In an article by Bruno Dupire of Bloomberg titled "Optimal Process Approximation: Application to Delta Hedging and Technical Analysis" (July 7, 2005), he looks at price changes irrespective of time. I cannot speak for him, but in reading the article, I had the distinct impression that Mr. Dupire was unaware that he was really describing point and figure filtering. That's believable, as I doubt he would read the P&F literature, since it is considerably beneath him. I also doubt that anyone who is seriously into P&F would read Dupire. Most of the P&F crowd are unaware that P&F is a non-linear, adaptive (and in some cases asymmetric) filter.

In our shop we have done lots of work with filtering data. Universally we have found that using P&F techniques create unnecessary lag. I say unnecessary, as there are other filters that do not distort, and yet do not create lag. So we were vexed by the time we wasted in researching P&F.

But, critically, we never researched the patterns of P&F. Then Vic and Laurel mentioned it, which got us thinking we may have dismissed P&F too soon. So we have recently embarked on creating a library of frequently-recurring patterns, and using those to deselect some of the stocks on our daily lists. That is, using P&F filtering on price data.

But P&F may have other possible beneficial uses. In many ways it is the opposite of a moving median. Whereas the median discriminates against an abrupt move, P&F immediately recognizes moves beyond a certain size and ignores periods of inactivity. That makes it extremely interesting, and for more than price.

Victor Niederhoffer remarks:

The probabilities that Dr. Rafter found are completely non-random and form the base for a simple model that can be tested against extensions for reductions of variation relative to hypotheses concluded. The positive sequence of length 12 in the Dow overrides the normal difficulties with drift and heightened chance of momentum due to starting out in each case nearer the next sequence than the reversal due to randomness. 

Paolo Pezzutti asks:

Is this true also if the number of observations is quite limited for every pattern?

Victor Niederhoffer explains:

The numbers can be tested with a broad brush, with a standard variance of expected number of observations for each classification. There's enough there for the simple model to be highly significant.





Speak your mind


Resources & Links