Cascading model for Forex market forecasting using fundamental and technical indicator data based on bert

Abstract

The foreign exchange rate market is the world's biggest and most liquid financial market, and it's where all currency pairs' exchange rates are set. Since foreign exchange (Forex) rates play a critical role in financial technology and business, many researchers are now interested in forecasting them. The characteristics of Forex data, that include fluctuation, non-linearity, and random walk phenomena, make it difficult for forecasting. Several related studies integrate fundamental data (FD) and technical indicator data to generate Forex forecasting signals (TI). TI is a price pattern-based signal, whereas FD is an indicator of the country's economic conditions. Nevertheless, when it comes to deployment, these two indicators have two major drawbacks. Gradient vanishing and information loss occur when modeling a sequential neural network. Furthermore, although FD has a big impact on currency prices, it was updated quarterly or monthly which is not as frequent as price change. This restriction is known as the FD releasing problem. Moreover, Forex forecasting with FD and TI is usually done in equal aggregation, which leads to inaccurate predictions due to unequal data changing frequency. In this paper (BERTFOREX), we introduce a cascading model for forex market forecasting using FD and TI based on BERT (BERTFOREX). The following are the steps in the BERTFOREX processing system: 1) BERT is applied to FD to extract hidden patterns. 2) Because the frequency of FD changes more slowly than that of TI, these hidden FD patterns are aggregated as additional weights for TI. 3) BERT is used to extract the aggregated pattern within TI and FD. 4) The BERTFOREX efficiency is demonstrated by feeding the aggregated pattern into a simple neural network for forecasting. From the experimental results, the proposed method outperforms other methods in terms of correct signal percentage, sensitivity, specificity, precision, and negative predictive value.