Inf in Variance value about onlinestats.jl HOT 7 CLOSED

Actually is there a bug here? https://github.com/joshday/OnlineStats.jl/blob/master/src/summary.jl#L65

from onlinestats.jl.

It works ok for EqualWeight but ExponentialWeight is completely wrong.

from onlinestats.jl.

Here's a simple example:

julia> x = randn(100);

julia> var(x)
1.162888972180564

julia> v = Variance()
▌ Variance 
  ▶     value: -0.0
  ▶      nobs: 0

julia> fit!(v, x); v
▌ Variance 
  ▶     value: 1.1628889721805642
  ▶      nobs: 100


julia> v = Variance(ExponentialWeight(10))
▌ Variance 
  ▶     value: -0.0
  ▶      nobs: 0


julia> fit!(v, x); v
▌ Variance 
  ▶     value: -1.976802951354775e189
  ▶      nobs: 100

from onlinestats.jl.

That is quite the bug! I'll take a look.

from onlinestats.jl.

var(o::Variance) now behaves a bit better:

julia> xvar = Float64[(fit!(v, xi); std(v)) for xi in x]
100-element Array{Float64,1}:
 0.0
 0.03677
 0.265324
 0.537096
...

In my rewrite, I apparently missed a method for ExponentialWeight(minstep::Integer), so your example was using a nonsensical weight of 10 for each update. This has been fixed.

julia> x = randn(100);

julia> v = Variance(ExponentialWeight(20));

julia> fit!(v, x); v
▌ Variance
  ▶     value: 0.9294099372585689
  ▶      nobs: 100

The line you pointed out comes straight from Knuth's Art of Computer Programming, but I'm not sure it's correct for non-EqualWeight updates. I'll leave this open while I look into it.

from onlinestats.jl.

http://www-uxsup.csx.cam.ac.uk/~fanf2/hermes/doc/antiforgery/stats.pdf confirms the update is the same for ExponentialWeight (see equation 142, equation 143 has a typo).

from onlinestats.jl.

Great thanks for following up!

On Sat, Feb 13, 2016 at 3:42 PM, Josh Day [email protected] wrote:

Closed #51 #51.

—
Reply to this email directly or view it on GitHub
#51 (comment).

from onlinestats.jl.