It's not like in the pre-HFT days market makers would just obligingly fill institutions for whatever volume trades they wanted to execute, regardless of how ham-handedly they broadcast their intent into the market. Any estimate of "costs" to institutions from HFT needs to be measured not against some Platonic ideal scenario where market makers earn no profits, but rather against the old specialist system that it replaced. In which case I suspect the costs would come out quite a bit on the negative side of $0.
There are some practices in HFT like paying for privileged newswire access that could use some reform, but the best solution to this fragmentation issue is for large institutions to stop sending out their orders like it's the year 1995.
In an earlier thread someone used the term "legalized front running" to describe HFT, and in thread the final pieces fell into place.
Fact 1: we know that HFT companies have special hardware in place. They have effectively built FPGA's (or even ASICs) that can process incoming orders while the message is still partially on the wire. This allows them to react to new orders and executions with a latency that is below the transmission delay of the protocol packet.
Fact 2: HFT companies have built directional radio links to get their messages faster between exchanges. Speed of light is still a limiting factor, and the distance traveled between exchange A to exchange B is somewhat longer along the fiber.
Fact 3: The basic rule of trading is "buy low, sell high".
So, if you can process order information faster than than it takes to receive the entire exchange message, you can use that as an advantage against the rest of the market. The HFT systems know the current price for a stock in the system (the lowest standing price), and when they receive an execution, they know that someone else just traded at that price. They can send out a message to other exchanges to quickly buy the said stock at the said price and put out offers at a slightly higher price.
The net effect? "I just bought all of the stocks under your nose. Here's what I'm willing to sell them for you."
As long as the price they're selling the stock out is within the limit, they'll get an immediate risk-free trade. The HFT systems already had a buyer ready before making their orders.
[1]: http://ra.ziti.uni-heidelberg.de/cms/images/pdfs/High-freque...
I have no doubt about that. Very few things are new.
But if you're deploying special hardware in an attempt to cheat the speed of light and gain an advantage of mere microseconds, the system is FUBAR (or recognition).
mutual fund X places two market orders to buy and take out the current offers, one on A and one on B. Its order on A executes at $100, taking out the liquidity and making the new best offer there $100.01
HFT firm Y sees this execution and cancels their offer of $100 on exchange B and replaces it with one at $100.01 all before the order that was placed on B arrives.
This is a market fragmentation issue, Reg NMS tries to address this but doesn't really work at these timescales, because you only have to comply based on the current quotes you are getting from the exchanges, which could be many milliseconds behind (and if your system is slow to process the quotes and makes more money as a result, you won't be getting a task from your boss to speed it up).
http://en.wikipedia.org/wiki/Regulation_NMS
There are many other types of useless latency arbitrage that don't depend on market fragmentation, like trading on the difference in price between ETFs (basically a basket of stocks usually trying to match something like the makeup of the entire S&P) and the underlying individual stocks. Each time the underlyings change, the ETF price changes (within reason, risk hedging activity etc. could make the ETF temporarily get out of sync, but that is just another latency arbitrage oppurtunity). Whoever is faster wins.
It's pointless, there are datacenters full of servers running wait loops on incoming data via DMA so that they don't have to pay the latency price of a system call and context switch to the kernel. Basically burning up part of that mutual fund's money into waste heat.
As a buyer of an ETF or as a market participant looking for trading venues I needn't spend lots of resources ensuring that the prices are in sync. Those costs have been born by the specialists in those arbitrage opportunities and it is a very cut throat efficient system.
There is some value to keeping things liquid, but that value isn't tied at all to the amount of money that arbitragers get compensated. Consider that whether the latency is 1ms or 5ms, the fastest arbitrager wins. But to bring things down to 1ms from 5ms may require 10X more expense and waste, lowering the compensation to the arbitrager. In a competitive enough environment they simply burn up all of the spreads into waste heat and redundant fiber infrastructure.
It's risk-free arbitrage...exploiting the difference in price between two exchanges that will exist for only a fraction of a second knowing that the rest of the order will arrive fractions of a second after they trade.