We are not dealing with a serious thinker or student of macro and predictive theories. All one has to do to reveal the flaw in predicting the future is read The Black Swan. As for accuracy in predictions, if a model was created which could predict the future accurately and reliably, it would influence the future for those people who had access to this model. By knowing the future, a person today could make decisions that predict it for their own benefit thereby influencing the future. All one can do is understand the past, prevent known flaws in the system from artificially creating undue risk in the future and hope that none of us lives through another Black Swan event in our investing lifetimes.
Taleb's criticism of forecasting is congruent in the context of financial markets, but it is incredibly useless with regards to many macroeconomic variables. We actually can produce rather impressive forecasts for macroeconomic aggregates such as employment and industrial production, just as we can do rather well with unemployment, housing starts, and inflation. Some of my own research suggests that estimating a nonlinear variant of a latent factor model using either kernel ridge regression or random forests is extremely good, especially at longer horizons (9, 12 and 24 months). We also found out that these nonlinear models have their biggest edge over simpler methods during times of rapid nonfinancial deleveraging, housing booms, tight credit conditions, and low-interest-rate environments. The only real annoying part with these methods is that you know they work well, but you cannot interpret their output. It's the number one reason machine learning is not more heavily used by central banks, for example. It's also part of the value-added by our paper.
The key point here is that if I try to evaluate how well forecasting models perform, it's incredibly hard to outperform some variants of a random walk for prices in equity markets. However, if you take a very long horizon, you'll beat that model in the currency market using structural macroeconomic models. That works across many currency pairs and over many time periods. Clearly, there is a problem with Taleb's comment: the problem is that Taleb is an options trader. When you look at data he often used in his work, the comment seems legitimate empirically and it also makes sense economically because people trying to capitalize on arbitrage would presumably eat away at most easily spotted regularities, insofar as things like transaction costs and liquidity issues do not forbid it. It also happens to be the case that our theories do not work well with financial data, save for the informal insight of the absence of arbitrage.
Lastly, a subtle point may be made with regards to knowledge of the underlying model of reality: your comment is almost always false. In economic models, agents are assumed to know the true model of the economy. This true model accounts for this knowledge, as well as their behavior. They are also assumed to take advantage of all the opportunities they face, in different ways and according to their preferences. Those things do not lead the model to become false. The parable you have in mind, however, makes sense in a specific context: financial markets and the existence of arbitrage.
If people try to take advantage of their knowledge and they know what is going on, how could an arbitrage opportunity exist? If a model could predict asset prices, its widespread use would invalidate it, as you point out, because people would presumably push prices in the opposite direction the model predicted by trying to take advantage of them. This is why most economic models enforce the efficient market hypothesis: mispricing can only exist if something prevents people from killing them by using them for profit. The parable, in other words, explains the efficient market hypothesis of Eugene Fama by telling you something doesn't make sense in that context. However, how do you profit from knowing the unemployment rate is likely to be close to 0.04 on average in the next 3 months? It's unclear how knowing that would push the unemployment rate far from that target value.
