Time-Series Forecasting in Retail Industry Using Bidirectional, Stacked, and Vanilla LSTMs

Recently, interest in deep learning research and its applicability to practical issues has grown significantly. Developing a time-series analysis model to comprehend sales and profits/losses, as well as forecast future values, is crucial for businesses and companies, whether they operate online or offline. The objective of this study is to construct a time-series analysis model that can comprehend sales and profits/losses while forecasting future values. To achieve an effective analysis, we have chosen Long Short-Term Memory (LSTM) deep learning architectures, including Stacked LSTM, Vanilla LSTM, and Bidirectional LSTM (Bi-LSTM). LSTM, in contrast to conventional recurrent neural networks, can handle time steps of varying sizes without encountering the issue of vanishing gradients. Additionally, they overcome the limitation of the stationarity assumption that is present in models like ARIMA, making them a more flexible and powerful tool for time-series analysis. The three distinct LSTM models are used to train the dataset and are compared with each other with respect to their accuracy measures. The conclusion of the thesis suggests that utilizing the Stacked LSTM deep learning architecture can greatly enhance the accuracy of sales prediction using financial data. Also, the thesis includes the forecast for the next 12 months. The implications of this thesis are significant for businesses and companies, as accurate sales prediction can help in making informed decisions related to production, inventory management, and marketing strategies. Furthermore, the findings of the thesis can also contribute to the realm of deep learning research, particularly concerning of time-series analysis.