Assembling the evidence…
We establish theoretical results about the low frequency contamination (i.e., long memory effects) induced by general nonstationarity for estimates such as the sample autocovariance and the periodogram, and deduce consequences for heteroskedasticity and autocorrelation robust (HAR) inference. We present explicit expressions for the asymptotic bias of these estimates. We show theoretically that nonparametric smoothing over time is robust to low frequency contamination. Nonstationarity can have consequences for both the size and power of HAR tests. Under the null hypothesis there are larger size distortions than when data are stationary. Under the alternative hypothesis, existing LRV estimators tend to be inflated and HAR tests can exhibit dramatic power losses. Our theory indicates that long bandwidths or fixed- b HAR tests suffer more from low frequency contamination relative to HAR tests based on HAC estimators, whereas recently introduced double kernel HAC estimators do not suffer from this problem. We present second-order Edgeworth expansions under nonstationarity about the distribution of HAC and DK-HAC estimators and about the corresponding t -test in the regression model. The results show that the distortions in the rejection rates can be induced by time variation in the second moments even when there is no break in the mean.
We show that time-variation in the mean induces low frequency contamination, meaning that the sample autocovariance and the periodogram share features that are similar to those of a long memory series
the short-memory SLS analysis supports contamination, but the large-lag autocovariance interpretation needs an explicit lag-growth qualification
We show theoretically that nonparametric smoothing over time is robust to low frequency contamination.
local smoothing mitigates contamination asymptotically, subject to break-window exceptions and the limits of the stated one-sided bounds
Under the alternative hypothesis, existing LRV estimators tend to be inflated and HAR tests can exhibit dramatic power losses.
analytic and simulation evidence supports the possibility in specified forecast-comparison designs with a restricted mean-shift window
long bandwidths or fixed-b HAR tests suffer more from low frequency contamination relative to HAR tests based on HAC estimators
the comparative power result is established for the particular Diebold–Mariano mean-shift design rather than the full range of HAR applications
Under the null hypothesis there are larger size distortions than when data are stationary.
example-specific bias evidence and unchanged asymptotic error order do not establish a general stationary-versus-nonstationary rejection-rate comparison
Derived from the full evaluation — not a separate score.
Strengths
The analysis links distorted global dependence estimates to inflated long-run variance estimates and losses of test power. This connects the estimator-level results in Section 3 with the inference consequences in Sections 4–5.
↳ Sections 3–5; Theorems 3–7
Theorems 1–2 distinguish windows without breaks from those containing breaks near boundaries or midpoints. These cases explain both the protection offered by time smoothing and the circumstances in which local contamination persists.
↳ Section 3.2; Theorems 1–2
The discussion distinguishes stationary fixed-b refinements from behaviour under nonstationarity and acknowledges the null-error cost of double smoothing. Prior evidence on nonmonotonic power is treated as context rather than presented as a newly discovered phenomenon.
↳ Introduction; Sections 4.2 and 5.1
Limitations
The abstract and conclusion assert larger size distortions than under stationarity. Theorem 4 preserves the stationary asymptotic order, while Tables 1–2 do not provide the direct stationary rejection-rate comparison needed for that general statement.
↳ Abstract; Theorem 4; Tables 1–2; Section 6
Section 3.1's positive large-lag autocovariance interpretation does not specify how lags grow relative to sample size. The regression framing in the abstract also extends beyond the Gaussian location-model setting used for the displayed Edgeworth analysis.
↳ Abstract; Section 3.1; Section 4
Section 5.3 specifies T=400, whereas Table 3 labels T=200 without identifying whether this denotes the evaluation subsample. Table 2 also reports unprewhitened DK-HAC null rejection rates of 0.117 and 0.154 at a nominal 0.05, limiting interpretation of unadjusted power rankings.
↳ Section 5.3; Table 3 headers; Table 2, M3–M4
The main contribution is the connected account of how nonstationarity alters dependence estimates and propagates into robust inference. Explicit assumptions and break-window cases support that account, but they do not resolve the missing lag-growth qualification or extend the Gaussian location-model derivations to general regression settings. The categorical null-size statement receives less support than the contamination mechanism because the theory retains the stationary error order and the tables do not directly compare stationary rejection rates. The forecasting sample-size labels and unequal null rejection rates also limit how confidently the simulation power rankings can be interpreted.
Nabu’s assessment, alongside the field’s view.
Are you an author of this paper?
Strong4.0
Confidence mediumThe score reflects a connected extension from nonstationary dependence estimates to higher-order HAR inference and power behaviour. The remedy builds on an existing DK-HAC framework, and the broadest implications require narrower interpretation than the summaries suggest.
The formal process definition, assumptions, break-window cases, and targeted comparisons support assessment of the central argument. Missing lag-regime qualifications, the narrower Gaussian location-model derivation, ambiguous simulation sample-size labels, and unadjusted size–power comparisons limit execution.
Sections 2–5 provide a followable progression from contamination mechanisms to inferential consequences. The categorical null-size comparison in the abstract and conclusion goes beyond the conditional theoretical and simulation evidence.
The paper distinguishes its results from spurious-persistence, stationary Edgeworth, HAC, and fixed-b work and discusses competing procedures' size–power trade-offs. Practical limitations are integrated around the relevant results, although broader comparisons with alternative nonstationary-inference strategies remain limited.
Caveats4 of 4 checks
2 reported values do not reconcile, but the key claims do not rest on them.
No adjudicated transparency or research-conduct finding is established. Numerical reporting notes are recorded under Internal Coherence rather than interpreted as evidence of misconduct.
92 of 94 checkable references verified
95 references in manuscript 1 are books, websites or datasets — counted, but not index-checkable
No retraction or expression of concern found in Crossref (includes Retraction Watch).
Sources: Crossref (incl. Retraction Watch) ✓
Where this paper’s evidence sits on the path from initial observation to real-world use.
Estimator definitions, bandwidth procedures, and Monte Carlo comparisons provide a methodological proof of concept. Accessible implementation guidance and validation in applied workflows are not demonstrated.
AI-generated, human-governed. Something look off? Contact us to request a review.
Keep reading
Other papers Nabu has evaluated blind, starting with the same field.
Hassan, Mertens, Francisco, et al. · 2024 · Economics and business
quantitative asset pricing models, built to address the equity premium and risk-free rate puzzles, systematically fail when applied to open economies: they cannot generate the large and persistent interest rate differentials we observe between risky and safe currencies.
SupportedMurillo Campello, Tomislav Ladika, Rafael Matta · 2018 · Economics
Gagliarducci, Paserman, Patacchini · 2026 · Economics and business
congress members from districts hit by a hurricane are more likely to support bills promoting more environmental regulation and control in the year after the disaster
Well-supported