Adding measures to limit the influence bots have on market raises the question of how you know whether a trade was prepared by a bot or by a person. Without this distinction, you can't filter out bots in the fast-paced transactional world. But on the other hand, if you add rules (e.g. maximum rate of change) across all transactions, you are subjecting natural people traders to the same rules, and thus making the market prices more stable and predictable by guaranteeing certain conditions. It changes the rules of gameplay, and more to the favor of AI as emulating limits is easier for a bot to analyze than a person. Is that a good thing? Probably not.
To draw the analogy: You don't even out the Starcraft battlefield by putting limits on what each player can kill (i.e. adding a maximum kill rate); you even it out by adjusting the abilities that the units have (in network terms, adjusting the weights). Likewise, to tackle the black swan financial problem, we shouldn't be looking at policy that can be applied across the board, but rather at adjusting these 'unit abilities' – e.g. minimum trade 'reload' interval before firing for each account. These weights can even be dynamic / random, just to baffle AIs even more.
I'd also expect humans to adapt to 'learn' AI influence. Gamers do by practicing against computer AIs, so why not in finance too?
Reshared post from +Jon Mallin
Wired.com on High Frequency Trading
Nanosecond Trading Could Make Markets Go Haywire
The afternoon of May 6, 2010 was among the strangest in economic history. Starting at 2:42 p.m. EDT, the Dow Jones stock index fell 600 poin…
This post has been reshared 1 times on Google+
View this post on Google+
+Sophie Wrobel Well of course it'd be easy to account for. The point is only that it'd screw up the profit model of ultrafast trading, just like eliminating bulk registration discounts would screw up the profit model of domain squatters. 😉
+Sai I was thinking about that too – I believe you had proposed this some days ago. 🙂
The problem is that a small transactional cost is something that AI can relatively easily account for – once the AIs are done retraining on each other, I'd suspect that we'd only end up with larger black swans as prediction models scale up price changes to continue nanotrading with the transactional cost as part of the high/low trendline fluctuation weight.
+Sophie Wrobel What about the simple fix: a small, fixed, per transaction cost?