WaveCorr: Deep reinforcement learning with permutation invariant convolutional policy networks for portfolio management

We present a new portfolio policy convolutional neural network architecture, WaveCorr, for deep reinforcement learning applied to portfolio optimization. WaveCorr is the first to treat asset correlation while preserving “asset invariance property”, a new permutation invariance property that significantly increases the stability of performance in problems where input indexing is done arbitrarily. A general theory is also derived for verifying this property in other fields of application. Our experiments show that WaveCorr consistently outperforms other state-of-the-art convolutional architectures.