Research on ground truth/back testing

Looking for links, papers, thoughts on how ground truth fits into model development for reporting current periods or forecasting. Applications to financial markets/hedge funds are the most 'obvious,' but need not be constrained.

For example, if one wants to estimate earnings at Best Buy (a la 4square), foot traffic data from their SDK doesn't represent 100% of the visitors, so to better train models it is important to know what percentage of the population they are sampling. Models need to be calibrated against a ground truth, an indisputable source, and this is likely doesn't come from the retailer. In short, having ~24 months of historical and ground truth data can establish sufficient confidence to make forecasts/estimates.

Another example is estimating storm surge/wind patterns. 30 years of historical/ground truth is required for insurance purposes (not legally required, but establishes enough statistical confidence).

I'm looking for papers, thoughts, directions, etc.... from various industries/application areas to understand the relationship between historical data, ground truth and accuracy/confidence.