean vector that identifies extreme tail values in a single-column xts time series or vector, over a rolling look-back interval.

Description

Calculate a Boolean vector that identifies extreme tail values in a single-column xts time series or vector, over a rolling look-back interval.

Usage

which_extreme(xtsv, look_back = 51, vol_mult = 2)

Arguments

xtsv	A single-column xts time series, or a numeric or Boolean vector.
look_back	The number of data points in rolling look-back interval for estimating rolling quantile.
vol_mult	The quantile multiplier.

Details

The function which_extreme() calculates a Boolean vector, with TRUE for values that belong to the extreme tails of the distribution of values.

The function which_extreme() applies a version of the Hampel median filter to identify extreme values, but instead of using the median absolute deviation (MAD), it uses the 0.9 quantile values calculated over a rolling look-back interval.

Extreme values are defined as those that exceed the product of the multiplier times the rolling quantile. Extreme values belong to the fat tails of the recent (trailing) distribution of values, so they are present only when the trailing distribution of values has fat tails. If the trailing distribution of values is closer to normal (without fat tails), then there are no extreme values.

The quantile multiplier vol_mult controls the threshold at which values are identified as extreme. Smaller quantile multiplier values will cause more values to be identified as extreme.

Value

A Boolean vector with the same number of rows as the input time series or vector.

Examples

# Create local copy of SPY TAQ data
taq <- HighFreq::SPY_TAQ
# scrub quotes with suspect bid-ask spreads
bidask <- taq[, "Ask.Price"] - taq[, "Bid.Price"]
sus_pect <- which_extreme(bidask, look_back=51, vol_mult=3)
# Remove suspect values
taq <- taq[!sus_pect]

algoquant/HighFreq documentation built on Feb. 9, 2024, 8:15 p.m.