To forecast sales there are many choices the person developing the model can make. But there is a relatively short list of categories of those choices.
The modeler can choose how accurate they want their forecast to be. This is called "model fidelity." An important clarification: the model builder does not actually get to say "I want a 99% accurate model." However, they do get to decide how complex a model to use as well as how much effort to put in to reducing uncertainty about the future. Said another way, the modeler can make qualitative judgements to work on a "more" or "less" accurate model.
The modeler can choose how much computational effort or data gathering effort to put into the model. This is called "model tractability." Again, the point is not that the modeler can decide to use deep learning but have the problem solve on a calculator. However, they can choose models that are easier to solve.
In forecasting specifically, another important question is the right level of aggregation for the forecast. For example the modeler may choose to forecast individual products at individual stores, or to forecast individual products aggregated for all stores and then dis-aggregate it.
Lastly, the modeler can choose many little things which impact the outputs of the model and any bias the model may have. For example, specifically for forecasting, we often need to choose how we want to handle lost sales (times we were sold out and so we sold none). If we choose to ignore lost sales, our forecast will be lower than if we try to fill in those gaps. However, in some cases those lost sales just wait until you are back in stock, in which case they would be double counted.
And so to solve the meta-problem we can try a set of tradeoffs and see what the optimal forecast model is for those tradeoffs. Depending on the result, we can update the tradeoffs in the interest of finding a better solution to the meta-problem.