Example: Rolling Correlations
Suppose we wish to add an analytic which (given a set of assets) does the following with the assets entered with the command.
- Produces a time series graph (
dygraph()) with the an average rolling correlation, as well as 25th and 75th percentiles - Produces a table (
gt()) with a correlation matrix of returns over the entire history. - Produces a table (
gt()) with a matrix of percentiles of the last rolling correlation over the entire history1.
Writing a function
Putting the above information together we can create the following function
my_corr <- function(todo,rv) {
# Get Data
tickers_to_get=strsplit(rv$assetline,";")[[1]]
roll_window <- fcoalesce(as.numeric(rv$todoargs),22) # default to 22 day rolling correlation
if(length(tickers_to_get)<3) {
avsh_quick_message("Need at least 3 tickers")
return() }
# Dont forget to add data in case it is needed
av_add_px(equitylist=tickers_to_get)
# Load the internal data store
allpx <- av_load_shinydata("pxd")[data.table(symbol=tickers_to_get),on=.(symbol)]
allpx <- allpx[,rtn:=c(NA_real_,diff(log(adjusted_close),1)), by=.(symbol)]
newdtstr <- FinanceGraphs::extenddtstr(rv$dtstr_hist,begchg=-ceiling(31/22*roll_window))
allpx <- allpx [,.(symbol,timestamp,rtn)] |> FinanceGraphs::narrowbydtstr(newdtstr)
# Make pairwise data
pairs <- CJ(var1=tickers_to_get,var2=tickers_to_get)[var1<var2,]
corDT1<- allpx[,.(timestamp, var1=symbol, rtn1=rtn)][pairs,on=.(var1),allow.cartesian=TRUE]
corDT<- allpx[,.(timestamp, var2=symbol, rtn2=rtn)][corDT1,on=.(var2,timestamp),allow.cartesian=TRUE]
# Rolling correlations
rollcor_DT <- corDT[,rcorr:=frollapply(.SD,roll_window,\(x) cor(x$rtn1,x$rtn2),by.column=FALSE), by=.(var1,var2)]
cornames <- c("corr_p25","corr_p50","corr_p75")
rollcor_toplot <- rollcor_DT[,
(cornames):=as.list(quantile(.SD$rcorr,probs=c(0.25,0.5,0.75),na.rm=TRUE)), by=.(timestamp)][
,.SD, .SDcols=c("timestamp",cornames)]
rollcorr_dyg <- fgts_dygraph(rollcor_toplot,title=paste0("Rolling percentiles of ",roll_window," bd correlations"),
roller=1,events=rv$ts_events)
# Overall correlations for period
corDT <- corDT |> FinanceGraphs::narrowbydtstr(rv$dtstr_hist)
allcorr <- corDT[,.(allcorr=cor(rtn1,rtn2,use="pairwise.complete.obs")),by=.(var1,var2)]
allcorr_gt <- dcast(allcorr,var1 ~ var2,value.var="allcorr") |> gt() |>
tab_header(title=paste0("Correlation matrix for ",rv$dtstr_hist)) |>
sub_missing(missing_text="--")
# Last Percentiles
corrpct <- rollcor_DT[,.(lastpctile=100*last(frank(rcorr,na.last=NA))/.N), by=.(var1,var2)]
pctlast_gt <- dcast(corrpct,var1 ~ var2,value.var="lastpctile") |> gt() |>
tab_header(title=paste0("Last correlation percentile ",rv$dtstr_hist)) |>
fmt_number(decimals=1) |> sub_missing(missing_text="--")
# Return list
return(list("GT3L"=allcorr_gt,"TS1"=rollcorr_dyg,"GT3R"=pctlast_gt))
}Note a few items about this code:
The app uses the user function av_load_shinydata() to get data from the internal data store. In this case, we need access to
pxdfrom the table above.The app uses two helper functions from FinanceGraphs.
extenddtstr()backs up the date string so that we can get a valid rolling correlation from the start of the analysis period (given as a string byrv$dttr_hist).narrowbydtstr()filters the data to the time periods desired.To control where the output goes, you can return a named list of outputs consistent with the above table. In this case, we return to matrices side by side by naming the outputs
"GT3L"and"GT3R". If named, the returned items don’t need to be in any order. If not named, the items are placed top to bottom.The function is
data.table()-centric, but does not need to be. Tidyverse idioms can be used, but are likely to be slower.
Registering and running the function.
We need to define how users will call this function, so a reasonable choice is “RCOR”. To add that function to the stable of those available, just run
Doing so will save the definition in the disk cache, so we just need
to rerun av_runShiny() prior to running the
RCOR function.
Once av_runShiny() is running with the registered
function, we need to ensure its definition is in the Global Environment
(.GlobalEnv()), and run it on a larger set of assets. For
example, to analyze a series of Factor ETFs over the past two years
using 2 month rolling correlations, we would type in the command line
QUAL;USMV;MTUM;VLUE;QUS;SMMV;SIZE;ESMV RCOR 44 to get `
