254 D. Pelusi
main issue for the successful application of technical trading rules is to find the best
association of a “training sets" (TN-S) and of “trading sets" (TD-S), for the highest
and possibly most stable profit streams.
In this paper, we propose a synthesis of the two traditional approaches in technical
analysis, outlined above, and use the chart pattern recognition technique for the best
association of training and trading phases. Some works [7,9,22,23,29] contain studies
on the information content of chart patterns.
Our target is to investigate the existence of non-linear configurations in the hourly
observations of the Euro-Dollar (EUR-USD), Dollar-Yen (USD-JPY) and Pound-
Dollar (GBP-USD) exchange rates. Our pattern recognition algorithm takes into ac-
count ten chart patterns which are traditionally analysed in the literature [7, 23, 28].
In Section 2 we describe the algorithm. The algorithm results are shown in Section
3, whereas Section 4 contains the conclusions.
2 Pattern recognition algorithm
As outlined above, we consider hourly exchange rates. The first task in the construction
of our algorithm is the recognition that some exchange rate movements are significant
and others are not. The most significant movements of exchange rates generate a
specific pattern. Typically, in order to identify regularities and patterns in the time
series of asset prices, it is necessary to extract non-linear patterns from noisy data.
This signal extraction can be performed by the human eye, however in our algorithm
we use a suitable smoothing estimator. Therefore, to spot the technical patterns in the
best way we use the kernel regression. Hardle [16] describes this smoothing method
which permits easier analysis of the curve that describes the exchange rate.
Generally, the various chart patterns are quite difficult to quantify analytically
(see the technical analysis manuals [2, 22, 28]). However, to identify a formal way of
detecting the appearance of a technical pattern, we have chosen the definitions shown
in the paper of Lo et al. [23]. In these definitions, the technical patterns depend on
extrema, which must respect certain properties. The use of kernel regression permits
easy detection of these extrema because the curve that describes the exchange rate is
smoothed. To identify these extrema we use a suitable method described by Omrane
and Van Oppens [28].
To detect the presence of technical patterns in the best way, we use a cutoff value
as in the work of Osler and Chang [30]. In this manner, the number of maxima
and minima identified in the data is inversely related to the value of the cutoff. In
other words, an increase or decrease of it generates a different series of maxima and
minima, which will result in a different set of chart patterns. For each cutoff value, the
algorithm searches the chart patterns HS, IHS, BTOP, BBOT, TTOP, RTOP, RBOT,
DTOP, DBOT on the basis of their definitions [23]. Considering a single pattern at a
time, the algorithm counts the patterns number of that type, foreach cutoff value.
To establish a similarity parameter, we define, for each jth technical pattern
(j= 1 , 2 ,dots,10), the coefficient that represents the similarity degree between