[leogau]

"Even if each individual algorithm makes perfect sense, collectively they obey an emergent logic—artificial intelligence, but not artificial human intelligence. It is, simply, alien, operating at the natural scale of silicon, not neurons and synapses. We may be able to slow it down, but we can never contain, control, or comprehend it"

I can see algorithms being used to automate decision making in almost every other industry, not just Wall St. It's both scary and exciting to see what happens with the resulting "emergent intelligence".

[jacques_chester]

The systems dynamics folks point out that systems with more rapid feedback often have more variability.

The classic example is a car dealership deciding to order cars from the factory. When should they be ordered, and how many should be ordered?

Suppose they decide that every day they place orders based on the 7 day moving average. Then they get a busy day and over-order for a week. That's an expensive mistake.

Next they decide to shrink the window to 3 days (more rapid feedback). Now after another busy day they find stocks dropping precipitously and order big to make up, follow by drops, followed by big orders. In an attempt to stabilise the system, they've made it worse.

The modern stock market has extremely short feedback times, on the order of microseconds. When a feedback loop forms on the market it can spiral out of control within seconds and minutes -- hence the 'flash crash'.

Two ways to deal with this might be to develop some kind of balancing feedback loop (for example, higher prices for more trades-per-second, or a progressive price for trade based on delta with the last trade) or to reduce the feedback rate. An economics blogger I host, Nicholas Gruen (you may know of him as Australia's Gov 2.0 inquiry leader), has suggested just that:

http://clubtroppo.com.au/2010/08/07/a-modest-proposal-to-rem...

http://clubtroppo.com.au/2010/10/08/a-self-denying-ordinance...

[yummyfajitas]

The systems dynamics folks point out that systems with more rapid feedback often have more variability.

You've got systems dynamics folks completely backwards. Smaller, more frequent controls tend to stabilized systems, not make them unstable. A simple example we've all seen: take a stable ODE. Now try to discretize it - if you are unlucky or uncareful, your discrete approximation can easily blow up exponentially.

The example you provide is different - you are describing two different control strategies, one of which fails to correct for noise (and note: any HFT who makes this mistake loses money FAST). If you made orders every day based on the 7-day moving average, it would be better than making orders once per week based on the 7 day moving average.

This is exactly what we saw with the flash crash - there was a large exogenous shock and the system self-corrected within minutes.

[jacques_chester]

Good followup, thanks.

[joshu]

Or to do more orderly crossings at regular intervals.

The market makers (and shadow market makers) on the ECNs get paid for providing liquidity, rather than having to pay.

This kind of stuff isn't supposed to happen on the NYSE, as all the trading goes through a single specialist.