Does HFT Behaviour Amplify via Stigmergy?


Back in late June, I was discussing algorithmic trading and HFT (high-frequency trading) with some friends. A sudden realization hit me, and I realized exactly why I've had this nagging idea that wide-spread HFT usage is fundamentally a bad idea due to the weird market behaviour it most likely leads to.

Here's a cleaned up mail I sent, with some links added. What do you think?

There is a major problem with HFTs and fast algorithmic trading and markets due to full computerization (no human in the loop) and the massive speed obtained through those methods. This time, I don't mean the obvious application of HFTs doing front-running.

Here I use the definition "HFT agents" to be any instance of an HFT/algorithmic trading algorithm running somewhere. Basically a computer reading and writing to the market.

So: I see the dynamics such that HFT agents both observe and influence the market. Yet they are part of that market as well. As they trade, they also influence each other (although indirectly)! And they do trading at such speed, that there is bound to be kind of "resonances" amplifying whatever direction was taken originally - sort of like everyone nudging the same direction a little bit. I think that this kind of system must be unstable.

Human traders can't jump in and fix up the situation as they can't keep up with the speed. So, in essence we get either flash bubbles or flash crashes. All this because the HFT agents actions influence each other, they end up modifying each other's algorithms through application of stigmergy when they modify the market (=they modify the environment they're in and read their own and others modifications as input to their algorithms).

It's like the synchronizing of fireflies (see the video and read Strogatz's book "Sync" for details), they adapt to each other until everyone blinks in unison. Except here the synchrony is a kind of "stochastic resonance" leading to spikes + valleys.

With slower trading the quick peaks and valleys even out, of course we still get bubbles and crashes over a longer period. But the same bubble/crash-creating phenomenon is at play, this time much quicker.

One example follows. There are surely other ways also, given that there is a big diversity in the algorithms.

  • HFT agents x and y scans the market. Some condition c holds (e.g. break in 30 day SMA).
  • HFT agent x triggers first and does whatever (sell/buy orders etc.), changing the market conditions.
  • HFT agent y might do the same action immediately after x, reacting to condition c or action of x (stigmergy comes into play). This "boosts" the direction, be it up or down.
  • Now imagine thousands of these agents. Some might react to the amplified direction signal (e.g. drop in price) further amplifying the system.
  • Result: flash crash / flash bubble

As long as there would be an overall balance of HFT trading strategies such that some are bearish and some are bullish, we would not see nearly as much volatility. But as it is now, the balance is tilted and therefore I think insane volatility is inevitable.

Note: Sometime after the mail exchange, the Zero Hedge blog blogged about a paper from Reginald Smith examining similar ideas. It is a very interesting article that makes the paper more accessible.