Estimating Directional Changes Trend Reversal in Forex Using Machine Learning

Most forecasting algorithms use a physical time scale data to study price movement in financial markets by taking snapshots in fixed schedule, making the flow of time discontinuous. The use of a physical time scale can make traders oblivious to significant

activities in the market, which poses risks. For example, currency risk, the risk that exchange rate will change. Directional changes is a different and newer approach of taking snapshot of the market, which uses an event-based time scale. This approach summarises data into alternating trends called upward directional change and downward directional change according to a change in price a trader considers to be significant, which is expressed as a threshold. The trends in the summary are split into directional change (DC) and overshoot (OS) events. In this work, we propose a novel DC-based framework, which uses machine learning algorithms to forecast when the next, alternate trend is expected to begin. First, we present a genetic programming (GP) algorithm that evolves equations that express linear and non-linear relationships between the length of DC and OS events in a given dataset.

Awareness of DC event and OS event lengths provide traders with an idea of when DC trends are expected to reverse and thus take appropriate action to increase profit or mitigate risk. Second, DC trends can be categorised into two distinct types: (1)

trends with OS events; and (2) trends without OS events(i.e. OS event length is 0). Trends with OS events are those that continue beyond a period when they were first observed and trends without OS event are others that ends as soon as they were observed. To further improve trend reversal estimation accuracy, we identified these two categorises using classification techniques and estimated OS event length for trends that belong in the first category. We appraised whether this new knowledge could lead to an even greater excess return. Third, our novel trend reversal estimation approach was then used as part of a novel genetic algorithm (GA) based trading strategy. The strategy embedded an optimised trend reversal forecasting algorithm that was based on trend reversal point forecasted by multiple thresholds. We assessed the efficiency of our framework (i.e., a novel trend reversal approach and an optimised trading strategy) by performing an in-depth investigation. To assess our approach and evaluate the extent to which it could be generalised in Forex markets, we used five tailored thresholds to create 1000 DC datasets from 10, monthly 10-

minute physical time data of 20 major Forex markets (i.e 5 thresholds * 10 months * 20 currency pairs). We compared our results to six benchmarks techniques, both DC and non-DC based, such as technical analysis and buy-and-hold. Our findings showed that our proposed approach can return a significantly higher profit at reduced risk, and statistically outperformed the other trading strategies compareds in a number of different performance metrics.