Allow me to begin with a paradox that never ceases to fascinate me. For decades, economists have tried to understand how one event gives rise to another: how an interest rate hike kills business activity, how an inflationary impulse spreads through the economy like ink in water. And all this time, the main enemy of their understanding has not been a lack of data, but noise. Or rather, the wrong kind of noise. Or, more accurately, a misunderstanding of what noise is at all.
A group of researchers has proposed a new way of working with this noise. Their method bears the intimidating name 'Sparse Heterogeneous Markov-Switching Heteroskedasticity.' Let's agree to call it simply SHMSH, as the authors themselves do. And I will try to explain why it matters – through analogies that don't require a dissertation.
What is a 'structural shock,' and why should we care?
Imagine you're listening to an orchestra. Violins, cellos, trumpets – all playing at once. You hear the overall sound, but you want to understand: what exactly did the trombonist do in the third bar? How did his note affect everything else?
The economy is the same kind of orchestra. Inflation, interest rates, industrial production – all these variables sound at once, mutually influencing one another. Economists use a tool called structural vector autoregression – SVAR in English – to 'disassemble' this orchestra into its instruments and hear each one separately.
A 'structural shock' in this context is precisely that trombone blast: an unexpected, isolated event that we want to trace throughout the score. For example, a monetary policy shock is an unforeseen change in the interest rate that the U.S. Federal Reserve made beyond what the market already expected. It is this 'unexpected part' that is the shock. And it's the impact of this part on prices and production that researchers want to measure.
The problem is that 'disassembling the orchestra' is a mathematically complex task. To understand the matrix of simultaneous interactions between variables (who influences whom instantly, without a delay), one needs additional assumptions. Traditionally, these were drawn from economic theory: for example, 'prices do not react to shocks instantaneously,' or 'production reacts with a lag.' But what if the theory is debatable? What if the constraints we impose are themselves distorting the result?
This is where identification through heteroskedasticity comes in. A complex term for a simple idea.
Heteroskedasticity as a Detective
'Heteroskedasticity' sounds like a curse in an unfamiliar language. But behind this word lies something quite mundane: volatility changes over time. In other words, markets are sometimes calm, sometimes turbulent. In the 1960s, the American economy was relatively stable. In the late 1970s and early 1980s, during the era of double-digit inflation and drastic actions by the Fed under Paul Volcker, volatility was off the charts. Then came the 'Great Moderation' – a period from the mid-1980s to the mid-2000s when the fluctuations of economic variables noticeably decreased.
And here is the key insight: if shocks change their 'volume' in different historical periods, this in itself carries information. Mathematically, this means that in different eras, the covariance matrix of residuals (roughly speaking, the 'pattern of interconnections between deviations') looks different. And if we know how exactly it changes, we can 'figure out' the structure of the orchestra – that is, determine who influences whom instantly – without any theoretical constraints.
This is reminiscent of detective work. Imagine several events have occurred in a room: someone dropped a glass, a door slammed, a telephone rang. If you know that in the first instance, the glass always falls louder than the door, and in the second, it's the other way around, you can reconstruct the sequence and causes without even seeing the events themselves. The different 'volume levels' in different periods are this heteroskedasticity, and it is what allows us to 'crack the case.'
Regimes and Markov Switching: When the World Changes in Fits and Starts
But how exactly can we describe this changing volatility? One approach is models with Markov switching. Their essence is this: the economy at any given moment is in one of several 'regimes' – for example, a 'calm period' or a 'turbulent period.' The transition between regimes happens randomly, but with certain probabilities. Today you are in a calm regime; tomorrow – with some probability – you switch to a turbulent one.
The name 'Markovian' refers to the Russian mathematician Andrey Markov, who described such memoryless processes in the early 20th century: the probability of the next state depends only on the current one, not on the entire history. It's a convenient and powerful abstraction.
However, previous models with Markov switching suffered from one limitation: they assumed that all shocks switch between regimes simultaneously. Imagine that the trombone and the violin must always play either loud together or quiet together – no independence. This is obviously unrealistic: an inflation shock can be highly volatile while a production shock is behaving quite calmly.
This is precisely where the authors introduce their first key innovation: heterogeneity. Each shock gets its own, independent Markov process. The trombone decides for itself when to play loud. The violin, for itself. This makes the model far more plausible from the perspective of real economic dynamics.
Sparsity: When Extra Regimes Simply 'Go Silent'
The second innovation is even more elegant. Imagine you want to describe a musical piece but don't know in advance how many movements it has. You could assume there are many – say, ten – and simply let the model itself decide which of them actually exist in the data and which remain empty.
This is exactly what is called sparsity in this context. The authors specify a deliberately large number of potential volatility regimes for each shock. But the model is designed so that the extra regimes simply 'collapse' – they are never activated on the real data, remaining empty. As a result, the model itself determines how many regimes are actually needed to describe a specific shock: maybe two, maybe three – depending on what the data says.
This eliminates a long-standing headache for econometricians: the need to choose the number of regimes in advance. Previously, this choice was arbitrary or required a laborious process of trial and error. Now, the model does it by itself – through Bayesian statistics, which 'penalizes' redundant regimes not supported by the data.
Verification: What if There's No Volatility at All?
Here we come to a crucial question, which the authors formulate with commendable honesty: what if there is no heteroskedasticity? What if volatility is actually constant, and the entire method of identifying through its changes is simply pointless?
This is called testing the hypothesis of homoskedasticity – that is, the hypothesis that the variance of shocks does not change over time. Applying identification through heteroskedasticity to homoskedastic data is like trying to identify a culprit by fingerprints when no one has left any.
The model's sparse structure turns out to be unexpectedly useful here. If, for a particular shock, all volatility regimes are statistically indistinguishable – meaning the 'volume' doesn't change in any regime – this in itself is strong evidence for homoskedasticity for that specific shock. The model, therefore, simultaneously assumes the possibility of switching and tests whether it is real. It's a built-in self-check mechanism.
How It Works in Practice: U.S. Monetary Policy from 1960 to 2007
The authors test their model on data familiar to any macroeconomist: the consumer price index, the federal funds rate, and the industrial production index for the United States. The observation period is from January 1960 to December 2007. The choice of the end date is not accidental: beyond it lies the Great Recession of 2008–2009, with its unconventional monetary policy measures that could distort any analysis claiming to capture 'normal' dynamics.
What does the model find? First, it confirms the presence of heteroskedasticity – specifically for the monetary policy shock. This means that identification through volatility changes is justified here. Second, the model identifies at least two volatility regimes for the monetary shock: a period of relative calm, corresponding to the 'Great Moderation' from approximately the mid-1980s to 2007, and a period of high turbulence – the late 1970s and early 1980s, when the fight against inflation forced the Fed to act with sharp, unpredictable methods.
Crucially, different shocks switched between regimes asynchronously. A price shock could be in a high-volatility phase while a production shock was in a calm one. This is exactly what the heterogeneous structure of the model predicts, and precisely what previous 'one-size-fits-all' models simply could not describe.
Third, the impulse response functions – that is, the economy's responses to an unexpected rate hike – turned out to be quite reasonable from an economic standpoint. A rise in the federal funds rate led to a decline in industrial production (with a delay of several months) and a subsequent slowdown in inflation. Nothing revolutionary in substance – but what's important is the precision and reliability of these estimates, achieved without theoretical constraints.
The Bayesian Kitchen: How This Calculation Is 'Cooked Up'
For those interested in the mechanics: the entire estimation is done using the Bayesian method with what is called a Gibbs sampler. This is an iterative algorithm that successively 'draws' samples from the distributions of each parameter while holding the others fixed. Step by step, iteration by iteration, the algorithm explores the space of possible values and gradually converges to the posterior distribution – that is, to what the data tells us about the model's parameters, given our prior assumptions.
The key innovation here is a specially chosen prior distribution for the variance parameters. This distribution is constructed in such a way that, on the one hand, it allows for any number of active regimes, and on the other, it 'penalizes' regimes that the data do not support. The result: the algorithm itself 'prunes' the unnecessary regimes, leaving only those that are truly needed.
The authors also note that their model's forecasting performance is comparable to that of stochastic volatility models – a powerful and computationally expensive tool traditionally considered the benchmark in forecasting financial and macroeconomic series. This is significant: the new model not only explains the past but also predicts the future no worse than its competitors.
Why This Matters Beyond Econometrics
Allow me to pause for a moment and ask the question that a student in the front row invariably asks me: 'Professor, who needs this besides academics?'
The answer is simple. Central banks make decisions about interest rates based, in part, on quantitative models. If a model misidentifies which shock is behind the observed changes – monetary or, say, a price shock – then the policy response will be wrong. In the 1970s, part of the U.S. Federal Reserve's policy errors was explained by precisely this misdiagnosis of the nature of shocks: inflation was mistaken for a supply-side shock when it was monetary in nature, and vice versa.
More accurate shock identification is not an academic game. It is a tool that ultimately affects how precisely a central bank 'hits' its target. And that target – price stability and employment – concerns everyone who earns a salary or pays rent.
Moreover, the method of verifying heteroskedasticity proposed by the authors solves a problem that was previously swept under the rug: researchers often applied identification through heteroskedasticity without even checking if it was present. This is like using a compass in a place with no magnetic field: the instrument will point in some direction, but you shouldn't trust it. Now there's a way to first check if the compass is working.
Different Shocks, Different Biographies
There is a detail in this work that strikes me as particularly beautiful from the perspective of the history of economic thought. For a long time, economists working with multidimensional systems were forced to assume that all variables 'live in the same time' – that is, they switch between regimes synchronously. This is mathematically convenient but deeply implausible historically.
Look at the American economy of the 1970s. The oil embargoes of 1973 and 1979 created price turbulence that was not necessarily accompanied by the same turbulence in production or monetary policy – at least, not in the same months or with the same intensity. Economic history is full of examples where different 'stories' unfolded at different speeds and with different rhythms.
The SHMSH model simply allows the data to tell us exactly this – that each shock has its own biography, its own rhythm of rising and falling. And this, in my view, is not just technical progress. It is a small step toward making mathematical models a little more like the real world – with its asynchronicity, its unpredictability, and its infinite variety of rhythms.
Because the global economy is not a monolith, marching in a single step. It is a polyphony, where each voice experiences its own storms and calms. And hearing them separately is a task we are only just beginning to truly tackle.
