Forecasting Forex Rates Using Optimized Machine Learning Models

The foreign exchange market has the characteristics of high liquidity and high leverage,and its unique characteristics bring high returns to investors as well as high risks.Foreign exchange transactions mainly profit by predicting the direction of the exchange rate between two currencies,so it is very important to judge the direction of the rise and fall.The direction prediction problem is different from the typical time series prediction problem,and the applicable prediction methods are also different.Studies have pointed out that machine learning models,especially the Long Short Term Memory(LSTM)series models,are very effective in foreign exchange direction prediction(Shen et al.,2015;Galeshchuk and Mukherjee,2017;Y?ld?r?m et al.,2021).Based on the theoretical research of disposition effect in behavioral finance,as well as historical data analysis,the exchange rate return has the characteristics of rising and falling asymmetry,which is used as a starting point to split the rise and fall for model improvement research.Recently,Y?ld?r?m et al.(2021)also used the LSTM model to study the improvement of the model for predicting the exchange rate of the Euro to the US dollar,but compared with this,this article considers more exchange rate forecasts,and selects a total of 8 major currencies such as the US dollar and the Australian dollar as the forecast.Goal,the improved model is more versatile.This paper selects 1,726 trading days as the sample period.Compared with the literature,the sample size selected in this paper is sufficient.The article published by Yildirim et al.in 2021 selected 1,214 trading days as the sample size and applied it to the LSTM model.In addition,this article innovatively proposes a research idea that the asymmetry of ups and downs has an important impact on foreign exchange forecasts.This paper first compares the predictive capabilities of the random forest model and the LSTM series model and selects the BLSTM with the best predictive capability as the basic model,and further models based on the asymmetry of fluctuations,analyzes the improvement effect through predictable directional indicators,and finally focuses on different Investors have constructed an upward strategy and a downward strategy to further test the model improvement results.The empirical results show that:(1)BLSTM is the optimal model when predicting multiple exchange rates.Random forest is not as effective as the LSTM system model in this field,and LSTM and BLSTM deep networks can better adapt to time windows of different lengths.(2)The asymmetry of the rise and fall of the foreign exchange rate of return has a significant impact on the foreign exchange forecast.Splitting the rise and fall into the forecast can improve the accuracy rate of up to14.61%.The improved BLSTM model's exchange rate forecast direction accuracy rate is mostly above 52% Therefore,it is necessary to consider the characteristics of ups and downs asymmetry in the foreign exchange forecasting model.(3)The improved model that has undergone the ups and downs splitting has economic implications,and can obtain an excess return of up to 7.4550%,and the income obtained by the ups and downs splitting is higher than that obtained by the undifferentiated ups and downs forecasts.Although machine learning models have been applied in the field of foreign exchange,compared with existing research,the main contribution of this article is to innovatively propose a research idea that the asymmetry of ups and downs has an important impact on foreign exchange forecasts,and provides a new evaluation method Measure the improvement effect of the model's directional forecast,focus on the important influence of the asymmetry of rise and fall on the foreign exchange forecast,and conclude that the asymmetry of rise and fall has a significant impact on the foreign exchange forecast,which provides an important basis for follow-up research.