Portfolio allocation with Deep Reinforcement Learning (DRL) has been the focus of many researchers. In investing, a portfolio optimization strategy is selecting assets that maximize return on investment while minimizing the risk. The task of asset optimization involves balancing risk and return, where stock returns are profits over a period of time and risk is the standard deviation value of the asset`s return. Many of the existing methods for portfolio optimization are essentially the expansion of diversification methods for assets in the investment. Signiant drawdowns and early entry into the share are still a challenge in portfolio construction. The idea here is that having a portfolio based on money net flow is less risky than allocating a portfolio based on historical data only and turbulence as risk aversion. In this paper, we propose a profitable stock recommendation framework for portfolio construction using DRL model based on the money net flow (MNF) indicator. We develop a new risk indicator based on the intelligent net-flow behavior of smart money to help determine the optimal market timing for buying and selling. The experimental results by real-world trading scenario validation show the model clearly outperforms all the considered baselines, and even the conventional Buy-and-Hold strategy. Moreover, in this paper, the effect of defining different environments made of various information with hyperparameter optimization on the performance of models has been investigated, and the performance of DRL-driven models in different markets and asset positions has been investigated. The empirical results show the dominance of DRL models based on MNF indicator.