Structural VAR Identification

A raw, fitted VAR (vector autoregression) gives you residuals — the leftover surprise in each variable each period, after accounting for its own recent history. The trouble is these residuals are usually correlated with each other: a surprise in oil prices and a surprise in interest rates might move together simply because news days often carry both kinds of information at once. If you want to ask "what happens to equities specifically because of an oil shock," you can't just use the raw, correlated residual — it's contaminated with whatever else happened that day. Structural VAR identification is the set of assumptions you add to disentangle these tangled residuals into genuinely separate, economically interpretable shocks, each one you can shock in isolation.

An analogy: untangling two microphones

Imagine two microphones in a room, each picking up a mix of two people talking at once — every recording is some blend of both voices. You can't tell from either recording alone "what did person A say" versus "what did person B say," because both signals are mixed into both microphones. To separate them, you need extra information: maybe you know person A is always louder in mic 1, or that person A never speaks at certain times. Structural VAR identification is exactly this kind of extra information, applied to economic shocks instead of voices — some assumption (a timing restriction, a sign restriction, an economic theory) that lets you unmix the tangled statistical residuals into individually meaningful "who spoke" shocks.

The mechanics, one symbol at a time

The plain (reduced-form) VAR residuals $u_t$ relate to the true structural shocks $\varepsilon_t$ through a mixing matrix $B$:

$$u_t = B\, \varepsilon_t,$$

where $u_t$ is the vector of statistically estimated, correlated residuals (what you actually observe from fitting the VAR), and $\varepsilon_t$ is the vector of structural shocks you actually want — mutually uncorrelated, each with a clean economic meaning (a "monetary policy shock," an "oil supply shock"). $B$ describes how each structural shock instantaneously spreads across all the observed variables. In plain English: the messy, correlated residuals you can estimate directly are some linear mixture of the clean, independent shocks you actually care about, and $B$ is the unmixing recipe. The problem is that fitting the reduced-form VAR alone only pins down the covariance of $u_t$, which isn't enough information to uniquely recover $B$ — there are infinitely many $B$ matrices consistent with the same observed covariance. Identification means imposing enough extra restrictions (short-run timing restrictions, long-run restrictions, or sign restrictions) to narrow this down to exactly one valid $B$.

Worked example 1: Cholesky (recursive) identification

The simplest scheme orders variables and assumes variable 1 reacts to no other variable within the period, variable 2 reacts only to variable 1, and so on — a "one-way street" of same-period causality. Suppose the ordering is oil price → interest rate → equity return, and the estimated (lower-triangular) $B$ matrix gives coefficients: a 1-unit oil shock moves the interest rate residual by 0.10 and the equity residual by −0.05 within the same period, while an interest-rate shock (after removing the oil-driven part) moves equities by −0.20, and an equity shock affects nothing else that period by assumption. If oil jumps by 2 standard deviations (a shock of size 2), the same-period equity impact via the oil channel is $2 \times (-0.05) = -0.10$, separate from — and addable to — the direct interest-rate channel. This is the exact scheme underlying most textbook impulse response plots.

Worked example 2: why ordering matters

Take the same three variables but swap the order to equity return → interest rate → oil price. Because Cholesky identification forces the first variable to react to nothing else within the period, equities are now assumed unaffected same-day by either rate or oil shocks — the opposite assumption from before. Re-estimating with this ordering, a rate shock's same-day effect on equities (previously −0.20) becomes exactly 0 by construction, and all of that comovement gets reattributed elsewhere. Two analysts using the same data but different orderings can report contradictory same-day effects, not because the data changed, but because the identifying assumption did — which is why the choice of ordering (or a better identification scheme) must be defended on economic grounds, not left as a coding default.

What this means in practice

Structural VAR identification is what separates a purely statistical VAR (good for forecasting) from one that supports causal claims like "monetary policy tightening cost equities X% within a week." Rates and macro desks rely on it to attribute market moves to specific shock types and to build impulse responses (see Impulse Response Functions) that can actually justify a trading or hedging decision, rather than just describing correlation.