Feature Article: The Potential Impact of Artificial Intelligence Technologies on Regime Probabilities and Asset Class Returns

Today there is no shortage of forecasts about the potential impact of artificial intelligence and other automation technologies on future employment and employability. However, there are far fewer frameworks to help investors assess the current state of AI technologies, and even fewer attempts to discern their potential impact on future asset class returns, which are this article’s two goals.

A broad definition of automation is “the technique of making an apparatus, process, or system operate with minimum human input.” Robotics is a subset of automation that “deals with the design, construction, operation, and use of robots (machines that can replicate human actions) as well as computer systems for their control, sensory feedback, and information processing.”

There is more confusion over the meaning of “artificial intelligence.” The Oxford English Dictionary defines it as, “The theory and development of computer systems able to perform tasks normally requiring human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages.” Amazon’s definition is similar: “the field of computer science dedicated to solving cognitive problems commonly associated with human intelligence, such as learning, problem solving, and pattern recognition.”

Early approaches to artificial intelligence involved attempts to capture and codify expert knowledge in computer code via massive rule sets. The inability of this approach to effectively scale led to declining interest in the field during the 1980s and 1990s, a period that became known as “AI Winter.” Interest was reinvigorated with the invention and development of a series of methods collectively known as “machine learning” or “ML”.

We prefer the definition offered by Tom Mitchell, Dean of the School of Computer Science at Carnegie Mellon University: “the study of computer algorithms that allow computer programs to automatically improve through experience.”

This improvement is driven by the use of various “learning algorithms”, including the following:

“Supervised Learning” uses input “training data” that has been previously labeled to aid in the automated development of algorithms that can, with some minimum degree of accuracy, discriminate between different categories – e.g., photos of cats versus cars. A key issue is the cost of having human beings label large quantities of training data.

“Unsupervised Learning” uses unlabeled training data, usually to identify patterns that can be used for the purpose of prediction. To be sure, humans also do this – e.g., when creating correlation matrices or testing regression models for different variables, or using data visualization methods for exploratory data analysis. The key difference is that the use of machine learning methods like deep neural networks enables the discovery of far more complex patterns in a data set, which in turn can increase predictive accuracy.

“Reinforcement Learning” is an iterative approach to learning to make better decisions in a given set of circumstances. Again, this concept is not new; other methods like linear optimization and evolutionary search have been used for the same purpose. The difference with machine learning methods is the scale and complexity of the problems that can be addressed, and the speed at which acceptable solutions can be found. However, like older approaches, the effectiveness of reinforcement learning critically depends on the creative design of the so-called reward function (e.g., maximizing gains subject to one or more constraints), which can quickly become problematic when multiple conflicting goals are being pursued.

“Generative Adversarial Networks” or GANs (and closely related “Actor/Critic” methods) harness the power of competition between two deep learning networks to accelerate learning. In the case of GANS, the first network is the “generator”, which creates an output. The second network, the “discriminator”, tests that output against some criteria to produce its own output. Over multiple iterations, the discriminator seeks to maximize its results on the specified criteria, while the generator attempts to produce outputs that minimize those results. A classic example is how Google DeepMind used GANs to play millions of games of Go in the development of its “AlphaGo” software that eventually defeated Lee Sedol, the reigning human champion. Note that while GANs essentially create their own training data, they still face the challenge of developing effective reward functions.

While the results achieved thus far using machine learning methods have been very impressive (particularly compared to rule-based approaches to artificial intelligence), it is also important to recognize the limitations of current ML technologies (which get far less publicity). Professor Judea Pearl’s “hierarchy of reasoning” provides an excellent way to do this (for a more detailed discussion, see "The Book of Why").

Pearl divides reasoning into three increasingly difficult levels. The lowest level is what he calls “associative” or statistical reasoning, whose goal is finding relationships in a set of data that enable prediction. A simple example of this would be creation of a linear correlation matrix for 100 data series. Associative reasoning makes no causal claims (remember the old saying, “correlation does not mean causation”). Machine Learning’s achievements thus far have been based on various (and often very complex) types of associative reasoning.

And even at this level of reasoning, there are many circumstances in which machine learning methods struggle and often fail. First, if a data set has been generated by a random underlying process, then any patterns ML identifies in it will be spurious and unlikely to consistently produce accurate predictions (a mistake that human researchers also make…).

Second, if a data set has been generated by a so-called “non-stationary” process (i.e., one that is evolving over time), then the accuracy of predictions are likely to decline over time as the historical training data bears less and less resemblance to the data currently being generated by the system. And most of the systems that involve human beings – so-called complex adaptive systems – are constantly evolving. In contrast, even in the case of very complex games like Go, the underlying system is stationary – e.g., the size of the board, ruled governing allowable moves, etc. – do not evolve over time.

Of course, a predictive algorithm can be updated over time with new data; however, this raises two issues: (1) the cost of doing this, relative to the expected benefit, and (2) the respective rates at which the data generating process is evolving and the algorithm is being updated.

Third, machine learning methods can fail if a training data set is either mislabeled (in the case of supervised learning), or deliberately corrupted (a new area of cyberwarfare). For example, consider a set of training data that contains a small number of stop signs on which a small yellow square had been placed, linked to a “speed up” result. What will happen when an autonomous vehicle encounters a stop sign on which someone has placed a small square yellow sticker?

In Pearl’s reasoning hierarchy, the level above associative reasoning is causal reasoning. At this level you don’t just say, “result B is associated with A”, but rather you explain why “effect B has or will result from cause A.”

In simple, stationary mechanical systems governed by unchanging physical laws, causal reasoning is straightforward. When you add in feedback loops, it becomes more difficult. But in complex adaptive systems that include human beings, accurate causal reasoning is extremely challenging, to the point of apparent impossibility in some cases.

For example, consider the difficulty of reasoning causally about history. In trying to explain an observed effect, the historian has to consider situational factors (and their complex interactions), human decisions and actions (and how they are influenced by the availability of information and the effects of social interactions), and the impact of randomness (i.e., good and bad luck). The same challenges confront an intelligence analyst – or active investor - trying to attach probabilities to possible future outcomes that an evolving complex adaptive system could produce.

Today causal reasoning is the frontier of developing machine learning methods. It is extremely challenging for many reasons, including, for example, requirements for substantial improvements in natural language processing, knowledge integration, agent-based modeling of multi-level complex adaptive systems, automated inference of concepts, and transfer learning (applying concepts across domains).

Despite these obstacles, AI researchers are making progress in the area of causal reasoning (e.g., “Causal Generative Neural Networks”, by Goudet et al, “A Simple Neural Network Model for Relational Reasoning” by Santoro et al, and “Multimodal Storytelling via Generative Adversarial Imitation Learning” by Chen et al).

At the top of Pearl’s hierarchy sits counterfactual reasoning, which answers the questions, “What would have happened in the past if one or more causal factors had been different?” and “What will happen in the future if assumptions X, Y, and Z aren’t true?”

One of my favorite examples of counterfactual reasoning comes from the movie Patton, where he has been notified of increased German activity in the Ardennes forest, at the beginning of what would become the Battle of the Bulge. Patton says to his aide, “There's absolutely no reason for us to assume the Germans are mounting a major offensive. The weather is awful, their supplies are low, and the German army hasn't mounted a winter offensive since the time of Frederick the Great — therefore I believe that's exactly what they're going to do.”

Associational reasoning would have predicted just the opposite, because, like most human beings (and unlike Sherlock Holmes), it struggles to recognize the significance of "the dog that didn't bark."

This example highlights an important point: in complex adaptive systems, counterfactual reasoning often depends as much (or more) on an intuitive grasp of human behavior learned from the study of history and literature as it does on more formal methods.

Counterfactual reasoning serves many purposes, including learning lessons from experience (e.g., “what would have worked better?”) and developing and testing our causal hypotheses (e.g., “what is the probability that effect E would have or will occur if hypothesized cause X was/is present or not present?”).

While Dr. Pearl has developed a systematic approach to causal and counterfactual reasoning methods (see, “The Book of Why”), this remains a continuing challenge for machine learning methods, and indeed even for human reasoning. For example, the Intelligence Advanced Research Projects Activity recently launched a new initiative to improve counterfactual reasoning methods (the “FOCUS” program).

Thus far, we have only discussed various artificial intelligence/machine learning technologies and the challenges they face. However, three other challenges are also critical.

The first is the hardware on which AI/ML software runs. In many cases, training ML software is more time, labor, and energy intensive than many people realize (e.g., “Energy and Policy Considerations for Deep Learning in NLP”, by Strubell et al). However, recent evidence that quantum computing technologies are developing at a “super-exponential” rate suggests that this constraint on AI/ML development is likely to be significantly loosened over the next five to seven years (e.g., “A New Law Suggests Quantum Supremacy Could Happen This Year” by Kevin Hartnett). This dramatic increase in processing power that quantum computing will provide could, depending on software development (e.g., agent based modeling and simulation), make it possible to predict the behavior of complex adaptive systems and, using GANS, devise better strategies for achieving critical goals. Of course, this also raises the prospect of a world in which there are many more instances of “algorithm vs. algorithm” competition, similar to what we see in some financial markets today.

The second challenge is “explainability”. As previously noted, the statistical relationships that ML identifies in large data sets are often extremely complex, which makes it hard for users to understand and trust the basis for the predictions they make. For this reason, the development of “explainable AI” algorithms that can provide a causal logic for the predictions or decisions they make is regarded as critical precondition for broader AI/ML deployment.

If history is a valid guide, organizational obstacles will present a third challenge to the widespread deployment of ML and other AI technologies. In previous waves of information and communication technology (ICT) development, companies first attempted to insert their ICT investments into existing business processes, usually with the goal of improving their efficiency. The results, as measured by productivity improvement, were usually disappointing.

It wasn’t until changes were made to business processes, organizational structures, and employee knowledge and skills that significant productivity gains were realized. And it was only later that the other benefits of ICT were discovered and implemented, including more effective and adaptable products, services, organizations, and business models.

In the case of machine learning and other artificial intelligence technologies, the same problems seem to be coming up again (e.g., “The Big Leap Toward AI at Scale” by BCG, and “Driving Impact at Scale from Automation and AI” and “AI Adoption Advances, but Foundational Barriers Remain” by McKinsey). Anybody in doubt about this need only look at the compensation packages companies are offering to recruit data scientists and other AI experts (even though the organizational challenges to implementing and scaling up AI/ML technologies go far beyond talent).

Finally, there are potential regulatory obstacles to faster AI/ML deployment, especially in the area of data privacy, which could limit access to training data sets.

Let us turn now to the competitive, sector employment, and economic output and price level effects that the (eventual) widespread deployment of more advanced AI/ML technologies may have, and the uncertainties that surround them.

At the firm level, in “Robots and Firms” by Koch et al, the authors “study the implications of robot adoption at the level of individual firms using a rich panel data-set of Spanish manufacturing firms over a 27-year period (1990-2016).” They find that “robot adoption [by firms] generates substantial output gains in the vicinity of 20-25% within four years, reduces the labor cost share by 5-7% points, and leads to net job creation at a rate of 10%.”

However, they also find “substantial job losses in firms that do not adopt robots, and a productivity-enhancing reallocation of labor across firms, away from non-adopters, and toward adopters.” It is not clear whether these reallocations involve the same people. Other researchers often find that they do not, because the workers losing jobs lack the skills that are sought by the firms hiring people.

This certainly aligns with many stories about changes in employment in the financial sector, where, for example, large firms have cut trading floor staff but substantially increased technology and data science hires as algorithmic trading has increased.

At the sectoral level, there is no shortage of forecasts on the potential impact of accelerating development and deployment of automation and AI technologies. However, the initial estimates of potential job losses appear to have been too pessimistic. A more realistic assessment was provided in 2018 by the OECD, in their report on “Automation, Skills Use and Training”:

"The implications for jobs and skills of the developments in Artificial Intelligence and Machine Learning have dominated recent debates on the Future of Work and the changes brought about by digital technologies. Since Frey and Osborne (2013) shocked analysts and policy makers worldwide with a study suggesting that 47% of jobs in the United States are at high risk of being automated, several other researchers and institutions have contributed to the debate, all produced estimates in the high double digits.

"All these studies stem from an assessment by experts of the risk of automation for a subset of occupational titles, based on the tasks these occupations involved. This allowed identifying the so-called bottlenecks to automation – i.e. the tasks that, given the current state of knowledge, are difficult to automate. These include: social intelligence, such as the ability to effectively negotiate complex social relationships, including caring for others or recognizing cultural sensitivities; cognitive intelligence, such as creativity and complex reasoning; and perception and manipulation, such as the ability to carry out physical tasks in an unstructured work environment. These bottlenecks were used to compute a risk of automation for occupational titles that were not included in the expert assessment and for countries outside the United States.

"More recent studies, exploiting the Survey of Adult Skills (PIAAC), brought the estimates of the share of jobs at risk of automation down significantly. These studies show that there is considerable variation in the tasks involved in jobs having the same occupational title and that accounting for this variation is essential to gauge the extent of the problem. Arntz, Zierhan and Gregory (2016), for instance, put this share to 9% in the United States. While this figure is only a fraction of the estimate provided by Frey and Osborne, it translates to approximately 13 million jobs across the United States, based on 2016 employment figures. As job losses are unlikely to be distributed equally across the country, this would amount to several times the disruption in local economies caused by the 1950s decline of the car industry in Detroit where changes in technology and increased automation, among other factors, caused massive job losses. The current study aims to go beyond providing an estimate of the share of jobs at high risk of automation by also highlighting the significant changes that jobs will undergo as a result of the adoption of new technologies. It also offers an analysis of the distribution of risk among different population groups and the role of training in helping workers transit to new career opportunities.

"Here are the study’s key findings: Across the 32 countries, close to one in two jobs are likely to be significantly affected by automation, based on the tasks they involve. But the degree of risk varies. About 14% of jobs in OECD countries participating in PIAAC are highly automatable (i.e., probability of automation of over 70%). Although smaller than the estimates based on occupational titles obtained applying the method of Frey and Osborne (2013) this is equivalent to over 66 million workers in the 32 countries covered by the study…[In addition], the regional concentration of the risk of automation could amplify its social and economic impact, particularly in countries where geographical mobility is low.

"Another 32% of jobs have a risk of between 50 and 70% pointing to the possibility of significant change in the way these jobs are carried out as a result of automation – i.e. a significant share of tasks, but not all, could be automated, changing the skill requirements for these jobs. The variance in automatability across countries is large: 33% of all jobs in Slovakia are highly automatable, while this is only the case with 6% of the jobs in Norway. More generally, jobs in Anglo-Saxon, Nordic countries and the Netherlands are less automatable than jobs in Eastern European countries, South European countries, Germany, Chile and Japan…

“There are upside and downside risks to the figures obtained in this paper. On the upside, it is important to keep in mind that these estimates refer to technological possibilities, abstracting from the speed of diffusion and likelihood of adoption of such technologies. Adoption, in particular, could be influenced by several factors, including regulations on workers dismissal, unit labour costs or social preferences with regard to automation. In addition, technology change will without doubt also bring about many new jobs…

“But there are risks on the downside too. First, the estimates are based on the fact that, given the current state of knowledge, tasks related to social intelligence, cognitive intelligence and perception and manipulation cannot be automated. However, progress is being made very rapidly, particularly in the latter two categories. Most importantly, the risk of automation is not distributed equally among workers…The occupations with the highest estimated automatability typically only require basic to low level of education…

“This unequal distribution of the risk of automation raises the stakes involved in policies to prepare workers for the new job requirements. In this context, adult learning is a crucial policy instrument for the re-training and up-skilling of workers whose jobs are being affected by technology. Unfortunately, evidence from this study suggests that a lot needs to be done to facilitate participation by the groups most affected by automation. The odds of participating in any type of training, on-the-job and outside the job, are found to be significantly lower among workers in jobs at risk of being automated…

“In parallel, the large share of workers whose jobs are likely to change quite significantly as a result of automation calls for countries to strengthen their adult learning policies to prepare their workforce for the changes in job requirements they are likely to face.”

Let us now turn to the potential impact of automation and AI technologies on macroeconomic aggregates, especially growth in productivity and aggregate demand, and the distribution of income within and across nations.

With respect to productivity growth, in “Artificial Intelligence and the Modern Productivity Paradox: A Clash of Expectations and Statistics”, Erik Brynjolfsson and his co-authors make the case that the productivity impact of automation and AI technologies will eventually be substantial, but will take time to appear because of the previously noted obstacles to their widespread implementation. A more pessimistic case is made by Robert Gordon in “Why Has Economic Growth Slowed When Innovation Appears to be Accelerating?” in which he argues that stagnating improvement in educational quality, declining research productivity, and the slow diffusion of automation and artificial intelligence technologies will all constrain productivity improvement, and thus the growth of potential economic output (aggregate supply) in an economy where the size of the workforce is not growing as fast as it did in the past.

There is substantially more disagreement over the future impact of automation and AI technologies on aggregate demand and the distribution of income.

In “The Macroeconomic Impact of Artificial Intelligence”, PWC presents an optimistic case:

“Global GDP is estimated to have been approximately $75 trillion in 2016. Our baseline projections suggest that that figure is estimated to be approximately $114 trillion by 2030…This could be up to 14% higher in 2030 as a result of AI – the equivalent of up to $15.7 trillion...

“The economic impact of AI will be driven by (a) productivity gains from businesses automating processes as well as augmenting their existing labour force with AI technologies (assisted, autonomous and augmented intelligence) and (b) increased consumer demand resulting from the availability of personalised and/or higher-quality AI-enhanced products and services. Approximately 58% of the 2030 GDP impact will come from consumption side impacts.”

A more pessimistic case is present by Gries and Naude in their paper “Artificial Intelligence, Jobs, Inequality, and Productivity: Does Aggregate Demand Matter?”

The authors begin by noting that, “there are still very few theoretical growth models that incorporate AI, and virtually none that consider demand side constraints. A second concern is that the predictions of AI causing massive job losses and faster growth in productivity and GDP are at odds with reality: if anything, unemployment in many advanced economies are at historical lows. However, wage growth and productivity is stagnating and inequality rising.”

With respect to aggregate demand, the authors note that:

“Aggregate demand has mostly been neglected in predictions and analysis of the impacts of AI so far, as the literature tends to take an overtly supply-side approach. While the supply side is of great importance, because innovations in AI technologies improve the productive capacity of the economy and determine potential economic growth, weak demand can restrict actual growth. If AI-automation generates economic growth from which labour does not earn more income, then consumption demand will not grow sufficiently to absorb the additional production capacity [i.e., potential supply].

A lack of aggregate demand will thus continue to restrict growth if: (i) AI and human labour are highly substitutable [i.e., if improvements in automation and AI lead to significant net job losses], so that labour's overall income share [and thus consumption spending] declines relative to the capital share; and (ii) those economic agents who gain income shares from AI - like technology providers or financial wealth [capital] holders - do not spend their additional income to absorb the growing potential production [note that this is what economic theory would predict, given the declining marginal utility of consumption]”.

With respect to the distribution of income, the authors conclude that:

“The substitutability between labour and AI is a vital parameter. High elasticities of substitution will lead to a decline in employment, a decline in wages and the labour share of income, and greater inequality with a larger share of income accruing to the providers of the AI. Because the latter [will not spend all their additional income], consumption will decline [at least in the absence of much more aggressive income redistribution].”

What then is the likely future rate at which automation and AI technologies will be substituted for labor?

Today, multiple interacting causes have depressed the price of labor and employee incomes, including weak aggregate demand, intense competition in globalized labor markets, rising corporate concentration and market power, and inadequate education systems. This has slowed the rate at which capital is being substituted for labor, producing higher levels of employment without significant growth in productivity and real wages. Substantial improvements in these causal drivers are unlikely.

In sum, at the aggregate level, the most likely scenario for the next five to seven years seems to be slower than expected deployment of automation and artificial intelligence technologies because of both the time required to overcome technical and organizational obstacles and the low relative price of labor. At the micro level, however, companies that buck this overall trend and successfully deploy these technologies should reap substantial rewards for their shareholders and employees. However, their success will also contribute to worsening income inequality.

Over a longer time horizon, as automation and AI technologies are eventually more broadly deployed, at the aggregate level they will likely produce fewer job losses than many expect. However, the job losses that occur are likely to disproportionately affect certain categories of employees, sectors, regions, and nations. In the absence of a prompt and effective government response, this is likely to substantially increase both overall macro uncertainty and demands for political action.

While the OECD report is hopeful that improved lifetime learning and retraining initiatives can significantly mitigate these negative effects, we are not; the culture of most national education systems (as they stand today) is too resistant to change. Thus the most likely outcome of accelerating deployment of automation and AI technologies is not the transfer of jobs from lagging to leading companies and sectors, but rather higher unemployment, increased transfer payments and other social safety net supports for displaced workers (and those who, because of their poor educations, cannot find jobs that pay enough to enable them to thrive in the future economy). In turn, higher social safety net spending will require higher, and (because of the worsening income distribution) likely much more progressive taxes which will induce greater migration to lower tax states and nations.

At the global level, the eventual broad deployment of automation and AI technologies will likely widen the gap between leading and lagging nation states, with the US and China (and a few others) among the former. In turn, this will likely trigger increased migration flows, which will come on top of those triggered by the disproportionate impact of climate change. However challenging today’s global migration issues are today, it seems likely they will worsen in the future.

Finally, let us turn to the implications of this analysis for the probabilities for different regimes, and returns on broad asset classes.

Implications for Regime Probabilities

Persistent Deflation

A slower rate of deployment for automation and AI technologies will reduce the rate of productivity growth and thus the rate at which potential supply increases relative to demand. At the margin, this will slow the build up of supply-driven deflationary pressures.

On the other hand, stagnant real incomes, weak aggregate demand, and increases in government social safety net spending will likely drive the continued accumulation of debt in a global economy that is already highly levered. This increases the likelihood that a cyclical economic downturn could trigger destructive debt-driven deflation, in which cascading bankruptcies lead to job losses and falling prices.

On balance, these complex effects will reduce the chances of being in the persistent deflation regime in the short term, but increase them over a longer time horizon (i.e., beyond 5-7 years).

High Inflation

If a debt-driven deflation occurs, tax collections will fall and governments will be forced to increase money supply growth to service their high (and rising) debt. At some point, increasing debt monetization could, as it has in the past, cause a substantial increase in inflation and loss of confidence in the currency.

High Uncertainty

All else being equal, uncertainties about the rate at which the capabilities of automation and AI technologies are improving, the rate at which they will be deployed, and their future impacts on employment, productivity, and aggregate demand will increase the chance of the global macro system being in the high uncertainty regime.

Normal Regime

It is possible to see how the development and deployment of advanced automation and AI technologies could lead to a return to the normal regime of strong economic growth and attractive equity returns. In this scenario, improved technology and more capable employees drive faster productivity and demand growth, whose benefits accrue to both capital providers (i.e., rising returns to shareholders) and workers (i.e., rising real household earnings). In turn, this reduces the pressure on governments to increase social safety net spending at the same time that faster economic growth leads to higher tax collections, which enable a reduction in high debt/GDP levels.

While this scenario is undoubtedly the most desirable of all potential outcomes, it critically depends on two changes that today seem unlikely: substantial improvements in labor quality and labor’s share of national income.

Implications for Broad Asset Class Returns

Equity

Over time, increasing disparities in economic growth and equity market returns across countries (driven by the rate at which they develop and deploy automation and artificial intelligence technologies).

In nations that are technology leaders (e.g., the U.S., China, and possibly others), continuation of the trend toward a power law distribution of returns across firms and sectors.

Fixed Income

Lower rates of aggregate demand growth and continuing high uncertainty will maintain downward pressure on real sovereign bond yields (except under a return to the Normal Regime scenario, in which case real yields would increase because of faster productivity and economic growth).

Differences in both national and company specific rates of automation and AI adoption as well as exposure to deflation risk (e.g., because of high debt levels) will have strong effects on the distribution and pricing of credit risk.

Commercial Property

All else being equal, weaker aggregate demand, and, in the worse case, reduced employment levels will depress property net operating income. However, continuing demand from yield seeking investors confronted with weak bond returns should support demand for prime properties.

Differences in both national and regional rates of automation and AI adoption will widen risk spreads across properties.

If debt deflation develops and leads to fears of subsequent high inflation, the price of real physical assets like property (as well as gold and timber) will increase due to their hedging value.

Commodities

Except in the case of a return to the Normal Regime, weak aggregate demand will depress commodity returns.

If the probability of the High Inflation regime developing increases, returns are likely to be highest on commodity products backed by relatively liquid physical assets, like gold and timber (and also, depending on how environmental regulation evolve, oil and gas reserves). Commodity futures and futures-based financial products that lack these features may be relatively less attractive due to concerns over counterparty risk and liquidity.