Recently, deep learning density networks have become the standard baselines for financial time-series forecasting. In this paper, we introduce density networks structured using the two-stage volatility-to-return models, in which the volatility forecasts are treated as an input to the scale parameter of the second-stage return model. The models employ a negative log-likelihood loss for both stages, which is constructed from the Weibull or a generalized beta distribution for stage one and normal (Gaussian)or Student-t for stage two. VaR is computed directly from the fitted returns distribution using mean regression. However, it is well known that effects on the mean can differ substantially from those on the extreme quantiles, the focus for risk management in investment portfolios. Hence, nonparametric quantile and expectile regressions are considered by extending the return loss to asymmetric Laplace and Gaussian proxy distributions. The time-varying variance from the first stage weights the observations inversely to their variances. Our empirical examination focuses on daily returns and Parkinson volatility measures of Bitcoin and Ripple. In comparison with the parametric VaR, the nonparametric quantiles and expectiles show great adaptiveness and accuracy to observed data.