Transforming Stock Price Forecasting: Deep Learning Architectures and Strategic Feature Engineering

This paper delves into the multifaceted landscape of stock prices, as their nature makes accurate forecasting a considerable challenge for investors and researchers. The analysis encompasses three key stock datasets: AAPL, AMZN, and MSFT. Each stock incorporates raw datasets with OHLCV information and enhanced datasets with subsets of raw data, technical, and macroeconomic indicators. Whilst, adjusted closing price is the target feature. First, we clarify the dominance of Deep Learning (DL) over Machine Learning (ML) and statistical models in predicting stock prices. While statistical and ML models, such as ARIMA, SARIMA, ETS, GARCH, SVM, RFR, and XGBoost, have shown commendable reliability, the rapid technological evolution prompts an examination of their current relevance. Eventually, their performance is surpassed by DL models, such as LSTM, CNN-Conv1D, and SimpleRNN. Preliminary findings on enhanced data assure SimpleRNN as the most precise model, exhibiting the lowest evaluation metrics (MSE and MAE). However, SimpleRNN, in comparison to additional RNN-based models, such as GRU, BiLSTM, Convolutional LSTM, and ConvLSTM2D, is outperformed by BiLSTM. Therefore, BiLSTM is applied to enhanced and raw datasets. The results confirm that BiLSTM, in association with the enhanced dataset, averagely reduce the MSE by 68.44% and the MAE by 47.78%. The overall outcome of this research is devoted to the future renovation of forecasting principles in statistics and algorithmic trading.