Machine Learning the Macroeconomic Effects of Financial Shocks

Florian Huber; Karin Klieber; Massimiliano Marcellino; Niko Hauzenberger

REVIEW 3 major objections 4 minor 18 references

Machine Learning the Macroeconomic Effects of Financial Shocks

T0 review · 3 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that US financial shocks transmit to the real economy through strong sign asymmetry — adverse shocks matter much more than benign ones of equal size — while responses scale almost proportionally with shock size.

desk verdict Plausible empirical results built on an unvalidated sequential estimator; deserves peer review but with a demand for validation. read the letter →

arxiv 2412.07649 v1 pith:IM2HNWPU submitted 2024-12-10 econ.EM

Niko Hauzenberger , Florian Huber , Karin Klieber , Massimiliano Marcellino This is my paper · ORCID

classification econ.EM MSC 62F1568T0791B84

keywords BayesianneuralnetworksnonlinearlocalprojectionsfinancialshocksexcessbondpremiumsignasymmetriesimpulseresponsesUSmacroeconomyshocksizeproportionality

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper develops a way to estimate nonlinear impulse responses to structural shocks by replacing the linear regressions of standard local projections with Bayesian neural networks. Applied to US financial shocks measured through the excess bond premium, the method yields a specific empirical claim: adverse financial shocks push inflation, industrial production, and employment down sharply, while positive shocks of the same size produce little or no response. The paper also claims that the response scales almost linearly with shock size: a three-unit contractionary shock generates about three times the response of a one-unit shock. If true, these results would show that sign asymmetry, rather than size nonlinearity, is the dominant form of nonlinearity in the transmission of financial shocks to the real US economy.

What carries the argument

The machinery is a Bayesian neural network (a flexible regression model in which the conditional mean is a neural network with horseshoe shrinkage priors and a mixture of activation functions) embedded in nonlinear local projections. For each horizon h, the h-period-ahead outcome is regressed on the shock, controls, and a learned nonlinear function f_h; responses are computed as the difference between expected outcomes under shock τ and under zero, averaged over R=400 randomly drawn histories. Because the intermediate shocks ϵ_{t+h} are latent, the paper uses a sequential updating scheme: estimate the h=0 regression, save the posterior of the residuals, treat those draws as fixed regressors when estimating h=1, and continue feeding previous-horizon shock draws forward to h=2, ..., H. The sequence of horizon-specific nonlinear functions is what lets the response curves bend differently for different signs and sizes.

What would settle it

Run the same BNN local projections on simulated data from a known nonlinear model with known true shocks and known asymmetric responses; if the sequential estimator's horizon-h responses deviate systematically from the truth, the empirical conclusions about sign asymmetry would be called into question. Alternatively, re-estimate the responses using an external-instrument financial shock series; if the sign asymmetry disappears, the finding is an artifact of the recursive identification.

Watch

Extended reading notes

Core claim

The central discovery, on the paper's own terms, is that the nonlinear impulse responses of US macroeconomic aggregates to financial shocks are strongly asymmetric with respect to sign but approximately proportional with respect to size. A one-unit contractionary shock to the excess bond premium reduces inflation by around one percentage point at its one-month peak, cuts industrial production growth by almost two percentage points around eleven months out, and lowers employment growth for about two years, whereas a benign one-unit shock is mostly insignificant across all three variables. When the contractionary shock is tripled to three units, the responses line up with the one-unit responses after rescaling by one third, indicating no discernible size asymmetry. These patterns are obtained without imposing a particular parametric nonlinear functional form, since the conditional mean is learned by the network.

Load-bearing premise

The results stand or fall on the sequential shock-updating estimator: the paper treats the residuals from the zero-horizon network as the true shocks and feeds them forward as fixed regressors at every later horizon, an assumption it asserts but does not prove or validate by simulation.

Editorial extensions

If this is right

Policymakers reading the result as structural would expect contractionary financial shocks to be the dominant risk to prices and employment, with symmetric positive shocks providing little offsetting stimulus.
Linear local projections, by construction, would force positive and negative shocks to have mirror-image effects; the paper's finding implies such models understate the cost of adverse financial conditions.
The proportionality result implies that, within the range studied, scaling a shock is equivalent to scaling the response, so no separate nonlinearity in shock size needs to be built into forecasting or stress-testing models.
The method offers a template for estimating nonlinear impulse responses without choosing a parametric nonlinear functional form in advance.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the sequential shock-updating step is the fragile link; a Monte Carlo study with a known data-generating process would show whether the h=0 residuals remain valid controls once the nonlinear function f_h changes with the horizon.
Editorial inference: because the shock is identified by ordering the excess bond premium first in a recursive VAR, the sign-asymmetry result is conditional on that identification; an external-instrument or high-frequency proxy robustness check would tell whether the asymmetry is an artifact of the ordering.
Editorial inference: the proportionality finding could be tested at more extreme shock sizes (beyond three units), where financial frictions might make large adverse shocks disproportionately damaging.
Editorial inference: the size-proportionality result suggests a separable structure — nonlinearity in sign but not in scale — that a parsimonious semi-parametric model could reproduce; whether the BNN's flexible fit actually matches such a functional form is testable.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Here's my read of Hauzenberger et al. (arXiv:2412.07649). The headline: the empirical claims are plausible and consistent with prior work, but the sequential latent-shock estimator in Section 3 is the load-bearing part and it is not validated. I think the paper deserves peer review, but only with the expectation of a serious methods check.

What's actually new: applying their BNN framework to nonlinear local projections, and a sequential scheme that propagates posterior draws of residuals forward as controls. The sign-asymmetry result repeats Barnichon et al. (2022) and Forni et al. (2024); the size-proportionality finding is the genuinely new empirical result. The paper is clearly written and the BNN machinery is well documented in a companion paper, so the econometrics is not opaque.

The soft spot is real. Equation (4) includes ε_{t+h} as regressors, and the paper constructs them by saving residuals from the h=0 BNN and plugging them into later horizons. No consistency proof, no Monte Carlo, no bootstrap. The stress-test concern is legitimate: with horseshoe shrinkage, the h=0 residual is not orthogonal to the financial shock ζ_t in finite samples, so for h=1 the estimated response can be partly attributed to the residual rather than ζ_t. Since adverse shocks are large and persistent, this bad-control mechanism could plausibly generate sign asymmetry even in a symmetric DGP. That is not a narrow point about standard errors; it goes to the center of the empirical claim.

I also note the first-stage VAR that produces the financial shock is estimated separately, and its uncertainty is never propagated. That is a lesser but real issue. No code or data are provided, which makes independent verification impossible. The asymmetry inference is visual, from 68% credible intervals, not a formal test. These are not fatal individually, but together they mean the paper's main result rests on an unproven estimator.

For all that, the paper is not a throwaway. If the estimator is valid, the size-proportionality result is a useful calibration target, and the BNN-LP framework is likely to be adopted by others. I'd send it to a referee who knows local projections and Bayesian nonparametrics, and ask for a Monte Carlo exercise and a consistency argument for the sequential update.

Bottom line: worth a serious referee, but I wouldn't bet the farm on the empirical results until the estimator is checked. I'd bring it to a reading group to discuss the bad-control problem.

Referee Report

3 major / 4 minor

Summary. The paper develops Bayesian neural network (BNN) nonlinear local projections for structural impulse responses. Equation (4) regresses y_{t+h} on the shock ζ_t, controls x_t, and latent past shocks ε_t,...,ε_{t+h-1}. Because the latent shocks are unobserved, the paper proposes a sequential procedure: estimate h=0, save posterior draws of ε_t, use them as fixed regressors for h=1, save ε_{t+1} draws, use them for h=2, and so on. The method is applied to monthly US data (1960-2020) with the excess bond premium (EBP) from a VAR ordered first as the financial shock. The main empirical finding is that contractionary financial shocks produce much larger declines in inflation, industrial production, and employment than benign shocks, while one- and three-unit contractionary shocks produce proportional responses.

Significance. The paper addresses an important question and proposes a flexible, prior-regularized alternative to parametric nonlinear local projections. A key strength is that the BNN posterior provides a full distribution of impulse responses rather than point estimates. If the sequential estimator is unbiased, the sign-asymmetry result is a useful contribution consistent with prior evidence. However, the central methodological innovation is not validated with either a consistency argument or Monte Carlo evidence, and the headline conclusions are drawn from informal comparisons of 68% credible intervals rather than formal tests. The significance of the empirical contribution therefore cannot be assessed until these gaps are addressed.

major comments (3)

[Section 3, Eq. (4) and the sequential latent-shock procedure] The sequential latent-shock estimator is the load-bearing component of the paper, but the manuscript provides no consistency proof, no asymptotic argument, and no Monte Carlo simulation supporting it. In Eq. (4), for h≥1 the regressor vector ε_{t+h}=(ε_t,...,ε_{t+h-1}) is generated from earlier horizons and then treated as observed. The h=0 residuals are obtained from a BNN with horseshoe shrinkage; they are not guaranteed to be independent of ζ_t conditional on x_t, because regularization and potential misspecification can leave part of the contemporaneous effect of ζ_t in ε_t. At h=1, including such a residual as a control can absorb part of the effect of ζ_t on y_{t+1}, biasing ψ_1. The problem compounds for h≥2, because the shocks used as controls are themselves residuals from biased earlier horizons. Footnote 3 cites Clark et al. (2024), but that reference concerns direct forecasts and does not address impulse-response bias from generated regressors in nonlinear local projections. Since every horizon h≥1 in Figure 1 depends on this procedure, the headline sign-asymmetry result could be an artifact of the bad-control mechanism. The authors should provide either a theorem with primitive conditions or, at a minimum, a Monte Carlo exercise using a DGP with symmetric responses to verify that the estimator does not manufacture asymmetry.
[Section 4, shock construction] The financial shock ζ_t is obtained from a first-stage structural VAR with EBP ordered first. The estimated shocks are then plugged into Eq. (4) as known regressors. The posterior intervals in Figure 1 therefore condition on the first-stage estimates and do not incorporate VAR estimation uncertainty or identification uncertainty. This can overstate precision, and it is especially relevant for the comparison of positive and negative shocks if the first stage is estimated imprecisely. The authors should propagate first-stage uncertainty, for example by drawing the VAR parameters jointly or by using a Bayesian VAR and passing posterior draws of ζ_t into the BNN estimation.
[Section 4, Figure 1] The central claims of sign asymmetry and size proportionality are based on visual inspection of posterior medians and 68% intervals. The paper does not report a posterior probability that the response to a negative one-unit shock differs from the negative of the response to a positive one-unit shock, nor a posterior probability that the rescaled three-unit response equals the one-unit response. Overlapping or non-overlapping intervals are not a formal test. Given that the abstract states these asymmetries as the main findings, the authors should provide explicit posterior summaries for the differences (or ratios) of responses across signs and sizes.

minor comments (4)

[Section 5] Section 5 contains typos: 'applies the method' should be 'apply the method' and 'ley' should be 'key'.
[Footnote 3] Footnote 3 states 'show that of ϵt+h only has a small impact'; it is missing a word and should likely read 'show that the inclusion of ϵt+h only has a small impact'.
[Figure 1 note] The note to Figure 1 is hard to follow because the negative-shock response is plotted after multiplication by -1; please state explicitly that the plotted green line is -NLP(h,-1) and that asymmetry is assessed by comparing NLP(h,+1) with -NLP(h,-1).
[Section 4, size asymmetries] Section 4 investigates size proportionality only for contractionary shocks; the conclusion that 'small and large shocks resulting in almost exactly proportional impulse responses' should be qualified to positive shocks or supplemented with evidence for benign shocks.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular step found: the estimated IRFs are outputs of the fitted BNN, not re-statements of its inputs; only minor methodological self-citations are present.

full rationale

No load-bearing circularity is present. Section 3 defines the nonlinear local projection as the difference between conditional expectations of Eq. (4) evaluated at ζt = τ and ζt = 0, with horizon-specific parameters ψh, γh, γ̃h and fh estimated separately. The reported sign and size asymmetries in Section 4 are computed by evaluating these estimated functions, not by plugging the fitted values back into the object being predicted. The latent-shock vector εt+h is constructed sequentially from saved residuals of the h = 0 regression, but this is an estimation algorithm with potential finite-sample bias, not a definitional identity: the saved residuals are regressors, while the response is the change in E(yt+h) with respect to ζt. The only self-citations are methodological (Hauzenberger et al. 2024 for the BNN prior and MCMC; Clark et al. 2024 in footnote 3 for the small effect of latent shocks on direct forecasts), and neither is invoked as a uniqueness theorem or as the source of the empirical conclusion. The paper makes in-sample structural estimates rather than out-of-sample forecasts, which limits the force of the word 'machine learning' but does not make the derivation circular.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The estimates depend on a recursive identification of the EBP shock, a fixed BNN architecture, and an unvalidated sequential algorithm for latent shocks. The ledger records the hand-chosen design parameters and domain assumptions that carry the results.

free parameters (5)

Number of hidden layers = 1
Chosen by hand; only one architecture is considered, so nonlinearity capacity is fixed rather than tuned.
Number of neurons per hidden layer = Q = K
Set equal to the number of covariates; no cross-validation or robustness checks are reported.
Candidate activation functions = leakyReLU, sigmoid, ReLU, tanh
The mixture set is fixed; the prior on the mixture indicator is uniform, and no alternative sets are tried.
Structural VAR lag order = not reported
The EBP shock series depends on the lag length used in the first-stage recursive VAR; the paper does not state it.
Shock sizes for asymmetry comparison = tau = 1 and tau = 3
The proportionality null is evaluated by rescaling the 3-unit response by 1/3, which presupposes linear scaling under the null rather than testing it formally.

assumptions (4)

domain assumption The recursive VAR with EBP ordered first identifies the EBP innovation as a structural financial shock.
Section 4: 'We order the EBP measure first and obtain the structural economic shock related to this variable, which can be interpreted as a financial shock.' If EBP reacts within the month to inflation, output, or employment, the shock is contaminated.
ad hoc to paper The BNN with one hidden layer and Q=K neurons can approximate the true conditional expectation without substantive misspecification.
Sections 2 and 4. The architecture is imported from Hauzenberger et al. (2024); no universal approximation argument, cross-validation, or sensitivity analysis is provided for this dataset.
ad hoc to paper The sequential latent-shock updating algorithm yields unbiased nonlinear local projections for h>=1.
Section 3, paragraph after Eq. (6). The h=0 residuals are saved and treated as fixed regressors in later horizons; no proof or simulation evidence establishes consistency.
standard math The monthly US series are stationary enough for local projection estimation.
Local projections as in Jorda (2005) assume stationarity or ergodicity; the paper uses levels of inflation and employment without unit-root or cointegration testing.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Machine Learning the Macroeconomic Effects of Financial Shocks." pith.science (2026). https://pith.science/paper/IM2HNWPU

@misc{pith2026241207649,
  author       = {Pith},
  title        = {Pith review of: Machine Learning the Macroeconomic Effects of Financial Shocks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IM2HNWPU}},
  note         = {Machine review of arXiv:2412.07649}
}

read the original abstract

We propose a method to learn the nonlinear impulse responses to structural shocks using neural networks, and apply it to uncover the effects of US financial shocks. The results reveal substantial asymmetries with respect to the sign of the shock. Adverse financial shocks have powerful effects on the US economy, while benign shocks trigger much smaller reactions. Instead, with respect to the size of the shocks, we find no discernible asymmetries.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 17 canonical work pages

[1]

Credit and economic activity: credit regimes and nonlinear propagation of shocks,

Balke, N. S. (2000): “Credit and economic activity: credit regimes and nonlinear propagation of shocks,” Review of Economics and Statistics , 82(2), 344–349

work page 2000
[2]

Theory ahead of measurement? assessing the nonlinear effects of financial market disruptions,

Barnichon, R., C. Matthes, and A. Ziegenbein (2016): “Theory ahead of measurement? assessing the nonlinear effects of financial market disruptions,” FRB Richmond Working Paper. (2022): “Are the effects of financial market disruptions big or small?,” Review of Economics and Statistics , 104(3), 557–570

work page 2016
[3]

Horseshoe regularisation for machine learning in complex and deep models,

Bhadra, A., J. Datta, Y. Li, and N. Polson(2020): “Horseshoe regularisation for machine learning in complex and deep models,” International Statistical Review , 88(2), 302–320

work page 2020
[4]

A macroeconomic model with a financial sector,

Brunnermeier, M. K., and Y. Sannikov (2014): “A macroeconomic model with a financial sector,” American Economic Review, 104(2), 379–421

work page 2014
[5]

Handling sparsity via the horseshoe,

Carvalho, C. M., N. G. Polson, and J. G. Scott (2009): “Handling sparsity via the horseshoe,” in Artificial intelligence and statistics , pp. 73–80. PMLR

work page 2009
[6]

In- vestigating Growth-at-Risk Using a Multicountry Non-parametric Quantile Factor Model,

Clark, T. E., F. Huber, G. Koop, M. Marcellino, and M. Pfarrhofer (2024): “In- vestigating Growth-at-Risk Using a Multicountry Non-parametric Quantile Factor Model,” Journal of Business & Economic Statistics , forthcoming

work page 2024
[7]

Nonlinear transmis- sion of financial shocks: Some new evidence,

Forni, M., L. Gambetti, N. Maffei-F accioli, and L. Sala (2024): “Nonlinear transmis- sion of financial shocks: Some new evidence,” Journal of Money, Credit and Banking , 56(1), 5–33

work page 2024
[8]

Model selection in Bayesian neural net- works via horseshoe priors,

Ghosh, S., J. Yao, and F. Doshi-Velez (2019): “Model selection in Bayesian neural net- works via horseshoe priors,” Journal of Machine Learning Research , 20(182), 1–46

work page 2019

Show all 18 references

[9]

Credit spreads and business cycle fluctuations,

Gilchrist, S., and E. Zakraj ˇsek (2012): “Credit spreads and business cycle fluctuations,” American Economic Review, 102(4), 1692–1720. Gonc ¸alves, S., A. M. Herrera, L. Kilian, and E. Pesavento(2024): “State-dependent local projections,” Journal of Econometrics , 105702

2012
[10]

Bayesian neural networks for macroeconomic analysis,

Hauzenberger, N., F. Huber, K. Klieber, and M. Marcellino (2024): “Bayesian neural networks for macroeconomic analysis,” Journal of Econometrics , forthcoming

2024
[11]

Local projections in unstable environments,

Inoue, A., B. Rossi, and Y. W ang (2024): “Local projections in unstable environments,” Journal of Econometrics , p. 105726. Jord`a, `O. (2005): “Estimation and inference of impulse responses by local projections,” Amer- ican Economic Review, 95(1), 161–182

2024
[12]

Ancillarity-sufficiency interweaving strategy (ASIS) for boosting MCMC estimation of stochastic volatility models,

Kastner, G., and S. Fr ¨uhwirth-Schnatter (2014): “Ancillarity-sufficiency interweaving strategy (ASIS) for boosting MCMC estimation of stochastic volatility models,” Computa- tional Statistics & Data Analysis , 76, 408–423

2014
[13]

L ¨utkepohl (2017): Structural vector autoregressive analysis

Kilian, L., and H. L ¨utkepohl (2017): Structural vector autoregressive analysis . Cambridge University Press

2017
[14]

Local projections, autocorrelation, and efficiency,

Lusompa, A. (2023): “Local projections, autocorrelation, and efficiency,” Quantitative Eco- nomics, 14(4), 1199–1220

2023
[15]

A simple sampler for the horseshoe estimator,

Makalic, E., and D. F. Schmidt (2015): “A simple sampler for the horseshoe estimator,” IEEE Signal Processing Letters , 23(1), 179–182

2015
[16]

FRED-MD: A monthly database for macroeconomic research,

McCracken, M. W., and S. Ng(2016): “FRED-MD: A monthly database for macroeconomic research,” Journal of Business & Economic Statistics , 34(4), 574–589

2016
[17]

Impulse response estimation via flexible local projec- tions,

Mumtaz, H., and M. Piffer (2022): “Impulse response estimation via flexible local projec- tions,” arXiv:2204.13150

2022 arXiv
[18]

MCMC using Hamiltonian dynamics,

Neal, R. M. (2011): “MCMC using Hamiltonian dynamics,” Handbook of Markov Chain Monte Carlo, 2(11), 2. 8

2011

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

Credit and economic activity: credit regimes and nonlinear propagation of shocks,

Balke, N. S. (2000): “Credit and economic activity: credit regimes and nonlinear propagation of shocks,” Review of Economics and Statistics , 82(2), 344–349

work page 2000

[2] [2]

Theory ahead of measurement? assessing the nonlinear effects of financial market disruptions,

Barnichon, R., C. Matthes, and A. Ziegenbein (2016): “Theory ahead of measurement? assessing the nonlinear effects of financial market disruptions,” FRB Richmond Working Paper. (2022): “Are the effects of financial market disruptions big or small?,” Review of Economics and Statistics , 104(3), 557–570

work page 2016

[3] [3]

Horseshoe regularisation for machine learning in complex and deep models,

Bhadra, A., J. Datta, Y. Li, and N. Polson(2020): “Horseshoe regularisation for machine learning in complex and deep models,” International Statistical Review , 88(2), 302–320

work page 2020

[4] [4]

A macroeconomic model with a financial sector,

Brunnermeier, M. K., and Y. Sannikov (2014): “A macroeconomic model with a financial sector,” American Economic Review, 104(2), 379–421

work page 2014

[5] [5]

Handling sparsity via the horseshoe,

Carvalho, C. M., N. G. Polson, and J. G. Scott (2009): “Handling sparsity via the horseshoe,” in Artificial intelligence and statistics , pp. 73–80. PMLR

work page 2009

[6] [6]

In- vestigating Growth-at-Risk Using a Multicountry Non-parametric Quantile Factor Model,

Clark, T. E., F. Huber, G. Koop, M. Marcellino, and M. Pfarrhofer (2024): “In- vestigating Growth-at-Risk Using a Multicountry Non-parametric Quantile Factor Model,” Journal of Business & Economic Statistics , forthcoming

work page 2024

[7] [7]

Nonlinear transmis- sion of financial shocks: Some new evidence,

Forni, M., L. Gambetti, N. Maffei-F accioli, and L. Sala (2024): “Nonlinear transmis- sion of financial shocks: Some new evidence,” Journal of Money, Credit and Banking , 56(1), 5–33

work page 2024

[8] [8]

Model selection in Bayesian neural net- works via horseshoe priors,

Ghosh, S., J. Yao, and F. Doshi-Velez (2019): “Model selection in Bayesian neural net- works via horseshoe priors,” Journal of Machine Learning Research , 20(182), 1–46

work page 2019