Last month I wrote about the challenges of regulating machine learning, and today I want to explore whether AI will actually highlight the different market structure between equities and FX – something that is a long running theme of this column.

The more I talk to people about the potential for AI the more I see it bifurcating along market structure lines – on one hand it can certainly be used for analysing mountains of data to discern future market moves, on the other it will be used to gain the smallest edge on a repetitive basis.

I would point out that either path involves a huge amount of data processing and therefore firms’ ability to handle that mountain of data will be a differentiator.

I must confess my thinking has shifted a little from my early attempts to analyse the impact of AI – although I still believe it will end up in more crowded trades than ever before, there is more room for nuance than I first argued – not least in temporal terms.

Advocates of AI stress how it can spot correlations and factors that humans may not have considered and to a degree this is right, however this only goes so far and, much as is the case in the human world, will experience ups and downs – it will not be infallible. I like to think of correlations as the carry trade; they work really well right up until the moment when they don’t – and it is at that point that the world falls apart as everyone tries to get through the same very small exit!

To me, the power of AI lies in its ability to handle ridiculous (to a human), amounts of data and it is here that its use may diverge along market structure lines. Equity markets are made up of millions of individual trading instruments that are traded relatively infrequently (some obviously are not), whereas FX is made up of relatively few instruments that are traded extremely frequently.

A great use of AI is in analysing data that is not necessarily available elsewhere – a popular explanation for this is studying satellite data from JC Penney car parks; the fuller they are during the quarter the more likely it is that sales will be higher. Another could be studying local weather events in the agricultural belts of the world – yes it might be raining on the coast of Australia, but is it raining 40km inland where much of the agriculture exists?

There is a huge amount of data to consume and analyse and the more you study the more opportunity there is for you to gain an edge in a slightly longer term trade. It has to be longer term because this information will eventually seep into the market, so the AI gives you the edge of quicker analysis and decision making – price reflects information and the person that waits for all the information is the ultimate top or bottom picker.

So in equities there is always somewhere that AI can make a difference and I suspect that is why there is more excitement (as far as I can see) in equities circles than there is in FX. For in the latter it is harder to come up with how AI can help you make a better long term decision.

Many of the inputs to the exchange rate are static – economic data, political events and the odd unexpected event (which AI is unlikely to pick up upon). AI can analyse data quicker, but all that really results in is a short term trade – which is normally reliant upon there being a slower player pricing into the market. There are other inputs obviously, such as market sentiment and positioning levels, but these have been discernible for some time to humans, AI will not make a huge difference.

There is also the lack of absolute data to consider – we may know that a record number of longs have been established in euro FX futures for example, but we don’t know the same data from the OTC world and, more to the point, we don’t actually know the level of supply to the market.

The world could be long dollars, but if a corporation hedges unexpectedly early or late (AI could feasibly pick up the hedging requirement from elsewhere) and buys more dollars, the logic is thrown out. Yes, AI could tell you the market is long, but can it tell you when it will turn? That’s doubtful and as an example why I would like to use the great example of the hilariously titled “Mr Yen” back in the 1990s. Euseke Sakakibara spent six months predicting the market was too long of dollars as USDJPY rose from 125 to 142. Yes, eventually he was proven right, when the market collapsed (thanks to US intervention – not Japanese) from 142 to 118 in two days but I’m not sure many would stand by their AI programme faced with that sort of mark-to market!

For me then, AI in FX becomes about analysing very short term inputs to help the pricing and risk engines – it is not about picking the direction of currencies. Thus, FX remains about the speed of processing data, whereas equities can utilise it more strategically. There will, no doubt, be areas where AI can help predict exchange rate patterns based upon certain datasets and circumstances, but I remain unconvinced this will happen often enough to justify the potentially enormous outlay on data and computing power.

Of course there is one very obvious counter to my argument – technical analysis. AI has to make certain assumptions around patterns repeating themselves, much as technical analysis does, so if it gains enough traction it could, possibly, achieve that same self-fulfilling character that technical analysis has in some circles.

It would also, I need to point out, bring crowded trades back into the agenda – for as much as technical analysis in the hands of real experts helps them make money, too many sheep merely follow and create “important levels” in the market, which to traders of a certain vintage actually means “targets” – break those levels and you get everyone running for cover.

I would also argue that if AI does go down this path, it will create many more crowded trades than technical analysis ever has – and I am not sure that is healthy.

I don’t hold with those that believe humans will have no role to play in the future of trading, not only would I point to my examples above about creating crowded trades and everyone trying to get through the same small door, but there are always outliers that a human will spot quicker.

Put simply, I believe there is a need to overlay the vast amount of analysis produced by AI with a human randomiser. Such a tactic – I will call it “The Donald Factor” can certainly help averting AI from going off down a blind alley because data and sentiment says one thing, but the political reality will be totally different. It can also, if AI incorporates “The Donald Factor” (trademark is pending on this) help a firm dilute some of their exposure to what could be a crowded, but still profitable, trade.

It will be a fascinating journey with AI and without doubt it will play a huge role in financial markets going forward. If something can be done more efficiently then the tools that make it so should be utilised – even though that will itself raise issues of firms’ capacity for data.

Looking across markets, though, I can’t help escape the conclusion that while on one hand there may be lofty ideals around the use of AI in terms of micro-analysing data that humans just couldn’t or wouldn’t be able to, on the other it will remain about micro-managing pricing and risk – in other words, how do we nick a tenth of a tick out of a shed load of volume…?

Colin_lambert@profit-loss.com

Twitter @lamboPnL

Twitter @Profit_and_Loss

Colin Lambert

Share This

Share on facebook
Facebook
Share on google
Google+
Share on twitter
Twitter
Share on linkedin
LinkedIn
Share on reddit
Reddit

Related Posts in