Prediction of Co2 Future Prices for Energy Risk Management via Neural Network Adapted Stochastic Processes

Abstract Energy risk management requires an efficient prediction of CO2 future prices. We apply a neural network calibrated stochastic model to predict day ahead future prices suitable for risk management with a given time horizon (e.g. 10 business days). The model is training neural perceptron layers to learn the weights of historical time series points within a past horizon equal to the the memory depth (e.g. previous 10 business days). With this weights, the parameters of the stochastic model are optimized, and the day ahead carbon future price is computed. Iterating these steps, we obtain price forecasts, which demonstrate the learning efficiency of our method, which essentially corresponds to a stochastic process with time-dependent parameters, the dynamics of the parameters being themselves learned continuously by the neural network. The back propagation in training the previous weights is limited by the memory depth. The latter is the analogue of the maximal time lag of an autoregressive processes.