HFT High frequency trading, dangers

In June 1998, I wrote an article on IT and the Future of the City,
http://futurizon.com/idp/future/thecity.htm. I'm quite pleased now as I read an article in today's Times that says pretty much the same, albeit 13 years later.

Fast computing and geographic location

With a factor of 1000 in computer speed and memory capacity per decade, in parallel with advances in software, computers can now make logical deductions from the flood of information on the internet, not just from Reuters or Bloomberg, but from anywhere. They can often assess the quality and integrity of the data, correlate it with other data, run models, and infer likely other events and make buy or sell recommendations. Much dealing can now be done automatically subject to human-imposed restrictions, and the speed and quality of this dealing far exceeds current capability. The technology is now called High Frequency Trading, or HFT. Very predictable back in the 90s, and it has arrived on cue.

It will bring problems…

Firstly, the speed of light is fast but finite. With these huge processing speeds, computers can make decisions within microseconds of receiving information. Differences in distance from the information source become increasingly important. Being just 200m closer to the Bank of England makes one microsecond difference to the time of arrival of information on interest rates, the information, insignificant to a human, but of sufficient duration for a fast computer to but or sell before competitors even receive the information. As speeds increase further over following years, the significant distance drops. This effect would cause great unfairness according to geographic proximity to important sources, but there are two solutions. Either there is a strong premium on being closest, with rises in property values nearby to key sources, or network operators could provide guaranteed simultaneous delivery of information, which I believe is now the case. However, the simultaneous delivery only applies to a small number of companies and in a restricted area. It does not apply to the large number of individuals or companies situated a long distance away.

Secondly, exactly simultaneous processing will cause problems. If many requests for transactions arrive at exactly the same moment, computers or networks have to give priority in some way. This is bound to be a source of contention. Also, simultaneous events can often cause malfunctions, as has been demonstrated perfectly by numerous system crashes across many computer systems and networks. Information waves caused by significant events are a network phenomenon that could potentially crash networks.

An interesting future side effect of this is that the predicted flood of people into the countryside may be averted. Even though people can work from anywhere, their computers have to be geographically very close to the information centres, i.e. the City. Automated dealing has to live in the city, human based dealing can work from anywhere. If people and machines work together, they must both work in the City.

Consumer share dealing and software

Ultra-powerful palmtop computers pick up and analyse information all day long, organising every aspect of their owners' lives. With finance applications able to show which shares are doing well, spot trends and act on their computer’s advice at a button push, markets will grow for tools to profit from shares, whether they be dealing software, advice services or visualisation software.

However, as we see more people buying personal access to share dealing and software to determine best buys, or even to automatically buy or sell on certain clues, we will see some very negative behaviours. Firstly, traffic will be highly correlated if personal computers can all act on the same information at the same time. We will see information waves, and also enormous swings in share prices. Most private individuals will suffer because of this, while institutions and individuals with better software will benefit. This is because prices will rise and fall simply because of the correlated activity of the automated software and not because of any real effects related to the shares themselves. Institutions may have to limit private share transactions to control this problem, but can also make a lot of money from modelling the private software and thus determining in advance what the recommendations and actions will be, capitalising enormously on the resultant share movements, and indeed even stimulating them. Of course, if the share-dealing public generally perceives this problem, the AI software will not take off so the problem will not arise. What is more likely is that such software will sell in limited quantities, causing the effects to be significant, but not destroying the markets.

A money making scam is thus apparent. A company need only write a piece of reasonably good AI share portfolio management software for it to capture a fraction of the available market. The company writing it will of course understand how it works and what the effects of a piece of information will be (which they will receive at the same time), and thus able to predict the buying or selling activity of the subscribers. If they were then to produce another service which makes recommendations, they would have even more notice of an effect and able to directly influence prices. They would then be in the position of the top market forecasters who know their advice will be self fulfilling. This is neither insider dealing nor fraud, and of course once the software captures a significant share, the quality of its advice would be very high, decoupling share performance from the real world. Only the last people to react would lose out, paying the most, or selling at least, as the price is restored to ‘correct’ by the stock exchange, and of course even this is predictable to a point. The fastest will profit most.

The most significant factor in this is the proportion of share dealing influenced by that company's software. The problem is that software markets tend to be dominated by just two or three companies, and the nature of this type of software is that their is strong positive reinforcement for the company with the biggest influence, which could quickly lead to a virtual monopoly. Also, it really doesn’t matter whether the software is on the visualisation tools or AI side. Each can have predictability associated with it.

It is interesting to contemplate the effects this widespread automated dealing would have of the stock market. Black Monday is unlikely to happen again as a result of computer activity within the City, but it certainly looks like prices will occasionally become decoupled from actual value, and price swings will become more significant. Of course, much money can be made on predicting the swings or getting access to the software-critical information before someone else, so we may see a need for more equalised delivery services, with more suppliers covered. Without equalised delivery, assuming a continuum of time, those closest to the dealing point will be able to buy or sell quicker, and since the swings could be extremely rapid, this would be very important. Dealers would have to have price information immediately, and of course the finite speed of light does not permit this. If dealing time is quantified, i.e. share prices are updated at fixed intervals, the duration of the interval becomes all important, since it strongly affects the nature of the market, i.e. whether everyone in that interval pays the same or the first to act gains.

Also of interest is the possibility of agents acting on behalf of many people to negotiate amongst them to increase the price of a company’s shares, and then sell on a pre-negotiated time or signal.

Such automated systems would also be potentially vulnerable to false information from people or agents hoping to capitalise on their correlated behaviour.

HFT is a great idea in some ways, but it is not without costs. We can expect severe problems, with lots of opportunities for instability, crashes, wild swings, and abuse of customers.