In response to the this

HFT often works in a way that is agnostic to the market itself.

In my experience strategies do not work in a way that is agnostic to the market. Usually strategies are formulated off of some premise related to the market. One example would be if there are more orders resting on the buy side then the market might move up. Or if the price has been moving up keep moving it up until reach some reversion point that you have calibrated by backtesting. These two strategies here may seem like they are agnostic to the fundamental price information of the stock, but they are not agnostic to the market itself. As a market maker you want to facilitate trading between buyers and sellers and using information that is coming into the market (like resting orders or recent trades) allows you to get a better idea of what the market participants are going to do next.

Now I imagine that there are other things that traders could do that are way out there. Attempting to correlate twitter traffic to price fluctuations seems particularly hard to me. Similarly doing anything that is deep machine learning like Neural networks or Genetic Algorithms run a bigger risk in my opinion because if the premise that you trained your learning algorithm on changes then your strategy might work in the complete wrong way. As an analogy imagine that you trained a neural network facial recognition algorithm in a room with only yellow incandescent light and then you bring in a test corpus to see if your training has worked. It will probably work pretty well but what if the yellow light gets changed to fluorescent light. If it's a neural network it will be hard to know if it will continue working. Similarly in the market there might be a certain market sentiment all of 2013 but if something big happens in 2014 like a sovereign default or another externality you have no idea how the market is going to behave moving forward. If you at least tie your strategy to something simple like order flow you are likely to run less risk.