optimize.portfolio.rebalancing with changing/dynamic stock universe [PortfolioAnalytics]

I see a few problems with your approach
1. Providing incorrect returns data for stocks when they are NOT in the
index won't work because you need the historical co-movements of the stock
with the rest of the investable universe. e.g. for Sharpe-Ratio
(Mean-Variance) optimization, if the stock first entered the index on July
1, you will need accurate returns data for at least the size of the window
used to compute a covariance for July 1 - e.g. if a 12-month window then
for the full year prior to the first entry into the index.
2. If your optimization problem allows negative weights (i.e. short
positions), then clearly using extreme negative returns for a stock will
identify it as a good candidate for inclusion on the short side.

Possible approach:
Major index membership changes relatively infrequently. e.g. the S&P500
membership is reconstituted annually.
If your 1/0 data shows that changes are infrequent you might be able to
simply perform several runs with each run having a "constant" investable
universe for the rebalancing dates for that run. Then stitch together the
results from the different runs.

Another approach:
Contact the authors of the package to see if they have suggestions. You are
raising a standard concern and they may have addressed it or have thought
about how to deal with it.

Regards,
Eric




On Fri, Oct 19, 2018 at 12:14 AM Simon Hovmark <simonhovmark at gmail.com>
wrote:

I am using the PortfolioAnalytics package to run a multi-period
optimization on a quite large stock universe (14 European indicies ~ 800
stocks) but I?ve run into a problem. All examples and vignettes using the
package has a fixed stock universe over the backtest period (like EDHEC)
while I?ve a dynamic universe. For example, stock XYZ might have been in

an

index from may, 2002 to October 2007 and after that excluded from the
index. However, the optimizer is only allowed to include stock XYZ in the
optimization if it was part of an index at the rebalance date.

From my understanding, it is not possible to set up a custom constraint
(like it is with objectives) to say that, stock XYZ should only be

included

in the optimization rebalancing if the rebalancing date is between may,
2002 to October 2007. Does anybody have an idea how to solve this?

Data
I?ve a dataset (daily data) with a 1 if the stocks were in their
respective index at the date and a 0 if they were not (see last

sentence).

Alternatively, I also have a list like the below with the precise dates:

Name            Entry            Exit
SBMO-NL        2003-03-03     2015-03-20
SMBO-NL        2016-03-21     2018-03-16
TA-NL           2008-03-26     2008-06-27
TLNL-NL          2000-12-29     2002-03-01
TLNL-NL          2004-03-02     2005-10-10

My returns series have the returns for the stocks for the complete

listing

period (not dependent on the stock being in an index or not). My current
(very very ugly) work-around was to multiply the 1/0 matrix and the

return

matrix, so I only had returns for the periods when the stocks where in an
index, and then replace all 0/NAs with -100. This -100 would penalize
stocks outside of an index to such a degree, that it would not be

included

in the index. However, this is ugly and incorrect plus it does not work.

Code
This is not my exact code but see this as an example of a simple
optimization:

returns <- read_excel("R", sheet = 1, col_names = TRUE)
returns <- xts(returns[,-1], order.by = as.Date(paste(returns$Dato,
format = "%Y.%m.%d")))
returns <- Return.calculate(returns, method = "log")
returns <- returns[-1,]

fund.names <- colnames(returns)
tranch1 <- portfolio.spec(assets = fund.names)

 # Constraints
tranch1 <- add.constraint(portfolio = tranch1, type = "leverage")
tranch1 <- add.constraint(portfolio = tranch1, type = "long_only")

# Objective
tranch1 <- add.objective(portfolio=tranch1, type="return", name="mean")
tranch1 <- add.objective(portfolio=tranch1, type="risk", name="StdDev")

tranch1_opt <- optimize.portfolio.rebalancing(R=returns,

portfolio=tranch1,

                                           optimize_method="DEoptim",
                                           #momentFUN = tranch1_boudt,
                                           #momentargs=ac.moments,
                                           rebalance_on = ?year?,
                                           training_period = 26,
                                           rolling_window = 26,
                                           trace=TRUE, traceDE=5,
search_size=2000)

Here is some data from my sample (dput):
structure(c(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 46.276, 46.327, 46.122, 46.924,
46.071, 46.856, 47.351, 45.645, 46.003, 47.436, 48.358, 49.501,
49.364, 49.433, 48.597, 48.204, 47.061, 47.266, 47.436, 49.211,
50.729, 51.514, 51.19, 49.996, 51.156, 51.856, 51.412, 52.385,
55.797, 55.371, 54.176, 53.784, 52.896, 53.511, 54.603, 55.115,
54.654, 52.658, 52.692, 54.603, 54.569, 56.258, 54.944, 54.654,
54.552, 54.603, 53.954, 53.852, 54.262, 54.262, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 123.067, 122.712, 123.422, 127.323, 127.855, 128.032,
124.84, 117.215, 121.826, 120.939, 117.747, 122.358, 119.875,
114.91, 115.619, 113.846, 116.328, 118.634, 119.698, 120.407,
119.166, 118.811, 120.584, 120.584, 120.584, 119.875, 119.166,
120.584, 123.422, 121.294, 119.52, 123.244, 122.89, 125.018,
126.082, 125.55, 126.082, 126.259, 125.55, 126.614, 126.082,
130.692, 132.288, 132.465, 134.061, 131.047, 131.401, 133.884,
134.771, 141.155, 38.773, 39.16, 39.419, 41.067, 40.356, 41.131,
41.099, 39.548, 40.453, 40.13, 39.419, 42.327, 42.392, 43.425,
43.684, 44.976, 44.944, 44.976, 44.136, 44.362, 43.813, 43.878,
43.942, 44.298, 44.912, 45.881, 45.235, 45.558, 47.658, 47.173,
45.59, 46.786, 47.335, 47.561, 47.27, 48.337, 47.82, 46.883,
45.396, 47.432, 47.917, 48.434, 48.66, 48.983, 48.983, 48.627,
49.015, 50.728, 51.697, 51.859), .Dim = c(50L, 5L), .Dimnames = list(
    NULL, c("BAER-CH-B4R2R5", "BALN-CH-7124594", "CFR-CH-BCRWZ1",
    "CIBN-CH-5196744", "CLN-CH-7113990")), index = structure(c(883353600,
883440000, 883526400, 883699200, 883958400, 884131200, 884217600,
884304000, 884563200, 884649600, 884736000, 884822400, 884908800,
885168000, 885254400, 885340800, 885427200, 885513600, 885772800,
885859200, 885945600, 886032000, 886118400, 886377600, 886464000,
886550400, 886636800, 886723200, 886982400, 887068800, 887155200,
887241600, 887328000, 887587200, 887673600, 887760000, 887846400,
887932800, 888192000, 888278400, 888364800, 888451200, 888537600,
888796800, 888883200, 888969600, 889056000, 889142400, 889401600,
889488000), tzone = "UTC", tclass = c("POSIXct", "POSIXt")), class =
c("xts",
"zoo"), .indexCLASS = c("POSIXct", "POSIXt"), tclass = c("POSIXct",
"POSIXt"), .indexTZ = "UTC", tzone = "UTC?)

As an alternative, I have also uploaded a full index here (
https://ufile.io/0cxn6 <https://ufile.io/0cxn6>) where the first sheet

is

the 1/0 matrix, the second the full return matrix and the third is the

1/0

matrix times the full return matrix.

        [[alternative HTML version deleted]]

        [[alternative HTML version deleted]]