Optimization vs. Stratified Sampling

[bos_IT_guy]

Schweser mentions that Optimization provides lower tracking risk than stratified sampling. However, I cannot find that mentioned in CFAI. Is it true? The only comparison I see in CFAI is that Optimization takes into account covariances of factors

 

[S2000magician]

bos_IT_guy wrote:Schweser mentions that Optimization provides lower tracking risk than stratified sampling. However, I cannot find that mentioned in CFAI. Is it true?

Probably.
Optimization is designed to minimize tracking error. Stratified sampling chops up the sample space into bins according to relevant factors, but within each bin the simple random sampling technique could produce significant skewness. If those skewnesses don’t cancel each other out, you could see (relatively) significant tracking error.

 

[Zpana]

Yes, Optimization tries to minimize Tracking risk and thats where there is more rebalancing involved (even when there are no changes in Benchmark) which will lead to more transaction cost.

 

[padniaki]

S2000magician wrote:

bos_IT_guy wrote:Schweser mentions that Optimization provides lower tracking risk than stratified sampling. However, I cannot find that mentioned in CFAI. Is it true?

Probably.
Optimization is designed to minimize tracking error. Stratified sampling chops up the sample space into bins according to relevant factors, but within each bin the simple random sampling technique could produce significant skewness. If those skewnesses don’t cancel each other out, you could see (relatively) significant tracking error.

And also, the more “bins” you chop everything into, the lower the tracking error.

 

[S2000magician]

padniaki wrote:

S2000magician wrote:

bos_IT_guy wrote:Schweser mentions that Optimization provides lower tracking risk than stratified sampling. However, I cannot find that mentioned in CFAI. Is it true?

Probably.
Optimization is designed to minimize tracking error. Stratified sampling chops up the sample space into bins according to relevant factors, but within each bin the simple random sampling technique could produce significant skewness. If those skewnesses don’t cancel each other out, you could see (relatively) significant tracking error.

And also, the more “bins” you chop everything into, the lower the tracking error.

Exactly.
And, likely, the higher the cost (because you’ll have to buy more securities, some may be illiquid, and so on).

 

[Allstatsandy]

Sounds right to me

 

[pupdawg82]

Which has lower transaction cost? optimization or ss?

 

[Allstatsandy]

i’m not sure it is cut and dried - depends on the degree each method is used.

 

[S2000magician]

pupdawg82 wrote:Which has lower transaction cost? optimization or ss?

Short run: probably optimization.
Long run: could be either.