| Summary: | 碩士 === 輔仁大學 === 統計資訊學系應用統計碩士在職專班 === 103 === With the rapid globalization of world economics, foreign exchange markets have become indispensable media and tools for international economic activities. In general, linear time series models such as the Auto Regressive Integrated Moving Average model (ARIMA) are used to forecast currency trading. However, most financial time series models are nonlinear and may not necessarily submit to normal distribution assumptions. Therefore, the predictive capabilities of linear time series models are inadequate in forecasting the fluctuating conditions of financial environments. Following the advancements in information technology, nonlinear nonparametric prediction techniques gradually emerged; among these techniques, Artificial Neural Network (ANN) techniques are prominent. Although ANNs can be employed to process nonlinear data, these techniques exhibit numerous shortcomings such as an excessive amount of time required to train data, black-boxing properties, and the inability to associate dependent and independent variables. In this context, this study endeavored to apply a novel algorithm introduced in the field of machine learning, namely, the Random Forest model (RF), to process nonlinear data. The RF effectively classifies and forecasts data and does not submit to the assumption criteria required in conventional statistical regression models. In addition, this model presents several advantages such as it evaluates the significance of variables, determines whether the variables can uniformly/non-uniformly classify the deviation of data sets, and accelerates learning processes. In contrast to previous studies, incorporated a large quantity of variables extracted from the financial market and technical indicator data into the present study. This study attempted to examine the performance of the RF in classifying USD/EUR currency trading trends, identifying key variables as well as assessing the benefits and risks of the model forecasts. For currency trading trends, the empirical results indicated that the proposed model could effectively distinguish between the three Forex classifications and extract variable importance measures. For currency trading forecast, the model forecasts were extremely accurate. Model testing results indicated positive yields with little downside risk. Therefore, the proposed model effectively aids investors in determining the trend in currency movement during trade.
|
|---|
