TCN-Driven Volatility-Robust Forecasting in Minute-Resolution Cryptocurrency Markets

Abstract

This paper proposes Temporal Convolutional Networks (TCNs) for cryptocurrency forecasting at minute-resolution. TCNs show better accuracy to XGBoost, LightGBM, and LSTM. However, TCNs are less robust than the tree models regarding volatility. While TCNs require much more computations at inference than LightGBM, they run faster than LSTMs. The performance of TCNs is best configured using a TCN with dilated convolutions to capture the temporal patterns, and with residual connections for the stability.