<p>The use of dropout as a regularization approach has been widely used in deep learning, especially in recurrent models such as LSTM. Conventional dropout strategies, however, when used on sequential data with high variability in time-series domain (i.e. with stock prices) tend to be less optimal. In this work, a new framework of Temporally Adaptive Dropout LSTM (TAD-LSTM) is introduced, and it is characterized by varying dropout probabilities per timestep based on the differences between the actual stock price at the current timestamp and the previous timestamps. This time conscious process increases regularization and resilience during present turbulent market environments. In order to provide even more flexibility, six variants are proposed: Attention-Guided TAD-LSTM (AG-TAD-LSTM) includes attention scores to guide the behavior of dropout; Reinforcement Learning-based TAD-LSTM (RL-TAD-LSTM) learns dropout strategies by search based on the reward; Hierarchical TAD-LSTM (HTAD-LSTM) applies dropout modulation across more than one LSTM layer; Temporally Adaptive Variational Dropout (TAVD-LSTM) provides a combination of the temporal adaptation Empirical analysis of real stock market data including NIFTY 50, NSE and AAPL stock market data show in detail enhanced next-day stock price prediction and generalization, which are better than fixed and variational dropout baselines in terms of measurement metrics, including MAE, RMSE, MAPE and R². The suggested framework provides a scalable and flexible way of dropout regularization of financial time-series modeling. In addition to next-day prediction, experiments are conducted based on longer-horizon evaluation (5-, 10-, and 20-day ahead), which indicates that the suggested TAD-LSTM variants preserve even better accuracy and strength when predicting more challenging scenarios.</p>