Act I · One different shock

What changes relative to the path we would otherwise have seen?

Where we are: the same system, a new question

C02 · Module 11’s connected autoregressive system

𝐘t=𝐀0+𝐀1𝐘t−1+𝛜t\mathbf Y_t=\mathbf A_0+\mathbf A_1\mathbf Y_{t-1}+\boldsymbol\epsilon_t

Mixing-board channel Module 12
Own and cross-variable memory Keep the VAR dynamics
Current connections State the shock’s identification
New question Trace one extra shock through the system
The object is a path of deviations from baseline.

Which coefficient tells the whole story?

C03 · Recall Module 11
Direct structural cross-lag Reduced-form cross-lag
xt−1x_{t-1} in the yty_t equation: 0 xt−1x_{t-1} in the yty_t equation: 8/35

The second coefficient combines connections in the system.

What happens now, next period, and after that when one disturbance changes?

Change one shock, once

C04 · Two copies of the same model
Before tt At tt After tt
Same history Add one specified shock to one copy Same subsequent innovations

The other copy is the baseline: what happens without that extra shock.

Does changing a disturbance mean holding the observed variable at a new level forever?

Two paths—and their difference

C05 · Known Module 11 system · default: +1 x structural-shock SD · periods
Static figure

The shocked y level is above its flat baseline; the difference decays toward zero.

One extra unit of the x structural disturbance; also one SD in this generator. Horizon 0 is impact.

The IRF is what remains after subtraction

C06 · Carry this identity through every figure
IRF = shocked path − baseline path
baseline + IRF = shocked path

𝐫h(j)(δ)=𝐘t+hshock−𝐘t+hbase\mathbf r_h^{(j)}(\delta)=\mathbf Y_{t+h}^{\mathrm{shock}}-\mathbf Y_{t+h}^{\mathrm{base}}

The IRF alone gives the deviation. It does not reveal the baseline.

A funds-rate response is relative

C07 · Hypothetical illustration · defined inflation shock · h = 4

Suppose the IRF of the funds rate is +0.5 percentage point.

Baseline rate IRF Rate with the extra shock
2.0% +0.5 pp 2.5%
4.0% +0.5 pp 4.5%
The same response is compatible with different rate levels.

“At rest” is a useful starting picture

C08 · The comparison can also have a moving baseline

At the system mean, with zero later innovations, our baseline is flat.

With a different history or common later innovations, it can move.

Could the rate rise over time and still have a negative response to the extra shock?

Yes—if it rises less than the baseline would have risen.

Act II · Today’s impact becomes tomorrow’s history

Where does the response come from when there is no new pulse?

Return to the system we know

C09 · Simulation gives us the structural truth

yt=0.2+0.5xt+0.3yt−1+0xt−1+ηy,t,xt=0.1+0.25yt+0.1yt−1+0.4xt−1+ηx,t.\begin{aligned} y_t&=0.2+0.5x_t+0.3y_{t-1}+0x_{t-1}+\eta_{y,t},\\ x_t&=0.1+0.25y_t+0.1y_{t-1}+0.4x_{t-1}+\eta_{x,t}. \end{aligned}

Add one unit to ηx,t\eta_{x,t}, holding the other disturbance fixed.

Here Var(𝛈t)=𝐈\operatorname{Var}(\boldsymbol\eta_t)=\mathbf I: one unit also equals one structural-shock SD.

Impact is a simultaneous solution

C10 · Subtract the baseline equations at h = 0

ry,0=0.5rx,0,rx,0=0.25ry,0+1r_{y,0}=0.5r_{x,0},\qquad r_{x,0}=0.25r_{y,0}+1

𝐫0=(4/78/7)≈(0.57141.1429)\mathbf r_0=\begin{pmatrix}4/7\\8/7\end{pmatrix} \approx\begin{pmatrix}0.5714\\1.1429\end{pmatrix}

Why does y move even though its own structural disturbance did not change?

One period later, the pulse is gone

C11 · Yesterday’s responses enter today’s equations

ry,1=25ry,0+835rx,0=2449≈0.4898,rx,1=15ry,0+1635rx,0=156245≈0.6367.\begin{aligned} r_{y,1}&=\frac25r_{y,0}+\frac8{35}r_{x,0}=\frac{24}{49}\approx0.4898,\\ r_{x,1}&=\frac15r_{y,0}+\frac{16}{35}r_{x,0}=\frac{156}{245}\approx0.6367. \end{aligned}

There is no new +1. What keeps the response alive?

The same rule applies again

C12 · Carry the previous response forward
Horizon h y deviation x deviation
0 0.5714 1.1429
1 0.4898 0.6367
2 0.3415 0.3890

(ry,2rx,2)=(2/58/351/516/35)(ry,1rx,1)\begin{pmatrix}r_{y,2}\\r_{x,2}\end{pmatrix} =\begin{pmatrix}2/5&8/35\\1/5&16/35\end{pmatrix} \begin{pmatrix}r_{y,1}\\r_{x,1}\end{pmatrix}

Matrices record the rule compactly

C13 · Impact first; propagation after that

𝐫0(j)(δ)=𝐁−1𝐪jδ\mathbf r_0^{(j)}(\delta)=\mathbf B^{-1}\mathbf q_j\delta 𝐫h+1(j)(δ)=𝐀1𝐫h(j)(δ)\mathbf r_{h+1}^{(j)}(\delta)=\mathbf A_1\mathbf r_h^{(j)}(\delta) 𝐫h(j)(δ)=𝐀1h⏟propagation𝐁−1𝐪jδ⏟impact\mathbf r_h^{(j)}(\delta)=\underbrace{\mathbf A_1^h}_{\text{propagation}} \underbrace{\mathbf B^{-1}\mathbf q_j\delta}_{\text{impact}}

The moving-average view collects the paths

C14 · Recall the univariate AR-to-MA representation

𝐘t−𝛍=∑h=0∞𝚽h𝛜t−h,𝚽h=𝐀1h\mathbf Y_t-\boldsymbol\mu =\sum_{h=0}^{\infty}\mathbf\Phi_h\boldsymbol\epsilon_{t-h}, \qquad\mathbf\Phi_h=\mathbf A_1^h

For standardized structural shocks:

𝚯h=𝚽h𝐁−1𝛀1/2\mathbf\Theta_h=\mathbf\Phi_h\mathbf B^{-1}\boldsymbol\Omega^{1/2}

Column jj of 𝚯h\mathbf\Theta_h gives the response to one structural-shock SD.

Let R repeat the known rule

C15 · Live code · same known structural shock
r_live <- matrix(0, H + 1, 2)
r_live[1, ] <- B_inverse %*% c(0, 1)  # h = 0
for (h in 1:H) {
  r_live[h + 1, ] <- A1 %*% r_live[h, ]
}
Which line changes if we choose another shock? Which line propagates it?

Act III · Read the experiment, then the curve

A curve needs a shock, a scale, a clock—and a baseline interpretation.

Read the caption before the shape

C16 · y response to +1 unit of eta-x · periods · noncumulative
The y deviation is 0.5714 at impact and decays toward zero over twenty periods.
Every point answers: how far from baseline at this horizon?

One unit, one SD, and 100 basis points

C17 · Different labels can define different experiments
Shock label What is fixed?
One unit The disturbance’s numerical increment
One shock SD Its scale under the stated identification
−100 bp policy impact A rate decrease of one percentage point
Does one SD of a reduced innovation equal one SD of a structural shock?

At a horizon—or accumulated through it?

C18 · Same +1-unit x structural shock · y response
The point response falls toward zero while the cumulative sum approaches a positive value.

Ry,h=∑s=0hry,sR_{y,h}=\sum_{s=0}^{h}r_{y,s}

The transformation decides the units

C19 · Check what entered the model before using cumulative = TRUE
Modeled variable Its point IRF Its cumulative IRF
Level yy Level-unit deviation Sum of level-unit deviations
Difference Δy\Delta y Difference in changes Level deviation, if initial levels match
Log difference Δlogy\Delta\log y Difference in log growth Log-level deviation
A response in level units is still a deviation, not the actual level.

What does linearity promise?

C20 · Predictions from the model

𝐫h(j)(−δ)=−𝐫h(j)(δ),𝐫h(j)(2δ)=2𝐫h(j)(δ)\mathbf r_h^{(j)}(-\delta)=-\mathbf r_h^{(j)}(\delta),\qquad \mathbf r_h^{(j)}(2\delta)=2\mathbf r_h^{(j)}(\delta)

Before changing the widget: what should −1 SD and +2 SD do to the entire path?

Reverse the sign. Double the deviation at every horizon.

Stability erases the extra shock’s footprint

C21 · A stable VAR forgets a one-time pulse

𝐫h(j)(δ)→𝟎\mathbf r_h^{(j)}(\delta)\longrightarrow\mathbf0

The shocked and baseline paths converge toward one another.

The observed series may keep moving as other shocks arrive.

If the cumulative response stays positive, must the point response stay positive forever?

Act IV · Identification changes the shock

Can the same reduced system support different response stories?

The covariance needs an economic interpretation

C22 · Hide the structural truth for a moment

𝛜t=𝐂𝐳t,Var(𝐳t)=𝐈,𝚺=𝐂𝐂′\boldsymbol\epsilon_t=\mathbf C\mathbf z_t,\qquad \operatorname{Var}(\mathbf z_t)=\mathbf I,\qquad \boldsymbol\Sigma=\mathbf C\mathbf C'

response matrices=𝚽h𝐂\text{response matrices}=\mathbf\Phi_h\mathbf C

A covariance factorization supplies a map. Restrictions give that map its economic meaning.

Change the restriction; keep the reduced system

C23 · Default: +1 SD of each model’s x-associated shock · choose size/source below
Static comparison

Different y response paths result from y-first, x-first, and the known simultaneous identification.

Same reduced dynamics and covariance. Both recursive zero restrictions are false for the known generator.

Both recursive restrictions are false here

C24 · We know because we built the generator
Assumption Excluded current connection Generating coefficient
yy first xt→ytx_t\to y_t 0.50
xx first yt→xty_t\to x_t 0.25

Both still reproduce the same reduced-form information.

Would an enormous sample make either zero restriction true?

What ortho asks R to compute

C25 · For an object returned by VAR()
Option Impact experiment Propagation
ortho = FALSE One-unit reduced innovation; other coordinates zero 𝚽̂h\widehat{\mathbf\Phi}_h
ortho = TRUE One-SD recursive shock, in the fitted order 𝚽̂h𝐏̂\widehat{\mathbf\Phi}_h\widehat{\mathbf P}

𝚺̂=𝐏̂𝐏̂′\widehat{\boldsymbol\Sigma}=\widehat{\mathbf P}\widehat{\mathbf P}'

Choose the experiment you intend. The switch does not validate an economic shock.

Make the choice explicit in code

C26 · Live R · Module 11 sample · n = 2000 · seed 637611
irf_reduced <- irf(fit, impulse = "x", response = c("y", "x"),
  n.ahead = H, ortho = FALSE, cumulative = FALSE, boot = FALSE)
irf_recursive <- irf(fit, impulse = "x", response = c("y", "x"),
  n.ahead = H, ortho = TRUE, cumulative = FALSE, boot = FALSE)
An identified SVAR() object already carries its impact map.
ortho = FALSE is not a generalized IRF. Optional notes explain that separate object.

Selecting the impulse does not reorder the model

C27 · The variable order belongs to the fitted system
fit_x_first <- VAR(var_data[, c("x", "y")], p = 1, type = "const")
irf_x_first <- irf(fit_x_first, impulse = "x", response = c("y", "x"),
  n.ahead = H, ortho = TRUE, cumulative = FALSE, boot = FALSE)
What changed: the reduced forecasting system, or the shock’s identifying assumption?

Act V · A curve is an estimate

How much of this apparent response is precisely estimated?

Bootstrap the response, not just one coefficient

C28 · Repeat the fitted-model experiment
Pseudo-data → refit VAR → recompute IRF

At each horizon, compare the responses across replications.

Sample residuals as whole vectors across equations.

Read size and precision separately

C29 · Estimated y response · one-SD y-first x shock · 95% pointwise band
The estimated response and band are zero at impact by restriction; the response rises at h1 and decays.
1,000 residual-vector bootstrap replications · seed 637612 · percentile intervals

Pointwise bands concern a selected horizon

C30 · Connecting intervals does not change their coverage claim

At a specified horizon: interpret its interval under the procedure.

Across the whole path: simultaneous coverage needs a different construction.

A response band is not a prediction interval for the future observed level.

Narrow bands do not validate the restriction

C31 · Sampling uncertainty is conditional on a model

The chosen variables, lag specification and identifying assumptions remain part of the claim.

If both recursive curves were estimated almost perfectly, which identification problem would remain?
Precision about a response under an assumption is not evidence that the assumption is true.

Act VI · Read Bianco as an applied researcher

What does a policy easing do to credit creation and destruction?

Net credit can hide gross movements

C32 · Bianco (2021), Monetary policy and credit flows

Some firms expand debt while others contract it.

The net change can conceal both movements.

Could monetary policy affect credit creation and destruction differently?

Know what the model measures

C33 · Quarterly Compustat · 1974Q1–2017Q1
Object Meaning
Credit creation Debt-weighted positive firm debt growth
Credit destruction Debt-weighted absolute negative firm debt growth
Macro controls GDP, unemployment, consumer prices
Policy rate Effective funds rate; Wu–Xia substitute at the lower bound
A rise in destruction means more credit contraction.

Translate the ordering into an impact claim

C34 · Rows: responses · columns: shocks · same ordered labels
  1. GDP
  2. Unemployment
  3. Consumer prices
  4. Policy rate
  5. Credit destruction
  6. Credit creation

𝐂=(*00000**0000***000****00*****0******)\mathbf C=\begin{pmatrix} *&0&0&0&0&0\\ *&*&0&0&0&0\\ *&*&*&0&0&0\\ *&*&*&*&0&0\\ *&*&*&*&*&0\\ *&*&*&*&*&* \end{pmatrix}

Which variables may respond on impact to the policy shock?

The paper also restricts a lag coefficient

C35 · A separate restriction from contemporaneous ordering

The first lag of the policy rate is excluded from the price equation.

The restricted system is estimated by iterated SUR.

An unrestricted VAR plus ortho = TRUE does not reproduce the full specification.

Read Figure 4’s experiment first

C36 · The caption is part of the result
Feature Figure 4
Shock One-time unexpected −100 bp policy easing
Clock Quarters; 0 = impact
Dynamics Two lags, selected by BIC
Bands 68%, mean-bias-corrected residual wild bootstrap
Every panel is a departure from the path without the added policy shock.

Output and unemployment respond over time

C37 · −100 bp policy easing · quarters · 68% bands
Published GDP and unemployment deviations after a 100-basis-point policy easing. GDP rises with a lag; unemployment falls. Original axes and 68% bands are retained.

Published GDP and unemployment IRFs. Original axes and bands retained.

Describe direction, timing and persistence relative to baseline.
Bianco (2021), Figure 4, p. 12 · Qualitative reading; retain source scaling.

The policy response is not a rate level

C38 · Consumer prices and policy rate · −100 bp easing · quarters · 68% bands
Published consumer-price and policy-rate deviations after a 100-basis-point easing. The policy-rate response is normalized to minus one percentage point at impact and later moves toward zero. Original axes and 68% bands are retained.

Published consumer-price and monetary-policy-rate IRFs, with their original scaling.

At impact: the rate is 1 percentage point below its baseline.
Bianco (2021), Figure 4, p. 12

Credit creation changes sign

C39 · Credit destruction and creation · percentage-point deviations · 68% bands
Published credit-destruction and credit-creation deviations after a 100-basis-point easing, in percentage points with 68% bands. Creation initially falls, turns positive by quarter six, and peaks around quarter eight.

Published credit-destruction and credit-creation IRFs.

Credit creation: −0.11 pp at impact → positive by Q6 → about +0.15 pp at Q8.

−100 bp policy easing · horizons in quarters · Bianco (2021), §4.2 p. 11 and Figure 4 p. 12

State the claim with its comparison attached

C40 · A response report needs more than a sign

Under [identifying restrictions], a [size and sign] shock produces a [receiving variable] deviation from baseline of [magnitude and units] at [horizon], with [uncertainty].

Complete the sentence for credit creation. Then name one assumption the figure cannot verify.

Keep the baseline in every interpretation

C41 · Key takeaways + next time
IRF = shocked path − baseline path

Specify the shock. Trace its propagation. Report the units, assumptions and uncertainty.

The curve alone does not tell you the baseline or actual level.

Next: FEVD asks how much shocks matter for forecast uncertainty. Module 13 also returns to system testing and cointegration/VECMs.