| dc.description.abstract | Genetic programming (GP) is increasingly popular as a research tool for applications in
finance and economics. One thread in this area is the use of GP to discover effective
technical trading rules. In a seminal article, Allen & Karjalainen (1999) used GP to find
rules that were profitable, but were nevertheless outperformed by the simple “buy and
hold” trading strategy. Many succeeding attempts have reported similar findings. This
represents a clear example of a significant open issue in the field of GP, namely,
generalization in GP [78]. The issue of generalisation is that GP solutions may not be
general enough, resulting in poor performance on unseen data. There are a small
handful of cases in which such work has managed to find rules that outperform buyand-
hold, but these have tended to be difficult to replicate. Among previous studies,
work by Becker & Seshadri (2003) was the most promising one, which showed
outperformance of buy-and-hold. In turn, Becker & Seshadri’s work had made several
modifications to Allen & Karjalainen’s work, including the adoption of monthly rather
than daily trading. This thesis provides a replicable account of Becker & Seshadri’s
study, and also shows how further modifications enabled fairly reliable outperformance
of buy-and-hold, including the use of a train/test/validate methodology [41] to evolve
trading rules with good properties of generalization, and the use of a dynamic form of
GP [109] to improve the performance of the algorithm in dynamic environments like
financial markets. In addition, we investigate and compare each of daily, weekly and
monthly trading; we find that outperformance of buy-and-hold can be achieved even for
daily trading, but as we move from monthly to daily trading the performance of evolved
rules becomes increasingly dependent on prevailing market conditions. This has
clarified that robust outperformance of B&H depends on, mainly, the adoption of a
relatively infrequent trading strategy (e.g. monthly), as well as a range of factors that
amount to sound engineering of the GP grammar and the validation strategy. Moreover,
v
we also add a comprehensive study of multiobjective approaches to this investigation
with assumption from that, and find that multiobjective strategies provide even more
robustness in outperforming B&H, even in the context of more frequent (e.g. weekly)
trading decisions. Last, inspired by a number of beneficial aspects of grammatical
evolution (GE) and reports on the successful performance of various kinds of its
applications, we introduce new approach for (GE) with a new suite of operators
resulting in an improvement on GE search compared with standard GE. An empirical
test of this new GE approach on various kind of test problems, including financial
trading, is provided in this thesis as well. | en_US |
