Can a binary graph model be used to model information flow in CDS trading?
In modeling information flow, it’s natural to think in terms of directed edges. However, since information can only flow from earlier times to later times, undirected edges are sometimes sufficient. The bar graph below shows the distribution of same-day execution times at a popular Swap Data Repository (SDR).
The question we want to examine is whether the AM group influences the PM group, or just the next AM group. And similarly for the PM group: Does it influence the AM group, or just itself on the next trading day?
To represent this, the first step is to imagine the AM and PM groups over several days. For convenience, we label the nodes with even numbers for the AM and odd numbers for the PM. We make a connection if a property of the first group is repeated or intensified in the subsequent group.
For example, the volatility of the prices may increase, or the executions may be clustered more closely. Some of the properties used in constructing a “fear index” could indicate a propagation of trader anxiety from one group to another.
In the graph shown as an example, the first AM group (node 0) has influenced the PM group, and both groups have influenced the subsequent day of trading. However, this does not continue to the third day (4=AM, 5=PM). In the following two days of the example, the groups influence themselves but not each other. In the final few days, node 9 (in the PM group) influences node 10 and 11.
From this starting point there are several ways forward. Cliques can be used to study the duration of an “influence”, from which some descriptive statistics of influence clusters could be worked out. However, this can also be treated as a flow problem, where a cut could be used to separate the “influencers” from the “influencees”. The “evenness” or “oddness” of the nodes in each set would indicate a net flow of influence from one group to another.
The question is, could this information be turned to practical value? And how many nodes would have to be included in the analysis?
cit.: “… since information can only flow from earlier times to later times, …”
Not necessarily, sir.
Google had not so long ago presented a big “success” in having used TensorFlow to predict, but was caught to just do an elementary “peeking-into-future” mistake while their algorithm has based it’s success ( well, a claimed success ) on an unfair benefit from a “negative lag” which will not work in real-world … ). >>> A fabulous Google error presented as a TensorFlow “success”
On the contrary your principal question is a dual-edged sword.
Let me rather skip the second, the lethal, one ( cit.: “could this … be turned to practical value” ), the first part could be assisted by some automated hypotheses generation framework ( GUHA, et al ) and cross-validated on observable history inside a prototyping controlled-loop.
The unstated premise here is that trading is done by human beings, or by machines that have been programmed by them. The trader observes the behavior of other traders for a period of time, and then modifies his or her own behavior for a subsequent period of time. In this sense, information has flowed from the earlier time to the later time.