Artificial Neural Nets for Prediction with Python (pybrain)
For my employer, I’m using neural nets to predict utility bills and optimally control a system that minimizes that bill by clipping the load. Utility bills are usually a highly nonlinear function of the peak power consumption for the day. So our system should save a lot of money and help save the planet by allowing utilities to run more efficiently and emit less greenhouse gas.
One way to get deeper nets to train on a highly-stochastic datasets (like ticker historical prices) is to “restart” the backpropagation training after it converges (or even if it doesn’t). In pybrain this seems to use a different portion of your dataset for the training and convergence-testing portions of the learning. This helps pybrain wander out of local minima and find smaller minima elsewhere in the terrain.