Ever since data science is growing so has its application. The portfolio allocation problem can be solved with ease using the Data Science framework. As there is no generalized method for portfolio allocation, the Modern Portfolio Theory (MPT) formulated by Markowitz has been a benchmark and widely used model (mean-variance). This model limits due to the highly stochastic nature of the market, whereas the existing model such as deep deterministic policy, random forest and other machine learning-based models lack in adapting to the stochastic market nature. In this study, a novel technique is proposed which is inspired by the success of the attention mechanism in natural language processing. The proposed model can quickly adapt to the changing market with its full encoder-decoder architecture, encoded layers, and gated components. It is trained and tested on three datasets i.e., Stock, ETFs & commodities. The accuracy of the model is checked on the widely used and industry go-to parameter Sharpe ratio, which measures the returns on risk. It has been found that using the proposed approach the sharp ratio, return and positive return % have been improved.