Predicting Financial Markets with Deep Learning Methods: Analyzing AAPL Stock, BTC-USD and EUR/USD Exchange Rates

Abstract A combination of advanced recurrent and convolutional networks is proposed in this study for financial asset price estimation. The performance of Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU) and the hybrid CNN-LSTM is evaluated, as the latter has not been thoroughly studied despite being able to find both spatial and temporal patterns [6]. This research covers three kinds of financial assets—Apple stocks (AAPL), Bitcoin cryptocurrency (BTC-USD) and the Euro to U.S. Dollar exchange rate (EUR/USD), in contrast to many that only analyze stocks [2,6,11,13,14], cryptocurrencies [16,19] or, very rarely, exchange rates. Such an approach allows us to apply predictive models to different markets. I gathered financial history from Yahoo Finance and then included four major technical indicators: SMA, EMA, RSI and MACD, while other studies might just use the raw data itself or a smaller number of indicators [15,17]. Standardized data was divided into sequences for a prediction period of 60 days. Assessing the model involved means of MAE and R², giving a more balanced result than obtained by using just RMSE or MAPE [13–15]. Choosing these methods allows for a broader and reusable approach for financial time series forecasting, as compared to previous studies limited in architecture or evaluation [2,14,15].