In this work, we present a new and efficient algorithm to perform a short-term market trend forecast, based on the Artificial Organic Networks (AON) metaheuristic machine learning framework. Regarding this goal, we present the concept of Artificial Halocarbon Compounds (AHC) or AHC-algorithm as a bio-inspired supervised machine learning algorithm based on the AON framework. Through our research, we contrast the forecast acquired with the proposed AHC model, to previously reported outcomes using the Artificial Hydrocarbon Networks (AHN) in similar tasks. The AHN algorithm is the first formally defined topology based on the AON, making the AHN algorithm a vital benchmark to contemplate. After comparing the AHC-algorithm to the original AHN-algorithm, we found out that due to the high computational complexity of the latter, the new topology is more convenient when modeling more complex systems; being this characteristic the main contribution of the AHC-algorithm, allowing it to be a more adaptable, dynamic, and reconfigurable topology. Likewise, we compared the results of the AHC-algorithm against the outcomes derived from an ARIMA model; we also made a cross-reference contrast against results concerning the prediction of other stock market indices using former state-of-the-art machine learning methods. The proficiency of the AHC-algorithm is assessed by doing a forecast of the IPC Mexico index obtaining good results, achieving a computed R-square of 0.9919, and an $$8\times 10^{-4}$$ 8 × 10 - 4 mean relative error for the forecast.

The use of algorithms in finance and trading has become an increasingly thriving research area, with researchers creating automated and pre programmed trading instructions utilising indicators from technical and sentiment analysis. The indicators of the two analyses have been used mostly individually, despite evidence that their combination can be profitable and financially advantageous. In this paper, we examine the advantages of combining indicators from both technical and sentiment analysis through a novel genetic programming algorithm, named STGP-SATA. Our algorithm introduces technical and sentiment analysis types, through a strongly-typed architecture, whereby the associated tree contains one branch with only technical indicators and another branch with only sentiment analysis indicators. This approach allows for better exploration and exploitation of the search space of the indicators. To evaluate the performance of STGP-SATA we compare it with three other GP variants on three financial metrics, namely Sharpe ratio, rate of return and risk. We furthermore compare STGP-SATA against two financial and four algorithmic benchmarks, namely, multilayer perceptron, support vector machine, extreme gradient boosting, and long short term memory network. Our study shows that the combination of technical and sentiment analysis indicators through STGP-SATA improves the financial performance of the trading strategies and statistically and significantly outperforms the other benchmarks across the three financial metrics.

We present a novel genetic programming (GP) algorithm that combines physical time and event-based time indicators to trade on the stock market. Rather than only using data in fixed intervals (e.g. daily closing prices), we use directional changes to transform physical time into events and allow the GP to make trading decisions based on when significant price movements have occurred. We use a two-objective fitness function, which simultaneously optimizes return and risk. To overcome challenges with the convergence ability of the multi-objective GP, we apply an a-dominance strategy, which is able to relax the strict Pareto dominance criteria. We run experiments on 110 stocks from 10 international markets and compare results against a single-objective GP, as well as strategies based on technical analysis indicators and buy-and-hold. Results show that the proposed GP algorithm offers statistically significant improvements when compared to the above benchmarks.

Deep reinforcement learning (DRL) offers the potential to make intelligent trading decisions in high frequency trading strategies in the foreign exchange (FX) market at a fraction of the time it takes humans. In this work, we use an event-based time sampling method referred to as directional changes (DC), which samples data only when there is a significant change in price, to build a DRL-based trading system. The enhanced price representation through DC sampling, combined with positional features reflecting the agent's trading account, provides the DRL agent with information about its exposure to market changes. With this representation of the environment we can train a DRL agent to profitably trade the FX market at high frequencies. Tick data from fourteen FX currency pairs is sampled using the DC framework and then split into windows to form 784 datasets. The novel trading system called PADRL, uses Proximal Policy Optimisation (PPO) to train agents that can autonomously generate considerable levels of profit without rule-based interjections. The resultant agents are compared to four different benchmarks including Buy and Hold, an existing successful DC-based DRL strategy (FDRL) and two popular technical analysis based strategies (MACD and RSI). Strategy performance is measured across three different performance metrics (namely Total Return, Maximum Drawdown and Calmar Ratio), with the novel PADRL system significantly outperforming them all for Total Return and Calmar Ratio.

Real estate is a favored investment option as it allows investors to diversify their portfolios and minimize risk. Investors can invest in real estate directly by purchasing a property, or through real estate investment funds (REITs) where they can purchase shares in companies that own and manage real estate. Investing in REITs has become increasingly popular because it eliminates some of the disadvantages associated with direct real estate investment, such as the need for a large upfront payment. When investing in mixed asset portfolios, it is crucial to predict future prices accurately to ensure profitable and less risky asset allocation. However, literature on price prediction often focuses on only one or two algorithms, and there is no research that explores REITs’ price prediction in the context of portfolio optimization. To address this gap, we conducted a thorough evaluation of 5 machine learning algorithms (ML), including Ordinary Least Squares Linear Regression (LR), Support Vector Regression (SVR), k-Nearest Neighbours Regression (KNN), Extreme Gradient Boosting (XGBoost), and Long/Short-Term Memory Neural Networks (LSTM), as well as other financial benchmarks like Holt’s Exponential Smoothing (HES), Trigonometric Seasonality, Box–Cox Transformation, ARMA Errors, Trend, and Seasonal Components (TBATS), and Auto-Regression Integrated Moving Average (ARIMA). We applied these algorithms to predict future prices for 30 REITs from the US, UK, and Australia, as well as 30 stocks and 30 bonds. The assets were then used as part of a portfolio, which we optimized using a genetic algorithm. Our results showed that using ML algorithms for price prediction provided at least three times the return over benchmark models and reduced risk by almost two-fold. For REITs, we observed that the use of ML algorithms led to a higher allocation to REITs diversified by country. In particular, our results showed that SVR was the best-performing algorithm in terms of risk-adjusted returns across different time horizons, as confirmed by our Friedman test results (Sharpe ratio). Overall, our study highlights the effectiveness of ML algorithms in predicting asset prices and optimizing portfolio allocation.

Directional changes (DC) have been shown to form an effective approach in algorithmic trading by converting fixed time series into event-based series and focusing on key events. Previous work has focused on forecasting the inflection point in the market and proposing new indicators under the DC framework, with just a handful of papers concerned with training and using DC indicators through machine learning. Earlier research has shown that genetic programming (GP) combining DC and physical time indicators could achieve positive returns with low risk. However, the fitness function used in that work is simply a risk-adjusted return. In this paper, we investigate whether a multi-objective optimisation approach could improve the performance of GP-based strategies in the market. We evaluate the cumulative return, risk, and rate of return of the proposed approach under 110 datasets from 10 different markets. Furthermore, we compare the proposed strategy against GP-based single objective optimisation (SOO) and buy-and-hold strategies. Our results show that the proposed approach significantly improves the cumulative return compared to SOO, from 14.29% to 62.04%, while also outperforming the buy-and-hold strategy.

This paper explores the use of the Directional Changes (DC) paradigm for financial forecasting. DC is an event-based alternative to the traditional approach of time-series with fixed intervals. In the DC approach, price movements are recorded when specific events occur, rather than in fixed time intervals, while significant price changes are identified using a threshold. Here, we consider a more general model that allows multiple weighted thresholds, and propose three novel trading strategies built within the DC paradigm. To optimize the weights of the thresholds, we use a genetic algorithm and manage to find strategies that outperform previously known single-threshold strategies under the common efficiency metrics. Furthermore, our method manages to create profitable trading strategies that outperform some traditional ones, such as buy-and-hold, MACD, and RSI.

This paper proposes a novel strongly typed Genetic Programming (STGP) algorithm that combines Technical (TA) and Sentiment analysis (SA) indicators to produce trading strategies. While TA and SA have been successful when used individually, their combination has not been considered extensively. Our proposed STGP algorithm has a novel fitness function, which rewards not only a tree's trading performance, but also the trading performance of its TA and SA subtrees. To achieve this, the fitness function is equal to the sum of three components: the fitness function for the complete tree, the fitness function of the TA subtree, and the fitness function of the SA subtree. In doing so, we ensure that the evolved trees contain profitable trading strategies that take full advantage of both technical and sentiment analysis. We run experiments on 35 international stocks and compare the STGP's performance to four other GP algorithms, as well as multilayer perceptron, support vector machines, and buy and hold. Results show that the proposed GP algorithm statistically and significantly outperforms all benchmarks and it improves the financial performance of the trading strategies produced by other GP algorithms by up to a factor of two for the median rate of return.

One of the most popular ways to reduce the risk of an investment portfolio is by holding shares of Real Estate Investment Trusts (REITs), which own and manage real estate. An important aspect of this process is to be able to forecast future REITs prices, as this allows investors to achieve higher returns at lower risk. This paper examines the performance of five different machine learning algorithms in the task of REITs price forecasting: Ordinary Least Squares Linear Regression, Support Vector Regression, k-Nearest Neighbours Regression, Extreme Gradient Boosting, and Long/Short-Term Memory Neural Networks. In addition to past REITs prices, we also use Technical Analysis indicators to assist the algorithms in the task of price prediction. While such indicators are very popular in stocks forecasting, they have never been used to forecast REITs. Our experiments show that (i) all ML algorithms produce low error and standard deviation, and are able to outperform the well-known statistical benchmark of AutoRegressive Integrated Moving Average (ARIMA), and (ii) the introduction of Technical Analysis (TA) indicators into the feature set leads to an error reduction of up to 50%.

Algorithmic trading has become an increasingly thriving research area and a lot of focus has been given on indicators from technical and sentiment analysis. In this paper, we examine the advantages of combining features from both analyses. To do this, we use two different genetic programming algorithms (GP). The first algorithm allows trees to contain technical and/or sentiment analysis indicators without any con-straints. The second algorithm introduces technical and sentiment analysis types through a strongly typed GP, whereby one branch of a given tree contains only technical analysis indicators and another branch of the same tree contains only sentiment analysis features. This allows for better exploration and exploitation of the search space of the indicators. We perform experiments on 10 international stocks and compare the above two GPs' performances. Our goal is to demonstrate that the combination of the indicators leads to improved financial performance. Our results show that the strongly typed GP is able to rank first in terms of Sharpe ratio and statistically outperform all other algorithms in terms of rate of return.

The main purpose of portfolio optimization is to reduce the risk, and/or maximize the return of a group of investments. Most of the works that have been done on port-folio optimization are based on the Modern Portfolio Theory introduced by Markowitz in 1959. Some of them have employed price predictions to compute optimal asset weights. It has been demonstrated that using price predictions, instead of historical data, might improve portfolio performance under a risk-adjusted perspective. However, contributions in the field mainly focused on stocks, while little attention has been given on multi-asset portfolios including real estate. In this paper, we fill this gap by running a genetic algorithm on 456 portfolios to demonstrate the added value of including price predictions in our asset allocation problem. To investigate this, we compare the theoretical case of having a perfect foresight, where the predicted price$p_{t}$is exactly the same as the expected price pt; under this case, the portfolio optimization task takes place in the test set (since we have assumed a perfect price prediction). We compare the results under perfect foresight with results derived from portfolio optimization that only took place in the training set, and the weights were then directly applied to the test set. Our goal is to demonstrate the theoretical advantages of using price predictions on mixed-asset portfolios that include real estate. Our results show that there can be significant improvements (up to 45 %) in sharpe ratio, rate of return, and risk, when using price predictions instead of a historical prices based portfolio.

Machine learning (ML) techniques have shown to be useful in the field of financial forecasting. In particular, genetic programming has been a popular ML algorithm with proven success in improving financial forecasting. Meanwhile, the performance of such ML algorithms depends on a number of factors including data analysis from different markets, data periods, forecasting days ahead, and the transaction cost which have been neglected in most previous studies. Therefore, the focus of this paper is on investigating the effect of such factors. We perform an extensive evaluation of a financial genetic programming-based approach and compare its performance against 9 popular machine learning algorithms and the buy and hold trading strategy. Experiments take place over daily data from 220 datasets from 10 international markets. Results show that genetic programming not only provides profitable results but also outperforms the 9 machine learning algorithms in terms of risk and Sharpe ratio.

The subject of financial forecasting has been re-searched for decades, and the driver behind its measured data has been fuelled by the selection of physical time series, which summarize data using fixed time intervals. For instance, time-series for daily stock data would be profiled at 252 points in one year. However, this episodic style neglects the important events, or price changes that occur between two intervals. Thus, we use Directional Changes (DC) as an event-based series, which is an alternative way to record price movements. In DC, unlike time-series methods, time intervals are constituted by price changes. The unique feature that decides the price change to be considered as a significant is called a threshold θ. The objective of our paper is to create DC-based trading strategies, and then optimize them using a Genetic Algorithm (GA). To construct such strategies, we use DC-based indicators and scaling laws that have been empirically identified under DC summaries. We first propose four novel DC-based trading strategies and then combine them with existing DC-based strategies and finally optimize them via the GA. We conduct trading experiments over 44 stocks. Results show that the GA-optimized strategies are able to generate new and profitable trading strategies, significantly outperforming the individual DC-based strategies, as well as a buy and sell benchmark.

Financial Forecasting is a popular and thriving research area that relies on indicators derived from technical and sentiment analysis. In this paper, we investigate the advantages that sentiment analysis indicators provide, by comparing their performance to that of technical indicators, when both are used individually as features into a genetic programming algorithm focusing on the maximization of the Sharpe ratio. Moreover, while previous sentiment analysis research has focused mostly on the titles of articles, in this paper we use the text of the articles and their summaries. Our goal is to explore further on all possible sentiment features and identify which features contribute the most. We perform experiments on 26 different datasets and show that sentiment analysis produces better, and statistically significant, average results than technical analysis in terms of Sharpe ratio and risk.

Real Estate Investment Trusts (REITs) is a popular investment choice as it allows investors to hold shares in real estate rather than investing large sums of money to purchase real estate by themselves. Previous work studied the effectiveness of multi-asset portfolios that include REITs via an efficient frontier analysis. However, the advantages of including (both domestic and international) REITs in multi-asset portfolios, as well as analyzing all the possible combinations of asset classes, has not been investigated before. In this paper, we fill in this gap by performing a thorough investigation across 456 different portfolios to demonstrate the added value of including REITs in mixed-asset portfolios in terms of different important financial metrics. To this end, we use a genetic algorithm approach to maximize the Sharpe ratio of the portfolios. Our results show that optimization via a genetic algorithm outperforms the results obtained from a global minimum variance portfolio. More importantly, our results also show that there can be significant improvements in average returns, risk and Sharpe ratio when including REITs.

The majority of forecasting methods use a physical time scale for studying price fluctuations of financial markets. Using physical time scales can make companies oblivious to significant activities in the market as the flow of time is discontinuous, which could translate to missed profitable opportunities or risk exposure. Directional changes (DC) has gained attention in the recent years by translating physical time series to event-based series. Under this framework, trend reversals can be predicted by using the length of events. Having this knowledge allows traders to take an action before such reversals happen and thus increase their profitability. In this paper, we investigate how classification algorithms can be incorporated in the process of predicting trend reversals to create DC-based trading strategies. The effect of the proposed trend reversal estimation is measured on 20 foreign exchange markets over a 10-month period in a total of 1000 data sets. We compare our results across 16 algorithms, both DC and non-DC based, such as technical analysis and buy-and-hold. Our findings show that the introduction of classification leads to return higher profit and statistically outperform all other trading strategies.

Rainfall is one of the most challenging variables to predict, as it exhibits very unique characteristics that do not exist in other time series data. Moreover, rainfall is a major component and is essential for applications that surround water resource planning. In particular, this paper is interested in extending previous work carried out on the prediction of rainfall using Genetic Programming (GP) for rainfall derivatives. Currently in the rainfall derivatives literature, the process of predicting rainfall is dominated by statistical models, namely using a Markov-chain extended with rainfall prediction (MCRP). In this paper we further extend our new methodology by looking at the effect of feature engineering on the rainfall prediction process. Feature engineering will allow us to extract additional information from the data variables created. By incorporating feature engineering techniques we look to further tailor our GP to the problem domain and we compare the performance of the previous GP, which previously statistically outperformed MCRP, against our new GP using feature engineering on 21 different data sets of cities across Europe and report the results. The goal is to see whether GP can outperform its predecessor without extra features, which acts as a benchmark. Results indicate that in general GP using extra features significantly outperforms a GP without the use of extra features.

Regression problems provide some of the most challenging research opportunities, where the predictions of such domains are critical to a specific application. Problem domains that exhibit large variability and are of chaotic nature are the most challenging to predict. Rainfall being a prime example, as it exhibits very unique characteristics that do not exist in other time series data. Moreover, rainfall is essential for applications that surround financial securities such as rainfall derivatives. This paper is interested in creating a new methodology for increasing the predictive accuracy of rainfall within the problem domain of rainfall derivatives. Currently, the process of predicting rainfall within rainfall derivatives is dominated by statistical models, namely Markov-chain extended with rainfall prediction (MCRP). In this paper, we propose a novel algorithm for decomposing rainfall, which is a hybrid Genetic Programming/Genetic Algorithm (GP/GA) algorithm. Hence, the overall problem becomes easier to solve. We compare the performance of our hybrid GP/GA, against MCRP, Radial Basis Function and GP without decomposition. We aim to show the effectiveness that a decomposition algorithm can have on the problem domain. Results show that in general decomposition has a very positive effect by statistically outperforming GP without decomposition and MCRP.

The deployment of fibre optics poses a huge investment risk, thus telecommunication companies are skeptical about replacing copper given the high cost of doing so. Over recent times, the usage of the internet has changed and led to a need for fibre optics. The decision on whether to deploy or not is made through the use of complex models. However, the problem being that deployment plans are manually predefined based on previous knowledge, this process does not guarantee that the plans are optimal. This paper demonstrates that the deployment of fibre optics can be optimised by using intelligent algorithms. We implemented a metaheuristic (Guided Local Search) to the problem to demonstrate the effectiveness and benefit of looking for an optimal deployment plan. Results indicate that Guided Local Search lead to a significant increase in the profit and can address the problem of finding an optimal deployment plan.

Standard Regression models are presented with n samples from an input space X that is composed of observational data of the form (xi, y(xi)), i = 1...n where each xidenotes a k-dimensional input vector of design variables and y is the response. When k ≫ n, high variance and over-fitting become a major concern. In this paper we propose a novel approach to mitigate this problem by transforming the input vectors into new smaller vectors (called Z set) using only a set of simple statistical moments. Genetic Algorithm (GA) has been used to evolve a transformation procedure. It is used to optimise an optimal sequence of statistical moments and their input parameters. We used Linear Regression (LR) as an example to quantify the quality of the evolved transformation procedure. Empirical evidences, collected from benchmark functions and real-world problems, demonstrate that the proposed transformation approach is able to dramatically improve LR generalisation and make it outperform other state-of-the-art regression models such as Genetic Programming, Kriging, and Radial Basis Functions Networks. In addition, we present an analysis to shed light on the most important statistical moments that are useful for the transformation process.

Hyper-heuristics have successfully been applied to a vast number of search and optimization problems. One of the novelties of hyper-heuristics is the fact that they manage and automate the meta-heuristic's selection process. In this paper, we implemented and analyzed a hyper-heuristic framework on three meta-heuristics namely Simulated Annealing, Tabu Search, and Guided Local Search, which had successfully been applied in the past to a Financial Forecasting algorithm called EDDIE. EDDIE uses Genetic Programming to extract and learn from historical data in order to predict future financial market movements. Results show that the algorithm's effectiveness has been improved, thus making the combination of meta-heuristics under a hyper-heuristic framework an effective Financial Forecasting approach.

Guided Local Search is a powerful meta-heuristic algorithm that has been applied to a successful Genetic Programming Financial Forecasting tool called EDDIE. Although previous research has shown that it has significantly improved the performance of EDDIE, it also increased its computational cost to a high extent. This paper presents an attempt to deal with this issue by combining Guided Local Search with Fast Local Search, an algorithm that has shown in the past to be able to significantly reduce the computational cost of Guided Local Search. Results show that EDDIE's computational cost has been reduced by an impressive 77%, while at the same time there is no cost to the predictive performance of the algorithm.

Financial forecasting is a vital area in computational finance. This importance is reflected in the literature by the continuous development of new algorithms. EDDIE is well-established genetic programming financial forecasting tool, which has successfully been applied to a variety of international datasets. Recently, we introduced hyper-heuristics to EDDIE. This was the first time in the literature that hyper-heuristics were used for financial forecasting. Results showed that this introduction significantly benefited the performance of the algorithm. However, an issue was encountered in the way that lowlevel heuristics were selected during the search process, because it was considered to be a static way. To address this issue, in this paper we further improve our algorithm by introducing a Choice Function, which is a score based technique that offers a more dynamic selection of the low-level heuristics. This paper presents preliminary results, after having tested the Choice Function approach with 10 datasets. These results show that the introduction of the Choice Function is beneficial to EDDIE, thus making it a very promising tool for future investigation on financial forecasting problems.

This work uses a Genetic Programming (GP) algorithm to co-evolve negotiation strategies of agents that have different preference criteria, namely optimizing price and optimizing negotiation speed. While GP and other algorithms have been extensively used for price-only optimization, the problem of price-speed optimization has not yet received the same amount of attention. In Cloud/Grid computing environments, any delay in acquiring resources will be considered an overhead, hence negotiation agents need to adopt strategies that will enable them not only to optimize resource price but also to reach early agreements. This research is the earliest work to apply a GP algorithm for evolving price-speed optimizing negotiation strategies. An important advantage of the GP is its representation, which allows solutions to be represented in terms of the problem parameters, rather than as binary or real-value code, as it has been the case until now with other algorithms. We apply the GP to different negotiation scenarios and compare its results to other previously published works on the problem of pricespeed optimizing negotiation agents. Results show that the GP 1) outperforms the algorithms from these previous works and 2) can evolve to an optimal or near optimal strategy.

EDDIE is a Genetic Programming (GP) tool, which is used to tackle problems in the field of financial forecasting. The novelty of EDDIE is in its grammar, which allows the GP to look in the space of technical analysis indicators, instead of using prespecified ones, as it normally happens in the literature. The advantage of this is that EDDIE is not constrained to use prespecified indicators; instead, thanks to its grammar, it can choose any indicators within a pre-defined range, leading to new solutions that might have never been discovered before. However, a disadvantage of the above approach is that the algorithm's search space is dramatically larger, and as a result good solutions can sometimes be missed due to ineffective search. This paper presents an attempt to deal with this issue by applying to the GP three different meta-heuristics, namely Simulated Annealing, Tabu Search, and Guided Local Search. Results show that the algorithm's performance significantly improves, thus making the combination of Genetic Programming and meta-heuristics an effective financial forecasting approach.

Guided Local Search (GLS), which is a simple meta-heuristic with many successful applications, has lambda as the only parameter to tune. There has been no attempt to automatically tune this parameter, resulting in a parameterless GLS. Such a result is a very practical objective to facilitate the use of meta-heuristics for end- users (e.g. practitioners and researchers). In this paper, we propose a novel parameter tuning approach by using Genetic Programming (GP). GP is employed to evolve an optimal formula that GLS can use to dynamically compute lambda as a function of instance-dependent characteristics. Computational experiments on the travelling salesman problem demonstrate the feasibility and effectiveness of this approach, producing parameterless formulae with which the performance of GLS is competitive (if not better) than the standard GLS.

Fiber optics is a relatively new technology, one which has not yet been extensively used, because of its high cost. In order to evaluate the viability of such a costly investment, techno-economic models are employed. These models evaluate the investment from both technical (e.g., optimal network design) and economical (e.g., profitability) perspectives. However, an area that has not received much attention is the deployment plans of a given fiber optic investment. Existing works usually compare manually predefined deployment plans that are considered profitable, and then apply techno-economic analysis. While this indeed offers valuable information, it does not guarantee that the examined plans are the optimal ones. This should be considered as a major disadvantage, because there could be other deployment plans that could offer significantly higher profit. This paper offers a first attempt at looking for the optimal deployment plan of fiber optics, based on profit. Our method can be considered as a framework that wraps around existing techno-economic models. We employ a Genetic Algorithm (GA), which creates a population of deployment plans. These plans can then be evaluated through the usual techno-economic approach. The GA then evolves the population of these plans and at the end of the process acts as a decision support tool that advises on the optimal deployment plan, without the need for any human interference in the decision-making process. For comparison purposes, we compare the GA's results with results under other profitable plans. Results show that the introduction of the use of the GA is very advantageous and leads to a significant increase in profit.

This paper extends a previous market microstructure model, where we used Genetic Programming (GP) as an inference engine for trading rules, and Self Organizing Maps as a clustering machine for those rules. Experiments in that work took place under a single financial market and investigated whether its behavior is non-stationary or cyclic. Results showed that the market's behavior was constantly changing and strategies that would not adapt to these changes, would become obsolete, and their performance would thus decrease over time. However, because experiments in that work were based on a specific GP algorithm, we are interested in this paper to prove that those results are independent of the choice of such algorithms. We thus repeat our previous tests under two more GP frameworks. In addition, while our previous work surveyed only a single market, in this paper we run tests under 10 markets, for generalization purposes. Finally, we deepen our analysis and investigate whether the performance of strategies, which have not co-evolved with the market, follows a continuous decrease, as it has been previously suggested in the agent-based artificial stock market literature. Results show that our previous results are not sensitive to the choice of GP. Strategies that do not co-evolve with the market, become ineffective. However, we do not find evidence for a continuous performance decrease of these strategies.

In this paper we present a new version of a GP-based financial forecasting tool called EDDIE. The novelty of this new version (EDDIE 8), is its enlarged search space, where we allow the GP to search in the space of the technical indicators, in order to form its Genetic Decision Trees. In this way, EDDIE 8 is not constrained in using pre-specified indicators, but it is left up to the GP to choose the optimal ones. We then proceed to compare EDDIE 8 with EDDIE 7, which is based on previous EDDIE versions; EDDIE 7 has a smaller space where the indicators are pre-specified by the user and are part of EDDIE 8's space. Results show that thanks to the bigger search space, new and improved solutions can be found by EDDIE 8. However, there are cases where EDDIE 8 can still be outperformed by its predecessor. Analysis shows that this depends on the nature of the solutions. If the solutions come from EDDIE 8's search space, then EDDIE 8 can find them and perform better; if, however, solutions come from the smaller search space of EDDIE 7, then EDDIE 8 is having difficulties focusing in such a small space and is thus outperformed by EDDIE 7.