Tracked Visits
—
Cumulative tracked website visits
National Manufacturing Overview
Aggregate figures across all 19 real manufacturing subsectors in the panel, 2015–2025. Every number below is summed directly from the underlying dataset — nothing here is estimated or illustrative.
Manufacturing subsectors
19
NAICS-3 level, all real
2025 Aggregate Output
$7,180.9B
up from $5,540.4B in 2015
2025 Employment
12,596.8k
thousands of workers
2025 Trade Balance
–$1,208.3B
exports $1,772.5B / imports $2,980.8B
Aggregate Gross Output, 2015–2025 ($ billions)
Trade Balance, 2015–2025 ($ billions)
Predictive Risk Monitor
Every subsector's Resilience Score alongside its two real underlying drivers: import dependence and import-origin concentration. See the callout below for exactly how the score is calculated.
| Industry | Resilience | Import Dep. | Supplier Conc. | Risk |
|---|
Illustrative composite, not a regression result
Resilience Score = 100 − 0.5×(Import Dependence %) − 0.3×(Supplier Concentration, HHI÷100) + 2.0×(year-over-year Productivity Growth %), clipped to 0–100. Real inputs, transparent round-number weights — but this score has not been statistically estimated or validated against actual disruption outcomes. The regression results in Panel Regression and Robustness & Hausman are the validated analysis; this score exists to make the same real inputs scannable at a glance. See Data & Methodology for the same disclosure.
Trade Exposure
Import dependence (imports ÷ [output + imports]) and supplier concentration (real HHI, rescaled to 0–100) for every subsector, ranked by exposure.
| Industry | Import Dependence | Supplier Concentration |
|---|
Manufacturing Resilience
Full resilience ranking across all subsectors, highest to lowest.
| Industry | Resilience Score | Risk |
|---|
Industry Competitiveness
Real year-over-year growth rates (2024 → 2025) across exports, labor productivity, employment, and output.
| Industry | Export Growth | Productivity | Employment | Output |
|---|
Two-Way Fixed Effects Panel Regression
Dependent variable: log(Gross Output). N = 209 (19 subsectors × 11 years). Entity-clustered standard errors (asymptotic). This is the paper's primary econometric result, reproducible from src/05_panel_regression.py. Because this panel has only 19 entity clusters, a wild-cluster bootstrap — a more statistically appropriate small-sample approach — is reported in the Robustness section, and that bootstrap result should be treated as the more credible inferential statement.
| Variable | Coefficient | Std. Error | t-stat | Asymptotic p-value |
|---|---|---|---|---|
| Employment (thousands) | 0.0002 | 0.0002 | 0.85 | 0.397 |
| Labor Productivity (Index) | 0.0005 | 0.0018 | 0.30 | 0.765 |
| log(Imports) | 0.1998 | 0.1023 | 1.95 | 0.052 |
| log(Exports) | 0.2747 | 0.1349 | 2.04 | 0.043 |
| Import-Origin Concentration (HHI, std.) | −0.0240 | 0.0220 | −1.09 | 0.275 |
R² (within)
0.619
F-statistic
F(5,175) = 17.96
p-value
< 0.001
Reported honestly
Under standard asymptotic clustering, imports and exports look significant or near-significant. But with only 19 entity clusters, a wild-cluster bootstrap is the more statistically defensible test — and it finds neither imports (p=0.14) nor exports (p=0.08) significant at 5%. Import-Origin Concentration is directionally negative, consistent with this framework's hypothesis, but not significant under any specification tested. None of these dots are marked "significant" on this page for that reason. See Robustness & Hausman for the full diagnostic picture.
Robustness Checks & Hausman Test
Five specifications estimated on the identical real panel, plus a corrected Hausman test, a Mundlak test, a wild-cluster bootstrap, and a trade × concentration interaction — all added or corrected following independent methodological review.
| Variable | Pooled OLS | Random Effects | 2-Way FE (no HHI) | 2-Way FE (Main) | Entity FE only |
|---|---|---|---|---|---|
| Employment | 0.0007** | 0.0005** | 0.0001 | 0.0002 | 0.0003 |
| Labor Productivity | −0.0193** | −0.0013 | 0.0007 | 0.0005 | −0.0010 |
| log(Imports) | −0.5432** | 0.1983* | 0.1934* | 0.1998* | 0.2824** |
| log(Exports) | 0.8430** | 0.4078** | 0.2526* | 0.2747** | 0.3611** |
| Import-Origin Concentration | −0.0721* | −0.0764** | — | −0.0240 | −0.0679** |
| R² (overall / within) | 0.861 | 0.693/0.704 | 0.329/0.553 | 0.339/0.619 | 0.710 |
* p<0.10, ** p<0.05 (asymptotic). Full computation in src/06_robustness_checks.py.
Table: Model-Selection Tests
| Test | Statistic | p-value | Conclusion |
|---|---|---|---|
| Hausman (like-for-like, year effects in both) | χ²(5)=5.36 | 0.374 | Fails to reject RE |
| Mundlak (joint significance of entity-mean regressors) | χ²(5)=48.97 | <0.0001 | Rejects RE; supports FE |
These two tests point in different directions and both are reported. The corrected Hausman test alone does not decisively favor FE. The Mundlak test, which directly tests whether entity averages correlate with the outcome, does support FE — and is treated as the more directly relevant test here. Fixed Effects remains the preferred specification, but as a reasoned judgment from mixed evidence, not a test that "confirms" it beyond doubt.
Table: Additional Diagnostics
| Diagnostic | Result | Interpretation |
|---|---|---|
| Wild-cluster bootstrap — log(Imports) | p=0.142 | Not significant at 5% |
| Wild-cluster bootstrap — log(Exports) | p=0.078 | Not significant at 5% |
| Wild-cluster bootstrap — Import-Origin Concentration | p=0.263 | Not significant, consistent with asymptotic |
| Variance Inflation Factors (all vars) | 1.08–4.41 | No multicollinearity concern |
| Pesaran CD test (cross-sectional dependence) | p=0.169 | Not significant; supports entity-only clustering |
| Serial correlation test (H₀: coef=−0.5) | p<0.0001 | Rejects null — evidence of persistence a static model misses |
Table: Trade × Concentration Interaction
| Variable | Coefficient | p-value |
|---|---|---|
| log(Imports) × Concentration (HHI, std.) | 0.0119 | 0.653 (n.s.) |
Honest null result
Added specifically to test whether concentrated subsectors are more vulnerable to a trade shock than diversified ones. Not significant — this panel does not currently support that hypothesis, despite it being intuitive and central to this framework's premise. Reported directly rather than omitted.
Table: Dynamic Panel & Employment-Impact Models
| Model | Key finding |
|---|---|
| Dynamic panel (lagged output) | ρ = 0.4149, asymptotic p = 0.033 (marginal under bootstrap, p = 0.063) |
| Employment-impact model | No coefficient statistically significant (genuine null result) |
Pseudo-Out-of-Sample Forecast Evaluation
Rolling-origin design, averaged across four genuine one-year-ahead forecasts (2022, 2023, 2024, 2025). Each forecast uses only data that would actually have been available at the time.
| Model | Avg MAPE | Avg RMSE ($B) |
|---|---|---|
| Naive persistence | 4.94% | 37.43 |
| Linear (entity FE + trend) | 8.04% | 49.85 |
| Gradient Boosting (ML) | 11.90% | 118.91 |
| Hybrid (Linear + GBR) | 8.51% | 82.66 |
Mean Absolute Percentage Error by Model
Honest finding, not a failure — now on genuinely sound footing
The naive persistence benchmark achieved the lowest average error of all four approaches, by an even wider margin than the earlier (methodologically flawed) version of this exercise suggested. This is reported directly: with genuinely available lagged predictors and this panel's size, neither the econometric nor the machine-learning approach can be shown to add forecasting value over simply carrying forward each subsector's most recent output figure.
Exploratory Model-Implied Scenario Tool
Illustrative association, not a causal policy forecast
This tool applies the fitted regression coefficients arithmetically to hypothetical shocks. It does not estimate a causal effect — genuine causal disruption analysis would require exogenous variation (tariff natural experiments, shift-share instruments, event-study designs, or a structural CGE model), none of which this exercise provides. Treat every number below as an illustration of the calculation, not a prediction.
Live calculation using the fitted coefficients from the panel regression (Table 4.2 in the companion paper). Choose a subsector and adjust a hypothetical import shock or concentration shift — this is genuine arithmetic on the actual regression output, not a mock interface, but the underlying coefficients are not statistically significant under small-sample-appropriate inference (see Robustness & Hausman).
−20%
0.0 SD
−4.36%
model-implied output change
from $425.6B → $407.0B baseline (Computer and electronic products, 2025)
2025 Baseline Output
$425.6B
Current HHI
1,299
Current Imports
$709.35B
Comparative Illustration: Alternative Interventions
A single shock is useful, but real policy questions usually compare options. This table applies the same real coefficients to three named, alternative interventions for Computer and electronic products (2025 baseline $425.6B) — still illustrative and non-causal, but now structured as a genuine comparison.
| Intervention | Assumption | Model-Implied Change | Resulting Output |
|---|---|---|---|
| A: Tariff-style import reduction | −20% imports | −4.36% | $407.0B |
| B: Supplier diversification | −1 SD import-origin concentration | +2.43% | $435.9B |
| C: Combined approach (A + B) | Both simultaneously | −2.04% | $416.9B |
This illustrates a genuine trade-off the fitted model implies — import reduction and supplier diversification pull output in opposite directions — but should be read cautiously: the concentration coefficient underlying Scenario B isn't statistically significant in the preferred specification, and this arithmetic treats the two interventions as independent, when reducing concentration would plausibly also change import volumes themselves. A capability that resolves this kind of trade-off with real statistical and causal confidence is future work, not something this illustration provides.
Data Sources & Methodology
Every variable in this framework traces to one of four federal sources. Full construction methodology and documented data-infrastructure gaps are in docs/data_sources.md.
U.S. Bureau of Economic Analysis
Gross Output by Industry
Dependent variable. Quarterly, averaged to annual, by NAICS-3 subsector.
U.S. Bureau of Labor Statistics
Current Employment Statistics
Employment, thousands, seasonally adjusted, monthly, averaged to annual.
U.S. Bureau of Labor Statistics
Industry Productivity Program
Labor productivity index (2017=100), annual, all-workers basis.
U.S. International Trade Commission
DataWeb
Imports, exports, and country-of-origin data used to compute supplier concentration.
Current Prototype Limitations
These are scoped-out future development modules, not gaps in documentation — each is already planned in the Future Research Roadmap.
- No formal dependency network linking specific foreign suppliers to specific domestic manufacturers — only aggregate concentration is measured.
- No causal identification strategy — trade and concentration coefficients throughout this dashboard are associational, not causal effects.
- No economic-impact models estimating disruption effects on prices specifically (only output and employment are modeled).
- No retrospective validation against specific historical disruption episodes (e.g., 2018–19 tariff actions, 2020 COVID shock).
- N = 209 is a modest panel by econometric standards, though genuinely real and multi-source.
Platform Usage & Adoption
Live indicators of public engagement with the Economic Intelligence Copilot.
Dashboard Downloads
—
EIC dashboard downloads
GitHub Stars
—
Repository stars
GitHub Forks
—
Repository forks
About These Metrics
Tracked Visits are reported through GoatCounter. Dashboard Downloads, GitHub Stars, and GitHub Forks are retrieved from GitHub.
Future Research Roadmap
This is an ongoing project, not a closed one. Everything below is drawn directly from the honest gap-analysis in the companion paper (Table 4.5) — stated plainly rather than glossed over.
Near-Term (extends the current real analysis)
| Direction | What it would add |
|---|---|
| Trade × concentration interaction terms | Tested — added and estimated (coefficient 0.0119, p=0.653, not significant). This panel does not yet support the hypothesis that concentrated subsectors are more vulnerable to a shock than diversified ones; re-testing with a larger panel is the next step. |
| Larger panel | More subsectors, more years, or finer NAICS detail — directly addresses why Import-Origin Concentration is directionally negative but not statistically significant under any specification tested, including the interaction term above; more cross-sectional units would sharpen statistical power. |
| Dynamic panel (GMM) specification | Arellano-Bond or Blundell-Bond estimation to properly capture persistence in manufacturing output, which a static model cannot. |
Medium-Term (new capabilities)
| Direction | What it would add |
|---|---|
| Formal dependency network | Map specific foreign suppliers to specific domestic manufacturers and their downstream industries — not just aggregate concentration, but who actually depends on whom. |
| Economic-impact models | Estimate disruption effects on prices and employment specifically, not just aggregate output. |
| Retrospective validation | Test the framework against real historical disruption episodes — the 2018–19 tariff actions, the 2020 COVID shock — to see whether it would have flagged them in advance. |
Longer-Term (national scaling)
| Direction | What it would add |
|---|---|
| Expansion beyond manufacturing | Apply the same framework to additional sectors (agriculture, energy, services) using the same reusable Python pipeline (see docs/methodology.md). |
| Formal interface evaluation | Test this dashboard's format against alternatives with real policymaker and manufacturer users, rather than assuming this is the right presentation. |
| Institutional partnership | Collaboration with federal agencies, universities, or industry associations to validate and formally maintain the framework at national scale. |
Get involved
This project is open-source under the MIT License specifically so that other researchers, agencies, and manufacturers can extend it. See the GitHub repository for the full codebase, or reach out to the author, Olusoji Adenekan, to discuss collaboration on any of the directions above.