The Financial & Risk business of Thomson Reuters is now Refinitiv

All names and marks owned by Thomson Reuters, including "Thomson", "Reuters" and the Kinesis logo are used under license from Thomson Reuters and its affiliated companies.

March 28, 2018

The Market Sentimentalist – Artificial Stupidity

by Amareos.

AI is very in.

If one includes the phrases Artificial Intelligence and blockchain together with a smattering of machine learning or (cherry on the top) deep learning in any capital raising spiel then one is almost guaranteed to have people throwing cash at you.

Appetite for such products in the data rich world of finance is especially high. Independent Fintech companies – like Amareos – that could not have existed a decade ago are being created to tap into this technological innovation; almost every major commercial bank is exploring opportunities in this space as evidenced by the proliferation of Fintech hubs and incubator programmes; and, asset managers are facing increasing competition from Roboadvisors, or as they preferred to be called, digital investors.

With all of this focus on technology and the advance of machine learning, it is not surprising that no one talks about artificial stupidity. In fact, the only passing reference we came across was in a recent episode of Amazon’s TV series Mozart in the Jungle after one of the lead characters encounters a robotic orchestra conductor/composer.

Perhaps, this is a mistake.

Often with the emergence of a new technology there tends to be a considerable amount of hype generated. Indeed, we are sure many of you have come across some version of the Gartner Hype Cycle for emerging technologies – reproduced below. (Anecdotally, the Hype Cycle itself seems to be experiencing a self-validating popularity boom.)

Exhibit 1: Gartner Hype Cycle

Source: Jeremy Kemp at English Wikipedia

If one were to change the y-axis variable from visibility to an asset price, its profile looks a lot like a speculative bubble. To demonstrate this, take a look at the exhibit below which shows the evolution of the NASDAQ stock price index in the pre/post dotcom bubble period.

Exhibit 2: NASDAQ During Dotcom Bubble


Although we have chosen the technology heavy NASDAQ index during the dotcom bubble period to illustrate the similarity we need not have. Technology is not the common factor driving these cycles rather it is human psychology and how we, as people, collectively respond to certain stimuli.

The surge in the hype cycle and the asset price bubble occurs because increasing numbers of people become convinced about the merits of the innovation/asset in question. Given the absence of precedent, there is considerable uncertainty as to what the future will bring. Opacity about the future is usually a negative thing. However, in this case, it serves as a propagation mechanism. If no one can be sure of anything, then everything is possible, even future expectations that in almost every other situation would be considered preposterous. After all, consider the spread of valuations for the world’s first crypto-currency, Bitcoin, which ranges from zero to more than a million US dollars[1].

What’s more, strong incentives are created encouraging individuals, companies and/or organisations to increase their exposure to whatever is “hot”. Pet companies add dotcom to their corporate title and valuations jump or, as has occurred more recently, old-school business models adopt blockchain technologies, ICO and raise tens of millions of dollars in a matter of hours. Such developments only add to the frenzy.

(Returning to Bitcoin for a second, we highlighted the extremely frothy nature of crowd sentiment towards this cryptocurrency in late December – see footnote 1 below and an updated sentiment chart below – days prior to a significant downward price correction).

Exhibit 3: Crowd-Sourced Sentiment – Bitcoin

Source: and

Some of these ventures succeed and turn into highly profitable companies (Amazon, Apple and Alphabet – the equity market version of AAA assets – being a good example). However, given the strength of the aforementioned incentives, inevitably there is a lot of BS created as well – a less polite synonym for hype.

Picking the winners during the boom is not easy. It is only when the economic equivalent of gravity kicks in and reality fails to live up to the earlier dreams that weakness in the business models of some of the emergent technology companies are revealed. Euphoria is replaced first by fear, then panic and ends in depression – see exhibit below.

Exhibit 4: Stylized Investment Psychology Cycle


AI will not be immune from such effects. Indeed, AI has experienced numerous booms and busts in the 70 years since computers were first created and the concept of machines matching human intelligence was born[2]. This slow progress is testimony to the difficulty of the task at hand. Nevertheless, it is clear that progress is being made and it is this progress that is causing the current resurgence of interest in AI.

Deep Learning architectures and algorithms have made impressive advances in fields such as image recognition and speech processing. Their application to Natural Language Processing was less impressive at first, but has now proven to make significant contributions, yielding state-of-the-art results for some common NLP tasks. Sentiment analysis, for example, is one of the problems where neural network models have outperformed traditional approaches. Perhaps most remarkable of all is the progress in machine translation.

For most people though, evidence of progress comes from the likes of Siri, Alexa and Google Assistant which operate as smart home personal assistants and the chatbots that are becoming the first line in customer services for increasing numbers of companies. The imminent arrival of electronic self-driving cars is another popular technology meme.

In finance, algorithmic trading has increased hugely in volume and via the application of machine learning techniques, these algorithms are increasing being refined via computers rather than humans. As one company operating in this space states in their literature (we will refrain from naming them), this will have a profound impact on the nature of trading.

“The impact of human emotions on trading decisions is often the greatest hindrance to outperformance. Algorithms and computers make decisions and execute trades faster than any human can, and do so free from the influence of emotions.”

We fully agree with the first statement. Human emotions are a key element in the investment process, perhaps the single greatest element. As Oaktree’s Howard Marks stated in one of his famous client memos[3],

“If I could know only one thing about an investment I’m contemplating, it might be how much optimism is embodied in the price (our emphasis).”

In fact, this is the core belief of Amareos which is why we track and analyse crowd-sourced sentiment indicators extracted from millions of finance articles posted online every day. We also agree that emotions can be an impediment to making good investment decisions. The “fight-or-flight” response triggered by fear is a sensible strategy when being chased by a lion on the savannah, but is not always the best response when a stock price drops unexpectedly.What’s more, it is abundantly obvious from the existence of High Frequency traders that machines can make and execute investment decisions faster than humans.

Where we fundamentally disagree, however, is in relation to the last statement that algorithmic trading is emotion-free. This is patently false.

For the time being[4], the owners of capital are humans and we are subject to all the psychological flaws that evolution has endowed us with (see previous lion comment). Machines may not feel financial pain, but humans do. This basic fact influences the design of algorithms. Cathy O’ Neil, a data scientist and author of “Weapons of Math Destruction”[5] wrote the following:

“Algorithms are not objective – the people who build them impose their own agenda on the algorithms.”

Moreover, this statement holds even with the advent of machine learning algorithms where the search and selection of trading strategies is fully automated and hence subject to less human design intervention that has previously been the case.

By construction, trading algorithms are given a set of data inputs and are then set to look for patterns in the data in order to predict a given target variable, eg. the algorithm learns how to predict one day-ahead returns etc. The human influence may be less obvious with machine learning algos but it’s still there because the input variables and output target variables given to the algos as part of the training set are determined by its carbon-based creator – psychologically-flawed, emotional, humans.

As long as humans are in control of the bigger What and the Why questions, rather than the How questions which machines are currently making good progress in answering, then emotions will continue to have a profound impact upon investment behaviour and market trends.

There is an even deeper, and therefore subtler, reason why emotions will continue to have this profound effect and it all relates to rationality. Emotion-free investing by machines promises rational decision-making, but it is extremely unlikely that such an approach would generate superior investment returns.

To see why we first need a suitable definition of rationality in the context of investing. The best definition we could come up with is when the market price of an asset is converging with its fundamental valuation.

As the saying goes, a picture paints a thousand words, so we have the following exhibit to illustrate the point. The two lines show the fundamental estimate of the asset price in question and its market price. The blue time periods are when the market price is converging with the fundamental valuation – rational behaviour on this definition. The white periods, in contrast, correspond to periods of irrational behaviour, where market prices are diverging from its fundamental value.

Exhibit 5 – Market Prices versus Fundamental Valuation: An Illustration


Consider point A in the above diagram. The market price of the asset is above its fundamental value. A rational investor, or unemotional algo, would consider this a possible shorting opportunity – the difference in prices (P1 – P0) representing the expected profit from such a trade. However, because in this example we are dealing with market prices determined by the interactions of emotionally-flawed, irrational, human investors the price of the asset actually continues to rise until it peaks at P2.

An algo following a perfectly rational investment strategy – selling an asset whose price is fundamentally overvalued – would lose money. The optimal strategy is obviously to ignore fundamental value ie. act irrationally and hold a long position until the price is at P2, at which point the long position is closed and (P2 – P1) profit is in the bag.

The rational thing to do is… to be irrational.

What is an algo to do?

In short, investment success depends upon behaving like the rest of the crowd almost all of the time. Acting rational when everyone else is irrational is a losing trading strategy because market prices are determined by the collective interaction of all participants.

Successful algos, defined as being best able to predict their target variables based on the inputs must, therefore, act human-like. To borrow Milton Friedman’s famous  phrase to be successful machine learning algos must act “as if” they have emotions and are, at times, irrational[6].

Artificial stupidity!

For those who consider that this example is constructed in such a way as to generate this outcome bear in mind that fundamental valuation models are estimated such that the residuals of the model (ie. market price – fundamental valuation) sum to zero over the estimation period. The extent of overvaluation is the same as the extent of undervaluation – anything other than this results in a bias.

The only situation where this is not true is in the special case of strong efficient markets where the market price is considered to be always in line with the fundamental value. Thankfully, for anyone in active asset management, strong efficient markets are a long way from describing reality as shown by the repeated asset price bubbles and busts that pockmark financial history.

The above illustration clearly is based on a situation where the majority of investment decisions are made by carbon-based investors, but let’s extend this thought experiment by assuming that machine learning algos are able to successfully mimic the irrational behaviour of carbon-based investor. What’s more, let’s credit them with superior pattern identification methods such that they are able to correctly identify market turns ex ante enabling them to exit trades at just the right time. Minimal drawdowns in other words, resulting in impressive Sharpe ratios.

Over time such consistently strong performance would see increased capital allocations to these machine learning algos and their influence on determining market prices would increase. Eventually this influence would rise to such an extent that when the algos, anticipating a market turn, change their investment position it actually triggers the turn in price.

At this point, the smarter, more rational, decision for the algo (or just as likely a competitor algo) would be to pre-empt this move by changing position immediately prior to the market turn being anticipated. Such reasoning is akin to the “backward induction” problem we discussed in an earlier Market Insight, albeit in a different context[7].

The outcome of such rational behaviour would be a market where asset prices move in a manner consistent with the strong version of the efficient market hypothesis. That is to say, prices would jump/fall almost instantaneously in response to news events (stepwise moves), but flat-line the rest of the time. As these price jumps would be random (or, better put, unpredictable) then there would be no way to generate systematic alpha.

The algos would no longer be fit for purpose.

It would be an extinction event.

That really would be Artificial Stupidity.


[1] For our Bitcoin valuation framework and estimate – see:

[2] We picked this fact up from Nick Bostrom’s book Superintelligence – a great read for those interested in learning more about machine intelligence – see:

[3] See:

4] If, or when, this changes as the academic and author Yuval Noah Harari suggests in his recent book “Homo Deus” could happen – then this argument is no longer valid – see:

[5] See: We came across this quote in a blog post by Dave Trott[5], a creative director and author with a very unique prose style – see:

[6] Infamous, perhaps better put, because this framework has been a toxic cancer to academic economics in our opinion – see:

[7] See:

Article Topics

Get In Touch


We have updated our Privacy Statement. Before you continue, please read our new Privacy Statement and familiarize yourself with the terms.×