On Increasing the Scope of Genetic Programming Trading Agents
MetadataShow full item record
This research investigates the potential for widening the scope of Genetic Programming (GP) trading agents beyond constructing decision trees for buy-hold-sell decisions. First, both technical indicators (temporal feature construction) and decision trees (action selection) are co-evolved under the machine learning paradigm of GP with the benefit of setting Stop-Loss and Take-Profit orders using retracement levels demonstrated. GP trading agents are then used to design trading portfolios under a frequent intraday trading scenario. Such a scenario implies that transaction costs have a more significant impact on profitability and investment decisions can be revised frequently. Furthermore, existing long term portfolio selection algorithms cannot guarantee optimal asset selection for intraday trading, thus motivating a different approach to asset selection. The proposed algorithm identifies a subset of assets to trade in the next day and generates buy-hold-sell decisions for each selected asset in real-time. A benchmarking comparison of ranking heuristics is conducted with the popular Kelly Criterion, and a strong preference for the proposed Moving Sharpe ratio demonstrated. Moreover, the evolved portfolios perform significantly better than any of the comparator methods (buy-and-hold strategy, investment in the full set of 86 stocks, portfolios built from random stock selection and Kelly Criterion). Transaction costs (explicit and implicit or hidden) are important, yet often overlooked, attributes of any trading system. The impact of hidden costs (bid-ask spread) is investigated. The nature of bid-ask spreads (fixed or floating) is demonstrated to be important for the effectiveness of the automated trading system and a floating spread is shown to have a more significant impact than a fixed spread. Finally, the proposed GP framework was assessed on non-financial streaming data. This is significant because it provides the basis for comparing the proposed GP framework to alternative machine learning methods specifically designed to operate under a prequential model of evaluation. The GP framework is shown to provide classification performance competitive with currently established methods for streaming classification, and thus its general effectiveness.