Models need sufficient data in order to learn typical behavior. A single or small number of time steps are typically not sufficient for training the neural network (unless the time series are very long). While a Loud ML model trained on a few hundreds data points will usually still generate sensible forecasts, standard forecasting methods such as ARIMA or ETS may be more accurate and stable. Where the ANN approach in Loud ML starts to outperform the standard methods is when your dataset contains thousands of data points and thus can be significantly more accurate with more data.
You are looking at documentation for an older release. Not what you want? See the current release documentation.