[kvcc01]

Yes, there’s a "frequency spectrum" of sorts and HFT is at the short end of it. Here, your models monitor and react to trading tick-by-tick, and execution speed is paramount. It’s also where you have the fiercest arms race for the fastest systems, colocation, FPGAs, etc. Next up comes what’s often called “statistical arbitrage,” where you have models that no longer look at tick-by-tick trading but may look at what happens at 30-second, 1-minute, or 5-minute windows. Here you have more interesting relations emerge between stocks and the market, e.g., what does IBM do relative to the tech index, or relative to MSFT, etc. (Such cross sectional relationships don’t seem to matter as much in HFT.) Actually stat-arb was the domain where the earliest statistical approaches such as "pairs trading" emerged. Next up come models that trade daily (or less frequently), and here you begin to see the long-short market-neutral relative-value type of approaches, where some quantitative [mutual] funds may operate. Next up will be the traditional mutual funds, and beyond that you have your Warren Buffett’s, etc.

One thing to keep in mind is that the higher your trading frequency, the smaller the price moves you can hope to capture, which limits how much capital you can deploy in your models. This is why HFT models are usually small in size but have high Sharpe ratios. As you reduce your trading frequency, you can expect to capture larger price movements and deploy more capital but you’ll also be exposed to more of the vicissitudes of the general market, so your Sharpe ratio will decline. Market participants usually carve themselves a happy spot on this frequency spectrum and stay there. I don’t know of any firm who is successful at every spot.

[mrchicity]

I think you are misconstruing holding period or predictive horizon with latency sensitivity. Many HFTs are looking at statistical relationships like the ones you mention to compute a fair price for making markets. The only trades where HFTs hold positions sub-second on average are pure arbitrages. Like you mention, there simply isn't enough price movement within that timeframe to generate a profit.

All the inputs to their pricing change rapidly, so their order prices must change quickly as well, but they can end up carrying risk for long periods of time. The Australian regulator looked at HFT activity in their markets, mind you probably less sophisticated than US stocks, and found the average holding period was 42 minutes: http://tabbforum.com/opinions/hft-concerns-are-overstated

[mathgenius]

I'm not sure the average holding period is a useful thing to measure. For eXample, if you are trading a spread it's how quickly you put on the second leg that matters, and who cares how long you hold the pair for. So, 42 minutes has nothing to do with it.