There is an entire field dedicated to robust statistics, here is a useful practical application. Where you can estimate higher moments without relying on any lower moments.
[url=/User%20Files/681/hwcv-092.pdf]Attached File: hwcv-092.pdf[/url]
The Skeptical eye usually leads to what I call "john wayne stats".
Can you please discuss an example of a use of a robust estimator that has been a disappointment? I'm not saying that robust stats should replace classical stats, all I am saying is that practitioners should be aware of its existence.
Why Variance ?
-
kubrick
- Posts: 0
- Joined: Thu Jan 01, 2004 12:00 am
Why Variance ?
Example: in almost every case I've tried, I've gotten better OOS performance using (carefully hand-tuned, cleaned, etc) OLS regressions to robust regression methods.
Alpha male
-
kubrick
- Posts: 0
- Joined: Thu Jan 01, 2004 12:00 am
-
Rabid
- Posts: 1
- Joined: Thu Jan 01, 2004 12:00 am
Why Variance ?
In practice <3 median and interquartile range for frugality reasons, but in analysis i usually resort to mean and mean absolute deviation for slightly better accuracy given that i have time to check for unusual implications
"Most of those people have no idea what you're doing, so no idea when to get nervous, so they get nervous a lot." LMAO
-
ZmeiGorynych
- Posts: 0
- Joined: Thu Jan 01, 2004 12:00 am
Why Variance ?
Median of absolute deviation from [the median of the data] is a measure of dispersion/scale that is robust with respect to outliers.
Once I have computed that, I use it to identify the outliers, and after killing or thresholding them use standard linear-ish methods on what's left.
Of course a sceptical eye is always a prerequisite Smiley
Once I have computed that, I use it to identify the outliers, and after killing or thresholding them use standard linear-ish methods on what's left.
Of course a sceptical eye is always a prerequisite Smiley