I'm trying to produce time-series forecasts from a set of numerical predictors.

The resources I've found describing the use of neural nets for time series prediction are autocorrelation-focused. They assume that the same trends, seasonality, etc. will continue into the future, and the predicted forecasts reflect this.

In my case, the forecasts need to reflect economic changes that haven't been made yet (implementation of carbon taxes). I'm hoping to use a more abstract blackbox approach to address this problem:

I have "ground truth" forecasts, produced through qualitative research, and want to fit my predictor variables to this data. That is, I want my trained net to act as a nonlinear function, somewhat-accurately mapping the predictor variables to the ground truth forecasts (which may have one or more abrupt local changes). Assuming the model converges, I could then extrapolate the net to new data, producing forecasts based on this blackbox function.

Does anyone have experience using a similar approach? I'm new to working with time-series, coming from image analysis where more straightforward CNNs were king, and sequentiality of data could be ignored. If someone could point me in the right direction I'd really appreciate it!


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.