We set up multiple LSTM networks to predict the weekly returns of ten different stocksduring the uncertainty of the Covid-19 pandemic. Three different models were proposed foreach stock, which were trained on historic pricing data of different frequencies and sequencelengths. Two of these models used 5-minute price data in sequences of 3 and 5 trading days,respectively. The third model used daily closing price observations in sequences of 90 tradingdays.Out-of-sample returns were forecasted to evaluate the economic value of the LSTM models.These were then used in an MVO framework to construct optimal portfolios based on arisk averse investor. A simple DCC-GARCH model was created to forecast the variance-covariance matrix of the ten stocks. All the forecasted returns and variance-covariancematrices were based on out-of-sample test data, ensuring that the networks had not seen thedata. We do not find any significant economic gain by using the LSTM based forecast inconstructing portfolios compared to a naive forecast during the out-of-sample period.