I'm trying to train a model to predict the final cost of a product being developed over a few months. I have historical data of similar products which will be used for training the model.
Some of the features include %planned, %developed, %purchase orders raised, pending purchase orders, material cost thus far, time period (sampled daily), overhead costs, total cost.
The aim is to predict the final cost of the project given the current status. I'm familiar with LSTMs and know that this is a multivariate-multistep prediction.
However the question remains,
the project end date is not fixed, so how do we select a window in this scenario. Some projects may last a month, others last several months.
now since the end date is not fixed we don't know mow far in the future we have to predict. So number of steps to be predicted changes as well. I figured I'd build a regression model to predict the project completion date according to the current project progress and then use that to predict the future time steps.
Also every LSTM model I've worked on and seen so far have a fixed window for prediction (predict the next 7 time steps). In this situation assuming that I use point #2 to define how many steps to predict in the future the number changes. Initially predict first 30 time steps, as we progress, predict next 12 time steps, etc. , how do I go about building a model for the same?
One approach I could think of is train the model to predict just one time step and loop over the current prediction for as many time steps as required. Now I do understand that this will decrease the accuracy tremendously if we predict too many time steps based on current predictions, (compounding error).
Let me know if any clarification is required in the comments. Thanks in advance.