Time-Dependent Series Variance Estimation via Recurrent Neural Networks

This paper presents a nonlinear model for computing the time-dependent evolution of the variance in time series of returns on assets. First, we design a recurrent network representation of the variance, which extends the typically linear models. Second, we derive temporal training equations with which the network weights are inferred so as to maximize the likelihood of the data. Experimental results show that this dynamic recurrent network model yields results with improved statistical characteristics and economic performance.