International Journal of Academic Research and Development

Authors
Yasir Ahmad Itoo, Ikhlas Ahmad Sheikh

Abstract
A time series is a series of data points ordered in timely manner. Time Series can be regular or irregular based on the intervals over which it has been collected. Usually, the time series data has been collected over a period of time at regular intervals and same is used to observe patterns among it thereby predicting or forecasting future events. Apart from the conventional regression-based [1] modeling, deep learning-based algorithms [2] are the well-meaning approaches in addressing prediction/forecasting problems in time series and where the latter technique have been shown to produce more accurate results than former. A Lot of Researches suggest that Recurrent Neural Networks (RNN) with memory, such as Long Short-Term Memory (LSTM), are superior compared to Autoregressive Integrated Moving Average (ARIMA) or Seasonal Autoregressive Integrated Moving Average (SARIMA) with a huge margin. In order to memorize large sequences, the neurons in LSTM-based models comes with three different gates. Are these gates alone enough for the purpose of memorizing? Will more training on data improve the prediction or forecasting? Bidirectional LSTMs provide further training on the input data twice. i.e., It first traverses the data from the left to the right and then traverses back from right to left. Whether BiLSTM, with additional training capability, outperforms regular unidirectional LSTM? Here we report the behavioral analysis as well as comparison of LSTM and BiLSTM models. The objective is to explore to what extend additional layers of training of data would be beneficial to tune the involved parameters. The results show that additional training of data and thus The results show that BiLSTM based modeling offers better predictions than LSTM based models. Even the BiLSTM based models provide a lot fine predictions when compared to the ARIMA as was observed practically. Another important thing to mention is that BiLSTM models take much time to reach the equilibrium than the regular LSTM-based models.