[R-SIG-Finance] HF strategy style change detection based on txn data

[Brian G. Peterson]

On Fri, 2012-02-17 at 16:15 +0400, Alex Bird wrote:
> Hi there!
> 
>   I have a problem described in short below but don't know where to start from.
>   If you have any ideas about possible solutions could you please
> point me to some working papers, R packages or your thoughts.
> 
>   The problem is as follows: let's suppose there's some high frequency
> trading strategy with long enough track record for which I have all
> the transactions data (trades, fees, additions/withdrawals, etc) and
> where I would like to test a null hypothesis of at least one trading
> style change during the strategy lifetime. The style change here is
> defined in a broad manner and can mean almost anything like execution
> change, trading frequency change, change of mutual dynamics between
> traded instruments, money management style change, etc.
>   From top-down point of view I would like to solve a fund of funds
> problem, i.e. hiring asset managers who come with their prêt-à-porter
> strategies but with some hidden risks which I intend to reveal.
> 
>   The only solution I see for the time being is to try to apply a
> price discovery techniques which intend to find a mapping functions of
> txn data to actual return earned and then track stability of the
> models parameters for a structural breaks.
> 
> Thanks in advance for any help!

I think you're mixing a couple of ideas up together here.  Not that that
mixing is a bad idea in and of itself, but that it will likely not help
you find resources to solve your problem.

In the first case, analyzing a strategy transaction (or return) track
record for shifts in style, there is a literature, broadly cited in the
ttrTests package, for analyzing trading decisions to separate 'luck'
from skill[1].  Pat Burns has also written pretty extensively on this
topic[2].

In the second case, actual 'style analysis' for manager returns has been
covered in the literature by everyone from Sharpe on down the road.
I'll broadly categorize these style analysis approaches as falling into
two main classes.  

The first is 'factor analysis', where you fit mutliple observed market
factors onto the manager, and monitor their factor exposures for change
over time, which likely signifies some sort of shift in investing
strategy or style.  See, for example, R-Forge package factorAnalytics[3]
or Eric Zivot's related seminar presentation from the 2011 R/Finance
conference [4] or his 2009 presetnation at [5].

The second main category is that of detecting so-called 'regime shifts',
and may be combined with factor models to give you automated warnings.
David Ardia has written extensively on this topic, see for example [6]
and [7], as have others.

Regards,

   - Brian

Ref:
[1] St. John, David ttrTests: Standard Backtests for 
	Technical Trading Rules in Financial Data
http://cran.r-project.org/web/packages/ttrTests/index.html

[2] Burns, Patrick Random portfolios for Evaluating Trading Strategies
	http://papers.ssrn.com/sol3/papers.cfm?abstract_id=881735

[3] http://r-forge.r-project.org/projects/factoranalytics/

[4] Zivot, Eric. Financial Risk Models in R
	Factor Models for Asset Returns. R/Finance 2011
http://faculty.washington.edu/ezivot/research/factorModelTutorial_handout.pdf

[5] Zivot, Eric. Using R for Hedge Fund of Funds 
	Risk Management. R/Finance 2009
http://www.rinfinance.com/RinFinance2009/presentations/PresentationChicago24_04_2009.pdf

[6] Ardia, David. Efficient Bayesian Estimation and 
	Combination of GARCH-type Models.
        http://repub.eur.nl/res/pub/19380/401D94D6d01.pdf

[7] Ardia, David. Financial Risk Estimation with Bayesian Estimation of
GARCH models.
http://www.scribd.com/doc/46504267/Financial-Risk-Management-With-Bayesian-Estimation-of-GARCH-Models

-- 
Brian G. Peterson
http://braverock.com/brian/
Ph: 773-459-4973
IM: bgpbraverock