Predicting Stock Market Prices using Fine-Tuned IndRNN

Abstract

Prediction and analysis of stock market data have a vital role in current time’s economy. The various methods used for the prediction can be classified into 1) Linear Algorithms like Moving Average (MA) and Auto-Regressive Integrated Moving Average (ARIMA). 2) Non-Linear Models like Artificial Neural Networks and Deep Learning. In this work, we are using the results of previous research papers to demonstrate the potential of some models like ARIMA, Multi-Layer Perception (MLP) ), Convolutional Neural Neural Network (CNN), Recurrent Neural Network (RNN), Gated Recurrent Unit (GRU), Long-Short Term Memory (LSTM) for forecasting the stock price of an organization based on its available historical data. Then, implementing some of these methods to check and compare their efficiency within the same issue. We used Independently RNN (IndRNN) to explore a better efficiency for stock prediction and we found that it gives better accuracy prevailing methods in the current time. We also proposed an enhancement to IndRNN by replacing its default activation function with a more effective function called Parametric Rectified Linear Unit (PreLU). Our proposed approach can be used as an alternative method for predicting time series data efficiently other than the typical approaches today

Review of ML and AutoML Solutions to Forecast Time-Series Data

“…Linear models [ 22 , 23 ] for time-series forecasting such as ARIMA [ 24 ] have been prominent for a long time and many researchers still rely on such models as they can predict efficiently and provide interpretability. However, advances in machine learning research concluded that DL and neural networks can be more powerful models [ 25 ], as they can give higher accuracy [ 7 ]. Most machine learning algorithms require extensive domain knowledge, pre-processing, feature selection, and hyperparameter optimization to be able to solve a forecasting task with a satisfying result [ 26 ].…”

“…Combining AR and MA models constitute another model called the Auto-regressive Moving Average (ARMA) [ 32 , 33 ]. Another similar, but more advanced, model to ARMA is the Auto-Regressive Integrated Moving Average (ARIMA) [ 7 , 34 ]. These five mainly used linear models are elaborated below:…”

“…Many works compared DL with linear models in predicting time-series data like cryptocurrency and stock market prices. McNally et al [ 93 ], Phaladisailoed et al [ 94 ], Ahmad et al [ 7 , 95 ]. The results in all of the four works showed that DL-based models are more accurate compared to linear models.…”

“…Linear models which are the simplest and usually the fastest to execute can perform predictions [ 6 ] but might result in low prediction accuracy. DL models can perform better [ 7 ] but as with all machine learning methods [ 8 – 10 ], DL requires experience to tweak hyperparameters and set up the model [ 11 ]. Moreover, the number of hyperparameters to optimize might become large and require considerable effort to optimizing.…”