Edit: If I may expand on the point above. Which may seem cryptic.
Complex systems are interisting in that they can be both predictable (in theory) and not predictable (in practice). What's more is that they can be mischaracterized by analyzing the data. Ie, this data is a mess, it must me unpredictable. There are a whole class of deterministic processes which generate these types of results. Whereas a normal, linear process can be infered from medium/high resolution data, even with higher-resolution data we can't infer the underlying logic of the complex system. We may, as a result, either oversimplify the model or proclaim the data to not support any deterministic process at all. The classic example is the class of processes that exhibit sensitive dependance upon initial conditions. ie, variations on the notion of deterministic Chaos. Their sesitivity is such that the resolution of the dataset required to deduce or infer the origination function would never be feasible if it was anything other than complete. Whereas, with deterministic processes that are more traditionally tractable, you can make progress in your knowledge with data-sets of increasing resolution. ie, you can run a regression to infer y=mx+b or a monte-carlo to fit a gaussian curve or what not. But you cannot "brute" force a fit to a choatic process from a montecarlo, becaue you will never have enough resolution nor enough precision in your data set to infer an origination function. [1]
The summary thought is tha sometimes gaps in data can be bridged with higher-level logic or heristic, but this is not always possible (either in theory or practice). Yet, we should not infer a problem is unsolvable or untractable just because of this. =D
[1] http://en.wikipedia.org/wiki/Chaos_theory#Sensitivity_to_ini...