It's worth noting that prediction algorithms in the Machine Learning literature, if stated formally, usually come with the assumption that the data points are sampled i.i.d. from some distribution. This distribution is badly violated when the predictions are used to take actions in the real world that affect future data.
For example, one might observe an arbitrage in currency prices, then trade on that opportunity, which then removes the opportunity for arbitrage. So the model will perform well, then the distribution will change, and it will stop performing well.
Note that this doesn't apply to say image classification or speech recognition. Predicting that an image is of a cat won't change whether the image is or is not a cat. This is probably a big part of why Machine Learning is so successful here.
On the other hand, there are lots of domains which exhibit the i.i.d. breaking feedback property. Recommendations are one such example. After Netflix recommends a movie to customers, they become more likely to see that movie, which changes the distribution of the observed data.