Macroeconomic Forecasting with VAR

Welcome! It's fantastic to have you here, especially with your background at the IMF. Since you have already used Vector Autoregressions (VARs), you know they are the workhorses of quantitative macroeconomics, but mastering the underlying machinery is what makes you stand out in rigorous policy and research interviews.

Before we build your personalized roadmap, let's decide how deep you want to go. I want to make sure we strike the perfect balance between the high-level intuition you need to explain things clearly to a panel and the mathematical rigor behind estimation and identification.

Let's choose the depth of coverage you'd like for this course.

An interview-focused review is a brilliant choice. It ensures we prioritize the high-yield theoretical intuition and structural identification logic that interview panels love to test, without getting bogged down in tedious algebraic proofs.

Since you have already used these models at the IMF, you know they are incredibly powerful. To kick off our first milestone—Foundations of VAR—we will establish a solid mathematical baseline that you can easily sketch on a whiteboard during an interview.

Let's start with a classic bivariate structural system involving inflation ($π_t$) and a policy rate ($i_t$). This is a reduced-form model where each variable is expressed purely as a function of its own history and the other variable's history.

In an interview, you want to show you can think in terms of matrix algebra immediately. If we collect our variables into a vector $Y_t = \begin{bmatrix} \pi_t \\ i_t \end{bmatrix}$, we can compress this system into a clean, professional matrix notation.

Understanding this transition from a system of individual equations to a compact matrix format is crucial because it sets up how we analyze stability and lag structures using lag operators.

Correct

Spot on. Your intuition is completely sharp—since all the variables on the right-hand side of a reduced-form model are strictly lagged, $\phi_{12}$ must capture the dynamic, time-delayed effect of the policy rate on inflation rather than any contemporaneous interaction.

Now, let's level up our notation. To comfortably discuss VAR stability, unit roots, and structural identification in an interview, you need to be highly fluent in the language of the lag operator ($L$). This algebraic tool allows us to manipulate time series vectors as if they were simple polynomials.

Let's apply the lag operator to our matrix-form VAR(1). Since $Y_{t-1} = L Y_t$, we can rewrite the system $Y_t = C + \Phi_1 Y_{t-1} + \epsilon_t$ by pulling all the terms involving $Y$ to the left-hand side:

By factoring out $Y_t$ on the left side, we can define a matrix polynomial in the lag operator, denoted as $\Phi(L)$:

This representation generalizes beautifully to any lag length $p$. For a general VAR($p$) system, the matrix lag polynomial is written as:

Once you can fluidly write down this polynomial notation, you are fully equipped to explain the stationarity and stability conditions that interviewers frequently use to test your foundational time-series knowledge.

Don't worry at all—this is exactly the kind of structural plumbing that is easy to glaze over when using software, but incredibly satisfying to lock down for a whiteboard presentation. Let's break down the exact algebra of how we transition between those two states.

The Algebra of Factoring Out

Let's first look at simple scalar algebra as an analogy. If you have the expression $x - a x$, you can factor out the $x$ to get $(1 - a)x$. We are doing the exact same thing here, but because we are dealing with vectors and matrices, we cannot just use the number $1$. Instead, we use the identity matrix ($I_n$).

To see this step-by-step, let's look at the "before" and "after" states of our left-hand side:

Because matrix multiplication is distributive, you can easily verify this by multiplying $Y_t$ back through the parentheses from the right side:

Taking the time to visualize how these matrices line up prevents simple mistakes during technical interviews when a panel asks you to write down the system's characteristic polynomial.

Not quite

Actually, multiplying a vector by the identity matrix $I_n$ does not change its values at all—just like multiplying a number by $1$ keeps it exactly the same. Let's look at why $I_n$ is mathematically required here by looking at the actual matrix dimensions.

Recall that $\Phi_1$ is a $2 \times 2$ matrix. In matrix algebra, you can only add or subtract matrices if they have the exact same dimensions. If you tried to write $(1 - \Phi_1 L)$, you would be trying to subtract a $2 \times 2$ matrix from a single scalar number, $1$. That is a mathematical mismatch—like trying to subtract a spreadsheet from a single cell!

By using the identity matrix $I_n$ (which is a $2 \times 2$ matrix filled with ones on the diagonal and zeros elsewhere), the dimensions match perfectly, allowing us to perform the subtraction:

Now that we have solid footing on this algebraic notation, let's make sure you feel completely comfortable. Do you feel ready to wrap up this first milestone on Foundations of VAR and move on to our next milestone, Stationarity & Stability?

You are completely right—at its core, this is pure algebra, just moving terms from one side of the equals sign to the other. Let's strip away any academic jargon and look at exactly why we do this shifting, step-by-step.

Why Do We Move Things Around

In applied macro forecasting, you want to solve for current variables ($Y_t$) using past data ($Y_{t-1}$). But in a theoretical or policy interview, the panel wants to know: Is this economic system stable, or will it explode?

To analyze stability, we have to treat the VAR as a dynamic system. By moving all the endogenous variable terms ($Y$) to the left-hand side and leaving only the random shocks ($\\epsilon_t$) and constants ($C$) on the right, we isolate the internal feedback loop of the economy on one side. This lets us solve for the "roots" of the system, which tell us if the variables will eventually return to equilibrium after a shock.

Step-by-Step Whiteboard Walkthrough

Let's trace the physical movement of the algebra on a whiteboard, step-by-step.

Once we have grouped the terms inside the parentheses, we give that whole block a collective name, $\Phi(L) = I_n - \Phi_1 L$. This is the characteristic polynomial of the VAR, and its properties completely dictate whether the system is stable.

Seeing the math laid out this way helps show that we are not introducing new economic information; we are simply rearranging the furniture so we can easily inspect the system's structural stability.

The ultimate takeaway is this: by using the lag operator $L$ and the identity matrix $I_n$, we can pack a complex system of multiple macro equations into a single, compact expression, $(I_n - \Phi_1 L) Y_t = C + \epsilon_t$.

This algebraic rearrangement isolates the internal feedback loop of the economy into the matrix polynomial term $(I_n - \Phi_1 L)$, which serves as the "DNA" of the system. This allows us to analyze whether the entire economy is stable or explosive by studying just this single term, rather than trying to track every individual equation simultaneously.

Now that we have locked down this foundational notation, let's advance to our second milestone: Stationarity & Stability.

Defining Stability

In a policy interview, a panel will often ask you to define what it means for a VAR system to be "stable" or "covariance stationary". These terms are closely related, but they mean slightly different things to a theoretician.

Intuitively, a VAR is stable if the consequences of any temporary economic shock (like a sudden policy rate hike or an oil price spike) gradually die out over time, allowing the system to return to its long-run mean.

If a shock causes the system's variables to swing wilder and wilder or permanently drift away without bound, the system is explosive (unstable). For a simple scalar autoregression ($n=1$), $y_t = \phi_1 y_{t-1} + \epsilon_t$, we know the system is stable if and only if the absolute value of the coefficient is less than one:

To evaluate this joint dynamic feedback holistically, we must look at the eigenvalues of our transition matrix, which we will jump into next.

That is the ultimate practical question. In macro forecasting and policy analysis, stability isn't just a mathematical convenience—it completely changes how we interpret the economy's behavior and make projections.

There are three core reasons why we care deeply about stability when building VAR models:

Why Stability Matters

1. Mean Reversion and Forecasting: If a system is stable, its forecasts will eventually converge back to the unconditional mean of the variables as the forecast horizon goes to infinity. If the system is unstable (or contains a unit root, like a random walk), your long-run forecasts will exhibit infinite variance, making long-term projections highly uncertain and sensitive to recent noise.

2. Economic Logic (Economic Shocks): Economists use VARs to run "what-if" policy experiments. We want to see how the economy responds to a temporary policy shock—for instance, a sudden interest rate hike. If the system is stable, this shock eventually wears off, and the economy settles back to normal. If it is unstable, a single interest rate hike would trigger a run-away feedback loop, sending inflation or interest rates to infinity. That is economically implausible for most stable policy regimes.

3. Mathematical Validity of Policy Tools: The entire suite of modern policy analysis tools—specifically, Impulse Response Functions (IRFs) and Forecast Error Variance Decompositions (FEVDs)—relies on being able to invert the VAR system into an infinite Moving Average ($VMA(\infty)$) representation. This mathematical conversion is only valid if the system is stable.

Understanding this economic motivation makes the mathematical machinery we are about to look at highly intuitive—we are simply checking if our model's math mirrors this real-world economic gravity.