Neural networks in finance: an information analysis

We classify financial applications of neural networks into two broad classes by stability and signal-to-noise ratio. We present two statistical measures typically applied to investment analysis: the information ratio (IR) and the information coefficient (IC); then we use Monte-Carlo simulations to critically examine neural net performance as a function of signal-to-noise ratio in characteristic investment domains. We thus measure the maximum noise level tolerable by neural nets during training on a representative class of investment problems