Feature Article: Understanding and Predicting Uncertainty Shocks: An Update

Back in 2010, The Index Investor published a two-part series on “Understanding and Predicting Uncertainty Shocks”. Since then, a lot more research on the root cause and effects of uncertainty shocks has been published. This month we briefly review highlights from this research, as well as their implications for investors.

At the outset, let’s all agree that the definition of “uncertainty” is notoriously hard to pin down. The broadest definition we’ve seen is “the absence of certainty.” A better definition may be “doubt that produces anxiety and inhibits action.”

As a practical matter, this definition situates uncertainty in the middle of a spectrum that runs from certainty to risk to uncertainty to ignorance. In the face of certainty, there is no doubt, and action is not inhibited. Nor is action typically inhibited in risky situations, where the full range of possible outcomes, along with their consequences and probabilities, are all believed to be well understood. Strangely enough, it is also usually the case that ignorance does not inhibit action. As John Maynard Keynes noted in Chapter 12 of The General Theory of Employment, Interest, and Money, in the face of ignorance we typically adopt “shared conventions” (i.e., common assumptions) that enable action. As Keynes noted, the most familiar of these “lies in assuming that the existing state of affairs will continue indefinitely, except in so far as we have specific reasons to expect a change.”

More so than risk or ignorance, it is uncertainty that is usually the greatest obstacle to action. More specifically, doubt and anxiety typically peak when we confront situations in which we do not understand the full range of possible future outcomes, and/or their consequences, and/or the relative likelihood that they will occur.

The extent of uncertainty we confront today is arguably greater than ever before. There are two main reasons for this.

The first is the inexorable tendency of systems (including organizations) to become more complex over time. Thanks to the second law of thermodynamics, in the absence of new injections of energy and/or information, open systems will inexorably become more disordered over time. Increasing disorder creates problems, and we use energy and information to create solutions to them. Yet these new solutions are often more complex than the ones that preceded them, and thus lead to new problems, necessitating even more complex solutions, whether technological or organizational.

As complexity accumulates, it becomes ever more difficult to fully explain the system effects we observe, because of the existence of multiple interacting causes, some of which produce non-linear and time-delayed effects. In the absence of causal understanding, prediction becomes more difficult, or devolves to associational/statistical reasoning, as in the case today’s deep learning methods whose underlying logic is often difficult or impossible to explain.

The second reason uncertainty is higher today than ever before is the dramatic increase in global connectivity over the last decade. While the increase in complexity has made the behavior of many systems more non-linear, the exponential increase in connectivity and the speed with which information flows across these connections has made many socio-technical systems more tightly coupled (both internally and externally) than ever before. Thirty-five years ago, in “Normal Accidents”, Charles Perrow famously observed that unexpected failure pathways (and the uncertainty they create) increase exponentially as system complexity and connectivity increase.

Recent Research

Research into uncertainty shocks has in recent years been greatly aided by the development of specific indexes for measuring uncertainty, based on the pioneering research of Baker, Bloom, and Davis in “Measuring Economic Policy Uncertainty”.

“Really Uncertain Business Cycles”, by Bloom et al uses a model of the economy with heterogeneous agents (rather than a single representative agent) to find that “uncertainty shocks can generate drops in GDP of around 2.5%”. Increased uncertainty also initially reduces the initial positive impact of policy actions taken to stimulate the economy in a recession.

In “The Nonlinear Effects of Uncertainty Shocks”, Jackson et al from the Federal Reserve Bank of St. Louis find that “when uncertainty is relatively low, fluctuations in uncertainty have small, linear effects. [However] in periods of high uncertainty, the effect of a further increase is magnified, and have a more pronounced effect on real economic variables…This is due to uncertainty propagating through both reductions in consumption spending and business investment and employment.”

“The Macroeconomic Impact of Financial and Uncertainty Shocks”, by Caldara et al finds that financial shocks are distinguishable from uncertainty shocks; however the latter “have an especially significant economic impact when they trigger a tightening of financial conditions, as in the Great Recession.” These same issues are also addressed in another paper, “The Real and Financial Impact of Uncertainty Shocks”, by Alfaro et al, which also concludes that uncertainty shocks can be particularly damaging when financial conditions are fragile.

In “Policy Uncertainty and Aggregate Fluctuations”, Mumtaz and Surico find that uncertainty about taxes and government debt have a “large and persistent effect on output, consumption, investment, consumer confidence, and business confidence.” In contrast, uncertainty about government spending and monetary policy has a significantly smaller impact on the real economy.

In “Uncertainty Traps”, Fajgelbaum and his co-authors show “how large by apparently short-lived uncertainty shocks can generate long-lasting recessions.” As noted by other researchers, an increase in uncertainty generates a fall in consumption and investment. “Since agents learn from the actions of others, information flows slowly in times of low economic activity, uncertainty remains high, further discouraging consumption and investment” in a self-perpetuating cycle.

Finally, in “The Evolving Impact of Global, Region-Specific, and Country-Specific Uncertainty”, Mutaz and Musso find that “common global uncertainty plays a primary role in explaining the volatility of inflation, interest rates, and stock prices.” This is similar to the conclusions in a recent IMF working paper “Media Sentiment and International Asset Prices”, which found that “not all news sentiment is alike. Changes in global news sentiment have the largest impact on equity returns around the world, which does not reverse in the short run.”

Similarly, in their new paper, “What are Uncertainty Shocks?”, Kozeniauskas et al begin by observing that “one of the primary innovations in modern business cycle research is the idea that uncertainty shocks drive aggregate fluctuations. A recent literature demonstrates that uncertain shocks can explain business cycles, financial crises, and asset price fluctuations with great success. But the measures of uncertainty used are wide ranging…If these disparate measures are really capturing a common underlying shock, what is it?”

To answer this question, they note that, “people become uncertain after observing an event that makes them question future outcomes. That raises the question: What sorts of events can make agents uncertain in a way that shows up in a wide variety of disparate measures?” They note that the fundamental driver of uncertainty shocks across multiple metrics is unexpected changes in the macro political economy (including perceived “disaster” or tail risk). This is also the exact focus of analysis and forecasting here at The Index Investor.

What Causes Uncertainty Shocks?

The economist Rudiger Dornbusch famously noted that, “crises take a much longer time coming than you think, and then happen much faster than you would have thought.”

Today’s hyperconnected world provides ample evidence of the accuracy of this insight. The more connected we become, the greater the potential for rapid shifts in opinion and behavior that create new strategic risks for many investors and organizations.

In order to better anticipate these sudden and substantial shifts, investors should understand the underlying processes that are at work. In practice, these often operate together, with positive (amplifying) feedback loops between them.

Our starting point is Daniel Kahneman’s distinction (made in his book, “Thinking Fast and Slow”) between two modes of thought.

“System 1” is rapid, instinctive, based on association, and generally unconscious. It generates fast conclusions and emotions that prime us for actions (e.g., fight or flight) that, from an evolutionary perspective, have been highly adaptive for the survival and success of the human species.

In contrast, “System 2” thinking is slower, deliberate, logical, and conscious. Because it is also more effortful, most human cognition is based on System 1, not System 2.

Rapid shifts in popular perception, opinion, and behavior that generate uncertainty shocks originate in both System 1 and System 2.

In the case of System 1, the most primal driver is human beings’ capacity to rapidly transmit fear, through a variety of non-verbal means, including facial expressions and odor.

A slightly more evolved System 1 response is the automatic increase in our desire to avoid disagreement or abandonment by a group that occurs when perceived uncertainty or adversity increase. In the evolutionary sense, this is adaptive behavior. However, in today’s environment it is the root cause of excessive conformity and groupthink in periods of high uncertainty, which, as we will describe below, reduces the diversity of narratives in a system and sets the stage for sudden, substantial change.

Conscious and effortful System 2 processes can also contribute to uncertainty shocks.

In “The Evolution of Social Learning and Its Economic Consequences”, Bossan et all demonstrate how over the course of evolution, social learning – deliberately imitating others – has became dominant. They note that this has had three critical effects. “First, for better or worse, the decisions of social learners are more exaggerated than those of individual [trial and error] learners. Second, social learners react with a delay to changes in the environment. Third, the behavior of social learners can become more and more detached from reality.”

In “Social Learning Strategies Regulate the Wisdom and Madness of Interactive Crowds”, Toyokawa et al ask, “why do groups of individuals sometimes exhibit collective wisdom and other times maladaptive herding?” They find that “challenging tasks [which generate greater uncertainty] elicit more conformity among individuals, with rates of social copying increasing with group size, leading to a higher likelihood of herding among large groups confronted with high uncertainty.”

Similar conclusions have been reached by two other research teams, including Escudero and Polavieja in “Adversity Magnifies the Importance of Social Information in Decision Making” and Rendell et al in “Why Copy Others? Insights from the Social Learning Strategies Tournament”.

Researchers have also found that while “loss aversion” is a normal human response when the result of a decision will remain private, when those results will be visible to others gains become more important (see, “Interdependent Utilities: How Social Ranking Affects Choice Behavior” by Bault et al). Combine this with human’s optimism and confirmation biases, and you can see why investors have a bias towards believing in positive market and macro narratives despite accumulating evidence that tells a more negative story.

Still other research finds that a relatively small fraction of agents in a population (about 10% or more) can produce and sustain significant opinion shifts (for example, see, “Social Consensus Through the Influence of Committed Minorities” by Xie et al, and “Committed Activists and the Reshaping of the Status Quo Consensus” by Mistry et al). Put differently, when a relatively small percentage of people in a group switch to a different narrative, or when an even smaller but more highly connected group does this, sudden and substantial changes in behavior and asset values – a classic uncertainty shock – can result.

Broadly speaking, decisions made by System 2 are based on three types of input: private information not available to everyone, public information, and social information.

As described in the work of a number of researchers (e.g., David Tuckett, Robert Shiller, Paul Ormerod, and Marvin Cohen), in the face of situations characterized by high complexity and uncertainty, humans tend to decide and act not on the basis of detailed quantitative analyses (which are impractical and often impossible), but rather on the basis of so-called “conviction narratives” or, more simply, the stories we tell ourselves and others to justify our beliefs and actions on the basis of relatively simple causal models.

As system complexity increases and causation becomes exponentially more difficult to fully understand, these simplified narratives rest on progressively more fragile assumptions (for more on this, see Gary Klein and Robert Hoffman’s papers on “Explaining Explanation”).

These narratives produce varying levels of certainty (or “conviction”), along with similarly varying degrees of emotional excitement or anxiety regarding the accuracy of our perceptions, the appropriateness of our proposed actions, and the likelihood they will achieve our goals (e.g., see: “News and Narratives in Financial Systems: Exploiting Big Data for Systematic Risk Assessment” by Nyman et al, and “The Economics of Radical Uncertainty” by Paul Ormerod).

From the aggregation and sharing of these individual conviction narratives and the net impact of their associated levels of emotional valence and arousal there emerges the phenomenon Keynes called “animal spirits," or what we more generally call the level of public confidence (and its complement, the level of uncertainty).

People have varying degrees of conviction about the accuracy and importance of the private information they possess. As we have noted, in the face of high uncertainty, or when they are members of large groups (which naturally arise in our age of hyper-connectivity) people will often weigh social information much more heavily than private information when constructing narratives to explain the past and forecast the future.

The surprising result is that increasing uncertainty therefore tends to produce a reduction, rather than an increase, in the number of competing narratives in a population, which makes a system more susceptible to sudden and dramatic change.

Under these circumstances, an unexpected public signal – for example, the announcement that a high-ranking government official has been indicted for a crime, or that a key party is abandoning a coalition government – can quickly undermine fragile but widely believed narratives, and cause dramatic changes in consumer and business confidence and behavior.

In a hyper-connected world, this result can also be produced endogenously, if, as noted above, a relatively small percentage of the population changes its view due to accumulating private information. Indeed, this is the underlying causal logic guiding the actions of those who attempt to change public opinion by targeting information (be it press releases or fake news) at highly connected people (e.g., “influencers”) in social networks.

Can We Anticipate Uncertainty Shocks?

After four years on the Good Judgment Project, there isn’t any doubt in my mind about that. The more important questions are how far in advance we can do this, at what level of aggregation, with what degree of accuracy, and whether and to what extent we act on our forecasts in a timely manner.

Let’s look first at the time horizon question. Perhaps the oldest technique for anticipating uncertainty shocks in financial markets is calculating broad asset class valuation ratios, and comparing them to historical averages. Closely related to this is the work of Didier Sornette and others on the identification of emerging financial market bubbles. However, when these techniques lead to the conclusion that an asset class is severely overvalued (and therefore likely to produce a future uncertainty shock), you will always be told that, “this time it’s different.” Reinhart and Rogoff looked at eight centuries of data and wrote a book with this title to remind you that it’s not.

The larger problem, as Keynes said, is that, “the market can stay irrational longer than you can stay solvent.” To which professional asset managers often add, “or sustain the career risk of being underweight in an asset class when both its returns and its overvaluation is increasing.”

That has led to the search for tools and methods that would enable a manager to more accurately time the arrival of uncertainty shocks. The simplest of these are based on differential returns between assets that perform best under different regimes (e.g., persistent deflation or high inflation). Other simple tools use debt market credit and liquidity spreads. We use all of these.

More advanced methods either make use of larger amounts of data and/or advanced methods of extracting insights from it. A simple example of this is our use of the month-to-month autocorrelation of returns across multiple asset classes, which is an indicator that research has found to be useful (in some, if not all cases) when predicting large changes in the behavior of complex systems.

Various machine learning techniques are now widely used to extract insights from the large amounts of data that is available today. For example, text data from traditional and social media feeds is mined to identify emerging issues and measure changes in various types of sentiment (see, for example, Richard Nyman’s excellent thesis, “An Algorithmic Investigation of Conviction Narrative Theory”). In some cases, insights derived from textual data are converted into standardized metrics, like the Economic Policy Uncertainty Index we use in our analysis of market stress levels.

However, these techniques are still limited by a number of important shortcomings. For example, text mining usually depends on the number of times certain words or phrases appear in the data set being analyzed, and/or the frequency of their use by agents in a network (e.g., what CEOs of public companies said in this quarter’s earnings conference calls with investment analysts). Simple frequency based approaches require that word or phrase use reach a certain level before being deemed significant. This creates the classic tradeoff between Type 1 and Type 2 errors, or false alarms and missed alarms. If the threshold word or phrase frequency trigger is set low, there will be a lot of false alarms; if it is set high, there will be more missed alarms.

Tracking word or phrase usage by important individuals (or high value nodes in network science terms) is one approach that has been used to improve the Type 1 vs. Type 2 tradeoff. However, placing more weight on fewer sources runs its own set of risks, from deliberate deception to lack of independence (e.g., as when many high value nodes are basing their opinions on the same set of flawed information inputs).

A far more fundamental limitation of machine learning techniques based on natural language processing is the subject of “The Book of Why”, by Dr. Judea Pearl, one of the preeminent scientists in the field of artificial intelligence. He posits a hierarchy of reasoning skills. At the lowest level is association or correlation. Today’s machine learning methods can recognize complex patterns in large data sets better than ever before. Yet they fundamentally operate in a statistical rather than model-driven mode.

The next highest level is causal reasoning that clearly connects causes and effects. This is something that today’s artificial intelligence techniques are far from mastering, particularly in the case of complex adaptive systems where the rules governing allowable actions are constantly evolving.

As we noted in last month’s feature article on the critical difference between threats and threat signatures, making AI more “context aware” (i.e., able to combine data with existing knowledge), and capable of creating abstractions and using them to reason across different situations (i.e., “transfer learning”) is the subject of multiple initiatives in both the private (e.g., computational narrative intelligence and story generation to explain data) and public sectors (e.g., DARPA and IARPA projects like AIDA, KAIROS, CREATE, and ICARUS).

At the top of Pearl’s hierarchy sits counterfactual reasoning, which builds on associational and causal methods and enables both retrospection and imagination. For example, the goal of IARPA’s FOCUS project is to substantially improve counterfactual reasoning methods.

Given the limitations of current artificial intelligence and other machine learning technologies, predicting uncertainty surprises, particularly in complex adaptive systems over longer time horizons, remains an activity where human cognition still reigns supreme.

Another important issue when predicting uncertainty shocks is the focus on your analytical efforts. More specifically, how do you divide your time between forecasting discrete variables (i.e., events) and continuous variables (i.e., trends)? For example, consider a company with a substantially overvalued stock price. In this case, a specific event – an earnings miss, a failed transaction, or publication of a critical research report – will likely trigger an uncertainty shock and a sharp fall in the company’s valuation. The key forecasting question is when one of those events is most likely occur.

Forecasting macro uncertainty shocks is different. To be sure, when they occur the media will always identify proximate events that they deem to be causal. They may well be, but only in the very narrow sense of the last grain of sand that causes a sandpile in which stress has been building to give way in a large slide.

The forecasting problem is that, in an integrated multilevel complex adaptive system characterized by constant changes in technological, economic, national security, social, and political conditions, there are an unknowable number of specific events that could trigger an uncertainty shock that produces large changes in financial asset prices. Trying to forecast them all is impossible.

Instead, our methodology focuses on forecasting the evolution of continuous variables, and how close we are to critical thresholds that will trigger uncertainty shocks and regime changes in the global macro system.

For a specific example of a forecast that was based on this approach, see our May 2007 article, “Why We Don’t Sleep Well These Days, and Moving Into Cash Looks Like a Good Idea.”

Can We Hedge Exposure to Uncertainty Shocks?

In broad terms, there are three strategies for hedging exposure to uncertainty shocks.

The first is to formulate “robust” plans that maximize the probability of achieving your goals across a wide range of possible future scenarios. In an investment context, the application of this principle leads to the recommendation to structure portfolios that are broadly diversified across a wide range of asset classes.

The second strategy is to ensure that a system or organization or portfolio has an adequate level “resilience” in case robustness fails. In investment terms, this takes the form of liquid assets; at a company this is more broadly interpreted as redundancy and hardening of critical systems, organizational techniques like critical incident teams and crisis training, and foregoing maximum efficiency in favor of maintaining a certain degree of “slack” resources that can be quickly deployed following a shock.

The final strategy is maximizing the adaptive capacity of an organization or portfolio. If robustness fails, and resiliency absorbs the initial negative impact, adaptation drives a quick return to previous levels of performance. Indeed, along with effectiveness and efficiency, adaptability is one of evolutions three core performance metrics, that apply to every organism, from the smallest bacteria to the largest organizations and nations.

There are three key elements to adaptability. Skill in anticipatory thinking continuously identifies potential threats, and enables the preparation and prioritization of options for responding to them. Tools like the “Met Office Model” we use at both the Index Investor and Britten Coyne Partners enable organizations to manage the complex dynamic of the speed at which threats are developing and the remaining time required to adequately respond to them. And developing organizational capacity for problem detection, solution design, and rapid implementation (often collectively termed “agility”) ensures that the adaptive process is continuous and effective.

So, to sum up:

(1) Macro uncertainty shocks are caused by individual and collective forces that are deeply rooted in our evolutionary past, and have been supercharged in our present age of high complexity and hyper-connectivity. As former Bank of England Governor Mervyn King has observed, we now live in “an age of radical uncertainty”.

(2) Macro uncertainty shocks have substantial and long-lasting effects, and their full impact usually occurs with a time lag.

(3) Macro uncertainty shocks can be anticipated, not by trying to forecast the exponentially large number of discrete events that can be their proximate cause, but rather by focusing on the trends that can push complex adaptive systems beyond critical thresholds (i.e., “tipping points”) and trigger sudden non-linear effects and regime change.

(4) By focusing on robustness, resiliency, and adaptability, we can hedge our exposure to uncertainty shocks.