This paper considers forecasting in a big data environment. We develop a category of nonlinear forecasting based on factor models to benefit from many potential predictors while accounting for any possible nonlinear dynamics within the environment. The problem of forecasting with factor models is a two-step procedure. The proposed model, at the first step, employs an autoassociative neural network to estimate nonlinear factors from a large panel of predictors, and at the second step, applies a nonlinear function on the estimated factors to predict a single time series. Such features can go beyond the covariance structure analysis and enhance the accuracy of forecasting. Applying this approach to forecast equity returns, the proposed model captures the nonlinear dynamic between equities to enhance the performance of the subsequent forecast. This offers a significant improvement to current univariate and multivariate models. We emphasize the fact that linear models can be seen as a special case of the proposed nonlinear model, which basically implies that in the event that nonlinearity is absent between series, the model will subsequently be reduced to a linear model. The empirical results on daily returns of equities on the S&P 500 index from 2005 – 2014 proved the superiority of the out-of-sample forecasting ability of this model vis-`a-vis competing approaches.