Portfolio optimization is a crucial task in investment management, where asset managers attempt to maximize the returns while simultaneously minimizing the risk of their investments. In this paper, we introduce different adaptations of Long Short-Term Memory (LSTM) models to predict the expected returns and volatility preceding portfolio optimization with Mean-Variance Optimization (MVO). We compare the performance of the different LSTM-MVO models using historical data on selected assets and evaluate their effectiveness on portfolio optimization. Our results show that LSTM models can predict the expected returns and volatility better than historical performance alone, which suggests ML models have a potential ancillary place in portfolio management.