Predicting stock prices over multiple future days is a well-known challenge because financial time-series data is inherently noisy, non-stationary, and often chaotic. Whereas current research tends to emphasize the use of sophisticated deep learning methods, the present study is focused on addressing an entirely different question to what extent does the design of a basic ANN structure affect its performance. Initially, a comprehensive ANN architecture search was performed for fifteen different ANN models ranging from single-layer ANNs to six-layer bottleneck ANNs in order to explore the effect of ANN depth and ANN width on the ability of ANNs to forecast multi-step ahead returns. All ANNs were trained using a walk-forward testing procedure based on the change-in-z score target model formulation. To add another layer of pressure testing to our results, we deploy and analyze several cutting-edge architectures like standalone LSTMs, CNN-LSTMs, Random Forests, ARIMA, as well as a novel Hybrid CNN-Attention-LSTM model with a proprietary gating feature module. An experimental analysis reveals a surprising result whereby the optimally fine-tuned ANN establishes a remarkably high performance ceiling that is difficult for deep sequence learning models to match. Though the hybrid models utilizing feature gating were able to reach parity with the optimal ANN baseline, none exceeded its performance. Particularly, the optimally tuned ANN either equalled or surpassed the best hybrid models on five out of seven tested assets. More elaborate variations of the hybrid model, such as those using residual connections and snapshot ensembling techniques, actively reduced model performance due to the identity mapping issue.
Tejas et al. (Fri,) studied this question.