TCN-finance

Conditional time series forecasting with convolutional neural networks. Inspired from:

• Conditional time series forecasting with convolutional neural networks by Anastasia Borovykh, Sander Bohte and Cornelis W. Oosterlee (2018)
• Probabilistic Forecasting with Temporal Convolutional Neural Network by Yitian Chena, Yanfei Kangb, Yixiong Chenc, Zizhuo Wangd (2020)

Why Temporal Convolutional Network?

• TCNs exhibit longer memory than recurrent architectures with the same capacity.
• Constantly performs better than LSTM/GRU architectures on a vast range of tasks (Seq. MNIST, Adding Problem, Copy Memory, Word-level PTB...).
• Parallelism, flexible receptive field size, stable gradients, low memory requirements for training, variable length inputs...

Visualization of a stack of dilated causal convolutional layers (Wavenet, 2016)

Arguments

• dataset: pandas.Dataframe. The networks are trained to predict information from this dataset
• dataset2: pandas.Dataframe. This dataset is used as additional information to help predict values from the first dataset
• device : cpu or gpu
• epochs: Integer. Number of training epoch
• kernel_size: Integer. Size of the kernel used in convolutions
• checkpoint: Integer. Number of epoch after which networks are saved.
• receptive_field: Integer. Number of previous dates information used to forecast the future
• n_layers_encod: Integer. Number of layer of the encoder
• n_outputs: list of Integer. Number of channel of convolutions of the encoder
• dilation: Integer. Dilatation parameter of the convolution
• stride: Integer. Stride parameter of convolution
• n_inputs: Integer. Number of asset
• n_channel_decod: Integer. Number of channel of the decoder
• future_size: Integer. Number of step forecasted in the future
• learning_rate: Float. Learning rate used for parameter optimization
• sep_train_test: Integer. Index of separation between training set and test set
• outf: String. Folder where networks are saved.
• resume: Boolean. True if networks have previously been trained and one wishes to resume training