This chapter first introduces a conditional quantile model that is capable of capturing both long-term and short-term dependencies in sequential data. The model is built upon the framework of sequential neural networks and is designed to capture interpretable dynamics. It is applied to time series data of asset returns from the 1960s to 2018. The analysis reveals that the model is effective in extracting not only the serial dependence structure in conditional volatility but also the latent memories in the tails of return distributions. The proposed model demonstrates superior performance compared to the GARCH family, as well as models such as filtered historical simulation, conditional extreme value theory, and dynamic quantile regression. These results suggest that the conditional quantiles of asset returns are influenced by risk sources that are distinct from those responsible for volatility clustering. These findings have significant implications for risk management, particularly in the context of forecasting tail risks. After these, in this chapter, we then introduce a novel, succinct, and effective approach for distribution prediction aimed at quantifying uncertainty in machine learning. Our method adaptively produces flexible distribution prediction of \(\mathbb {P}(\textbf{y}|\textbf{X}=x)\) in regression tasks. Specifically, we enhance the quantiles of this conditional distribution with probability levels spreading over the interval (0, 1), using additive models that we have designed with intuition and interpretability. Our approach seeks to achieve an adaptive balance between the structural integrity and the flexibility of \(\mathbb {P}(\textbf{y}|\textbf{X}=x)\) . We observe that the Gaussian assumption often results in a lack of flexibility for real-world data, and approaches that are highly flexible (e.g., estimating quantiles separately without a distribution structure) may have inherent drawbacks and potentially lead to suboptimal generalization. To address these challenges, we propose an ensemble multi-quantiles approach, termed EMQ, which is entirely data-driven. EMQ is capable of gradually deviating from the Gaussian assumption and discovering the optimal conditional distribution during the boosting process. We evaluate the performance of EMQ on a wide range of regression tasks using UCI datasets and demonstrate that it achieves state-of-the-art performance compared to many recent uncertainty quantification methods. Furthermore, our visualization results underscore the necessity and the merits of such an ensemble model in capturing the underlying distributional characteristics of the data effectively.