Forums > Risk Management > Identifying Overlapping Strategy Risk Factors with Graph Theory

 Page 1 of 1
Display using:
 Corey Total Posts: 234 Joined: Feb 2008
 Posted: 2012-08-05 03:22 So in an era of "big data," I am trying to figure out a quantitative method for identifying possible overlapping risks between strategies. As a firm, before any strategy is launched, we sit down and qualitatively discuss possible risk overlaps between strategies, but I would love some way to quantitatively figure it out.The first step, for me, feels like doing some sort of return simulation analysis (stress testing in similar environments, et cetera), using stats as simple as correlation to more complex relationship identifiers like the mutual information coefficient. But this falls short of what I am looking for. It feels like the risk will only be identified when it comes to fruition. My thought was some sort of "operation" graph. Inputs would be at the top, strategies at the bottom, and "operations" in the middle. Ideally, I could use some sort of graph traversal algorithm to identify significant overlaps in strategies. It almost seems to me like it might be possible to determine for each strategy how "influential" the inputs are by determining how they are traced into the final "output" for the strategy. Then you could determine joint points of failure by seeing if two strategies share very influential inputs.Ideally, this could extend to the "temporary" inputs constructed after operations as well ... but I think that might be a little harder. Frankly, from my computer science background, it starts to feel like an abstract-syntax tree, and trying to find and identify replica node-groups & operations between different trees.This is sort of wishy-washy, but I wanted to provide an idea of what I was thinking. Has anyone every messed around with something similar? "Then there was the man who drowned crossing a stream with an average depth of six inches." W. I. E. Gates
 Nonius Founding MemberNonius Unbound Total Posts: 11311 Joined: Mar 2004
 Posted: 2012-08-05 18:34 ive built a simulator that operates on a graph of assets, hedge funds that invest in those assets, and fund of funds that invest in those hedge funds. i monte carlo the assets and let the effects ripple through the graph, accounting for margin calls and redemptions. dont know if that is similar. Geen idee why it's a rainy day.
 Corey Total Posts: 234 Joined: Feb 2008
 Posted: 2012-08-06 00:35 Sounds very similar. I think the only difference would be that since all the funds are in-house, we can exactly model how each fund operates with any changes to the inputs. I'd be very interested in hearing more about any lessons learned and whether you thought it was a good use of your time. "Then there was the man who drowned crossing a stream with an average depth of six inches." W. I. E. Gates
 Nonius Founding MemberNonius Unbound Total Posts: 11311 Joined: Mar 2004
 Posted: 2012-08-09 23:26 unfortunately, the biggest obstacle im facing is politics/marketing of the prototype, which took a good year to build. that, and another recommendation....avoid the general graph implementations, we did that and its slow. i'll have more later, but i generally think its the right idea. Geen idee why it's a rainy day.
 athletico Total Posts: 894 Joined: Jun 2004
 Posted: 2012-08-10 19:39 Hi Corey - you're thinking of something like a bayes net or probabilistic graph model if I understand correctly?  You might be interested in Plight of the Fortune Tellers.  Rebonato's description of associative models (e.g. the usual covariance matrix or joint pdf) vs causation models (bayes nets and PGMs) agrees fairly closely with what you wrote.
 Corey Total Posts: 234 Joined: Feb 2008
 Posted: 2012-08-10 20:22 Thanks. Just ordered a copy from Amazon -- I'll check it out.Nonius -- I'd be interested in hearing how you ended up implementing if it wasn't in a graph structure. Seems, at the very least, a graphical organization structure might provide a unique visual insight into strategy risk overlaps if you color "connecting" edges uniquely for each strategy. "Then there was the man who drowned crossing a stream with an average depth of six inches." W. I. E. Gates
 sharpend Total Posts: 270 Joined: Aug 2007
 Posted: 2012-08-10 20:43 "Plight of the Fortune Tellers"? Or "Coherent Stress Testing: A Bayesian Approach..."?
 athletico Total Posts: 894 Joined: Jun 2004
 Posted: 2012-08-10 23:08 I was thinking PotFT...  hadn't seen Coherent Stress Testing until now that you mentioned it, thanks  (just placed my order)
 Nonius Founding MemberNonius Unbound Total Posts: 11311 Joined: Mar 2004
 Posted: 2012-08-11 11:15 Corey, i did in fact use a graph structure, but it is slow.every node object points to a collection of other node objects. nodes can value, manage and initialize. value first values the children then itself. it is an abstraction of computing NAV, or MtM. manage is an abstraction of managing positions, selling assets, curing margin calls, servicing investor redemptions. initialize is the obvious resetting at beginning of scenario run. all cases work on recursion on calling value, manage, intialize, etc, with recursion called before the parent for a computation that naturally should start "at the bottom" and otherwise called after if at the top, which is analogous to diffs between explicit and implicit pde methods. in fact, i sort of view the whole market/graph as this meta-derivative.by the way, i do ensure that i dont do a function call on a child if it's been called through some other path in the graph. Geen idee why it's a rainy day.
 lmog Total Posts: 78 Joined: Mar 2010
 Posted: 2012-09-09 18:23 Nonius,I remember we chatted about graph transformation before...can't remember if I mentioned GrGen to you.Very nice graph transformation package (in many senses, the only truly *general* graph transformation package). You can specify edge/node types, attributes and get a pretty rich set of transformation semantics...Not sure how your prototype works, but it might be fun to just think about how your prototype would look in GrGen...I'm using it for some graph based simulations now (non-finance).
 Nonius Founding MemberNonius Unbound Total Posts: 11311 Joined: Mar 2004
 Posted: 2012-10-22 16:55 Sorry bait belated. I ended up using Gephi but ill check that out as well. Geen idee why it's a rainy day.
 Page 1 of 1