MQ-CNN

MQ-CNN (Multi-horizon Quantile - Convolutional Neural Network) is a convolutional neural network that uses a quantile decoder to make predictions for the next forecast horizon values, given the preceding context length values. It also uses time features, automatically computed based on the selected time frequency. The model forecasts the same quantiles as the ones selected for the evaluation metrics.

Training parameters

Batch size

The size of the batches to be used for training and prediction.

Epochs

Use default nb. of batches per epoch

Default value is the number of batches for which on average each time step appears in one sample per epoch.

Number of batches per epoch

Random state

Using a fixed random seed allows for reproducible result.

Model parameters

Use default context length

Set the context length to the model's default value (context length = 4 * forecast horizon).

Encode identifiers

Encode time series identifiers and use them as external features.

MLP layer sizes

Dimensions of the Multi Layer Perceptron layers of the decoder. Specify one value per layer.

Number of channels

Number of channels for each layer of the encoder (which is a stack of dilated convolutions). Specify one value per layer. More channels usually means better performance and larger network size.

Convolution dilations

Dilation of the convolutions in each layer of the encoder. Specify one value per layer. Should be of same length as the number of channels. Greater numbers correspond to a greater receptive field of the network, which is usually better with longer context lengths.

Kernel sizes

Kernel sizes (i.e. window sizes) of the convolutions in each layer of the encoder. Specify one value per layer, each must be strictly greater than 1. Should be of same length as the number of channels.

Scale target values

For each time series, rescale the target by its average absolute value.