[R-SIG-Finance] What does PortfReturns return?
Brian G. Peterson
brian at braverock.com
Wed Dec 19 04:23:17 CET 2012
On 12/18/2012 07:46 PM, Robert A'gata wrote:
> I also have another question about table.Arbitrary. What does it expect as
> input return series? A series of r(t) or 1+r(t)?
Either one. The user is presumed to know how they want to chain their
returns together, as is described repeatedly in the documentation.
For high frequency data, you probably want to use simple arithmetic
returns rather than log returns, and you probably want to use arithmetic
chaining rather than continuous geometric compounding, as geometric
compounding with high frequency data will rapidly lead to nonsense +/-
Inf numbers, especially with leverage or a long/short portfolio.
> Also when I put Sharpe ratio and Calmar ratio into table.Arbitrary, I get
> an error that periodicity is too high. My return time series is trade by
> trade time series which is a high frequency returns. Does that mean I
> should roll them up to daily returns first? The problem with this is that I
> cannot backtest it with too many days (says a year) using high frequency
> data because data amount becomes too large to store in memory for me. Is
> there any way to go around these issues? Thank you.
Trying to take intraday 'returns' and automagically turn them into
longer periodicity for something like an annualized Sharpe is an
exercise in mathematical futility. If you have a whole year of returns,
you know what your Sharpe was for the year, no need to annualized the
number using silly approximations like the square root of time rule,
with its associated Gaussian assumption.
You are, as you've guessed, best off aggregating into daily, as would
your broker or investors, and using those numbers. You don't need to
have all the data in memory at once, you put your transactions into your
P&L at a lower frequency than tick (you're not the entire market), and I
would assume you would mark the portfolio at a lower frequency still.
So the total volume of data goes down rapidly. Holding a quarter or a
year even of 1-second data n memory is quite simple, and you likely
don't need or want to mark the portfolio every second if what you're
ultimately looking for is some set of comparable numbers that will make
sense to management or an investor.
Brian G. Peterson
More information about the R-SIG-Finance