Temporal Convolutional Network (TCN)
A sequence model that applies stacked, dilated, causal convolutions to time-series data, capturing long histories in parallel without the step-by-step recurrence of an RNN.
A temporal convolutional network is a deep-learning architecture for sequences that uses convolutional filters instead of recurrence. Rather than walking through a series one step at a time, a TCN slides learnable filters across the time axis to detect local patterns, then stacks many such layers to build up a view of longer history. Two design choices make it work for time series. First, the convolutions are causal, meaning each output depends only on the current and past inputs and never peeks at the future, which prevents look-ahead leakage. Second, the layers use dilation, where filters skip over inputs at exponentially increasing gaps, so a modest stack of layers can see very far back without an explosion in parameters.
The main advantage of a TCN over a recurrent network is that its computations across time steps are independent and can run in parallel, making training faster and more stable. It also sidesteps the vanishing-gradient issues that plague basic RNNs, because the path from a distant input to the output is short and direct through the dilated structure. The effective length of history a TCN can use, called its receptive field, is set explicitly by the number of layers, the filter size, and the dilation schedule, which gives a designer clear control over how far back the model looks.
For markets, a TCN is well suited to detecting patterns that recur at multiple time scales, such as short-term mean reversion overlaid on a longer trend, because different layers specialize in different horizons. Its causal design is particularly valuable in a financial setting, where accidentally using future information would produce backtests that are impressive but impossible to trade. As with any flexible model, the receptive field and capacity must be matched to the data to avoid overfitting.
Hedgewing.ai includes a TCN as one of the four models in its forecasting ensemble, alongside an LSTM, a GRU, and a Transformer. The TCN contributes a fundamentally different inductive bias from the recurrent models, which is the point of an ensemble. It reads the same 45 engineered features and is evaluated through nightly walk-forward backtesting, and its strictly causal structure aligns with the platform's discipline of avoiding look-ahead bias in every prediction.
Related terms
Neural Network · Deep Learning · Ensemble Model · Look-Ahead Bias
Back to the hedgewing.ai glossary · See AI stock forecasts