Speaker
Description
inancial forecasting is a widely researched problem domain which is known to
be quite challenging and exhibits characteristics of uncertainty. Genetic Pro-
gramming (GP) [1] has been shown to be an effective tool for financial forecast-
ing [2]. However, the manual design of GP classification algorithms for financial
forecasting still remains the popular approach despite it being shown to be an
error-prone, time-consuming task influenced by human bias. The volatility of
financial forecasting problems require a tool that can respond to changes in a
timely manner while also maintaining acceptable prediction rates. In this re-
search, we propose the use of grammatical evolution (GE)[3] to configure GP
classification algorithms for financial forecasting. Grammatical evolution and
genetic programming are population-based algorithms and part of their function-
ality is evaluating the fitness of each individual at each generation. Inevitably
this leads to high run-times furthermore, the algorithms are stochastic in na-
ture which means a number of runs have to be performed in order to obtain a
normal distribution of results. To evaluate the proposed grammatical evolution
systems, we made use of the CHPC distributed architecture. Fifteen stocks were
selected from the NASDAQ, NYSE, XETRA and HKSE stock exchanges. A
varied selection was made because different industry stock have a varied volatil-
ity, for example stock from the technology sector is more volatile than stock
from the banking sector. Each dataset comprises of data from 1500 trading days
03/01/2012 to 05/03/2018 (1000 training and 500 test).The GE algorithms are
distributed over the cores and multiple runs performed on multiple cores. Certain
operations of the algorithm that can be performed independently are distributed
over the cores, such as population generation or fitness evaluation. For example,
the GE system consists of a GE algorithm and GP algorithm running simulta-
neously. The population generation of the GE is distributed over a fixed number
of cores. Each member of the GE population then initialises a GP algorithm
which is also distributed over a specified number of cores, for a fixed number
of runs. The GE system also has to perform a number of runs and these are
performed on the cores. The system is coded in the Java programming language
mainly using the threads API. Each individual of the GE is represented as a
core therefore as a thread. The large queue on the Lengau Cluster is used, using typical values of cores ranging from 600 to 1800.
