The price forecast accuracy chart is from a modelling engagement we had taken up for a client, hence have shared the actual numbers after anonymizing the product names.

We used simple neural networks here and not RNN/CNNs. The non-stationary data was converted into stationary before feeding to the model. However, we’ve come across cases where you could feed in the non-stationary data directly, depends on the context. Heteroscedasticity wasn’t much of a concern for non-linear, black box models like the neural networks.