From my experience, LSTM or other recurrent neural network models only work "well" at forecasting bounded and periodic or oscillating time series. Might work for something like seasonal sale data, but would fail spectacularly with unbounded and chaotic time series like stock prices.