Rare Earth Price Risk in U.S. Manufacturing

Introduction

29S&P 500 manufacturers across three exposure tiers

ρ = 0.02daily correlation between MVREMXTR and physical spot prices

34–37%average R² under shock-conditioned QQR vs. <3% monthly OLS

14firms with sign divergence — the mispricing signal

$76Bminimum unacknowledged tail risk in our sample

Rare earth elements are everywhere in the hardware that runs the modern economy — in the magnets inside EV motors, the semiconductors in 5G equipment, the guidance systems in defense platforms. Almost all of them come from China. When China tightens supply, manufacturers pay more. The question this research asks is simple: do stock prices know that?

To answer it, we built a pipeline that filters out market noise before looking for the signal. We use daily returns for 29 S&P 500 manufacturers, remove volatility clustering using a GJR-GARCH model, and run quantile-on-quantile regression on the days when rare earth prices actually moved — not the average day, but the shock days that matter for risk. We do this across three different rare earth price measures to stress-test every result.

The most important thing we found before running a single regression: the standard equity-based rare earth index (MVREMXTR) has a daily correlation of ρ = 0.02 with actual physical metal prices. Essentially zero. Most existing research on this topic has been regressing manufacturer stock returns on a variable that tells you almost nothing about what manufacturers pay for inputs. We show the consequences: 14 firms in our sample have positive sensitivity to the equity index but negative sensitivity to the commodity price — their stocks look like they benefit from rare earth price spikes when the physical cost data says the opposite.

Data & Sample

The sample assigns every S&P 500 company to one of three exposure tiers based on GICS sub-industry. High-tier firms — semiconductors, aerospace & defense, industrial machinery, copper mining — form the primary 29-firm estimation sample. Control-group firms verify that the pipeline doesn't generate spurious results for companies with no rare earth exposure.

Tier	Key Sub-Industries	Role
High	Semiconductors, Semiconductor Mat. & Equip., Aerospace & Defense, Auto Manufacturers, Electrical Components, Copper/Mining, Industrial Machinery	Primary QQR estimation sample (29 firms)
Medium	Health Care Equipment, Communications Equipment, Airlines, Life Sciences Tools	Robustness / extended regressions
Control	Restaurants, Insurance, Broadline Retail, Pure Software, Tobacco, Beverages	Placebo — verify no spurious sensitivity

Series	Source	Identifier	Frequency	N
29 firm prices (OHLCV)	Refinitiv/LSEG	RIC, TRDPRC_1	Daily	≈2,597
MVREMXTR	Refinitiv	.MVREMXTR	Daily	2,597
CSI 930598	Refinitiv	.CSI930598	Daily	2,422
China RE PPI	Refinitiv	aCNCNHVGWM / VALUE	Monthly	117
SMM spot (12 metals)	Refinitiv/SMM	SMM-REM-* / SMM-MIN-*	Daily	—
VIX	CBOE	VIX_History.csv	Daily	2,597
Geopolitical Risk Index	Caldara & Iacoviello (2022)	GPRA / GPRT	Monthly	117

Figure 1 — SMM RE Index with Event Months Highlighted (2015–2026)

SMM equal-weighted composite RE index with shock months highlighted. Yellow = up-shock (z > +1.5σ), red = down-shock (z < −1.5σ). Bottom panel: rolling z-score. The 2020 spike (+6.8σ) reflects COVID supply chain disruption.

Figure 2 — Top 15 Firms by RE Exposure vs. SMM Index, Normalized (excl. Nvidia)

All series normalized to 100 at April 2016. White = SMM RE composite. The flat white line vs. 500–4,000% equity returns illustrates why full-sample OLS fails — the RE signal is dominated by the tech bull market on non-shock days.

Basis Risk

Before estimating firm-level sensitivity, we must establish whether the standard RE equity index is a valid proxy for what manufacturers actually pay. The answer is more damaging than previously documented.

MVREMXTR tracks mining company stocks, not the prices manufacturers pay for metals. When rare earth prices rise, miners make more money and their stocks go up — but manufacturers face higher costs. A regression on MVREMXTR conflates both signals and produces a fundamentally biased estimate of manufacturer exposure.

We built two physical spot price indices from SMM data: an equal-weighted composite across 12 metals, and a usage-weighted index overweighting the metals most critical to our sample (neodymium, dysprosium, gallium, and germanium at 20% each).

CORE FINDING — BASIS RISK

At daily frequency, MVREMXTR vs. SMM Equal Weighted: ρ = 0.0201. MVREMXTR vs. SMM Usage Weighted: ρ = 0.0220. These correlations are statistically indistinguishable from zero. The 60-day rolling correlation ranges from −0.38 to +0.30 — meaning MVREMXTR is frequently negatively correlated with the physical prices it is supposed to proxy.

Index	Construction	ρ vs. SMM (daily)
MVREMXTR	Global RE mining equity index	0.0201 — near zero
CSI 930598	Chinese RE processor equity index	Better proxy — tied to domestic quotas
SMM EW	Equal-weighted, 12 metals	1.000 (definition)
SMM UW	Nd/Dy/Ga/Ge 20% each, Pr/Tb 10%	1.000 (definition)

Figure 3 — RE Index Comparison: Cumulative Returns, Rolling Correlation, Individual Metals

Top: cumulative returns (base=100). Middle: 60-day rolling Pearson correlation between MVREMXTR and SMM indices — frequently negative. Bottom: individual metal returns showing the 2021–22 spike concentrated in gallium and germanium.

Figure 4 — Basis Risk Scatter: MVREMXTR vs. SMM Physical Spot (Daily, N ≈ 2,400)

Left: MVREMXTR vs. SMM Equal Weighted (ρ = 0.0201). Right: vs. SMM Usage Weighted (ρ = 0.0220). Regression lines are nearly flat. The vertical SMM cloud reflects 59–99% zero-return days (stale prices), while MVREMXTR moves independently as an equity instrument.

Methodology

Standard OLS on monthly data gives an average R² below 3% and identifies only 9% of companies as significantly exposed. That's not a finding — that's a methodology failure. Our pipeline addresses each failure mode explicitly.

Why OLS Fails

OLS BASELINE (REJECTED)

\[ r_{i,t} = \alpha_i + \beta_i \, r_{RE,t} + \varepsilon_{i,t} \]

ARCH effects. Time-varying volatility makes OLS standard errors inconsistent. Ljung–Box tests reject white-noise residuals for all 29 firms on raw returns.
Omitted variable bias. Full-sample RE price moves are dominated by the green energy macro cycle, biasing full-sample OLS upward. The full-vs.-event delta scatter shows this directly.
Regime non-linearity. RE–stock sensitivity is qualitatively different during supply shocks vs. calm periods. OLS averages across both regimes, characterizing neither correctly.

Five-Stage Pipeline

Log Returns & Stationarity

Convert all series to log returns $r_t = \ln(P_t/P_{t-1})$. Verify stationarity via ADF and KPSS for all 32 series. Jarque–Bera tests confirm non-normality, validating quantile methods over OLS.

GJR-GARCH(1,1) Volatility Filtering

Remove volatility clustering from all series. GJR captures the leverage effect — negative shocks generate larger subsequent volatility. $\hat\gamma > 0$ in 26/29 firms (89.7%). Output: standardized residuals $\hat z_t = \hat\varepsilon_t / \hat\sigma_t$.

Shock-Day Identification

Flag days where the RE index's standardized residual exceeds |z| > 2.0 (rolling 60-day window, ±2-day expansion). Result: 643 MVREMXTR shock days, 626 CSI shock days — ~25% of sample.

Quantile-on-Quantile Regression

Run QQR on shock days across $\mathcal{Q} = \{0.05,0.10,0.25,0.50,0.75,0.90,0.95\}^2$. Gaussian kernel, bandwidth h = 0.15. Produces a full $\hat\beta(\tau,\theta)$ surface. Average R² rises to 34–37%.

China RE PPI Robustness

Monthly OLS with five macro controls (VIX, US 10Y, DXY, WTI, Copper) and HC3 robust SEs, one-month PPI lag. The sign of $\hat\delta^{PPI}$ relative to $\hat\delta^{MV}$ is the primary mispricing diagnostic.

GJR-GARCH(1,1) — Glosten, Jagannathan & Runkle (1993)

\[ \sigma^2_t = \omega + \alpha\varepsilon^2_{t-1} + \gamma\varepsilon^2_{t-1}\cdot\mathbf{1}(\varepsilon_{t-1}<0) + \beta\sigma^2_{t-1} \]

QQR KERNEL-WEIGHTED OBJECTIVE — Sim & Zhou (2015)

\[ \min_{\alpha,\beta} \sum_{t\in\mathcal{S}} K_h\!\left(\frac{F_n(z_{RE,t})-\theta}{h}\right)\rho_\tau\!\left(z_{i,t}-\alpha-\beta z_{RE,t}\right) \]

MISPRICING INDICATOR & DOLLAR EXPOSURE

\[ \text{SR}_i = \mathbf{1}(\hat\delta_i^{MV}\cdot\hat\delta_i^{PPI} < 0), \qquad \text{Exp}_i = |\hat\delta_i^{MV}|\times\hat\sigma_{RE}^{\text{ann}}\times\text{MCap}_i \]

Figure 5 — GJR-GARCH Conditional Volatility: NVDA, AAPL, RTX, FCX, MVREMXTR (2016–2026)

Estimated $\hat\sigma_t$ from GJR-GARCH(1,1). Red dashed lines mark the 2020 COVID shock and 2022 Fed rate cycle. NVDA peaks at ~12% daily vol in March 2020. The standardized residuals from these models are the inputs to QQR.

GJR-GARCH Parameter Estimates — All 29 Firms

Ticker	ω (×10⁻²)	α	γ (leverage)	β	Persistence	LB p-val	Clean?

Persistence = α + γ/2 + β. LB = Ljung–Box p-value at lag 10. TSLA fails (p = 0.040). INTC persistence = 1.000.

Results

24/29significant under MVREMXTR shock-day regression

4/29significant under CSI — a better commodity proxy

34–37%average R² under shock-conditioned QQR

−5.79worst-case β(0.05,0.95) — Northrop Grumman on shock days

When we use the equity index (MVREMXTR), 24 of 29 manufacturers appear to benefit from rare earth price spikes — their stocks go up with the index. When we use actual commodity prices, many of those same firms show negative sensitivity. The equity market is treating a cost shock as a demand signal. That's the mispricing.

Figure 6 — Full Sample vs. Event-Based Delta: All S&P 500 Companies

Points below the diagonal = event regression finds more negative RE sensitivity than full-sample OLS, confirming the green energy macro trend biases full-sample estimates upward. First Solar is the most dramatic: full-sample δ = +0.45, event-month δ = −1.05.

Figure 7 — Average RE Sensitivity by GICS Sub-Industry (Full S&P 500)

IT Consulting (−0.304) and Application Software (−0.279) most negative. Aerospace & Defense (+0.177) appears positive under the equity index — consistent with the mispricing hypothesis.

Figure 8 — Event Regression Delta: CSI 930598 vs. MVREMXTR (29 Firms)

Left: CSI (sig: 4/29, avg R² = 0.336). Right: MVREMXTR (sig: 24/29, avg R² = 0.371). The near-universal positive sign under MVREMXTR is the empirical signature of basis risk.

Three-Index Delta — All 29 Firms

Rows in red have sign divergence between MVREMXTR and PPI — the mispricing candidates. ** p < 0.05.

Ticker	Company	δ̂ᴹᵛ	Sig	R²	δ̂ᶜˢᴵ	Sig	R²	δ̂ᴾᴾᴵ	Sig	Sign

QQR Surfaces

The QQR surface $\hat\beta(\tau,\theta)$ maps the sensitivity of stock return quantile τ to RE index return quantile θ. Each firm's surface is computed on three subsamples: Full Sample, Shock Days, and Normal Days. The red dot marks the worst-case cell β(0.05, 0.95) — the stock falls hardest when the RE index spikes most.

KEY SURFACE CELLS

\[ \beta(0.05,\,0.95) = \text{worst case: stock crashes when RE spikes} \qquad \hat\delta_i = \overline{\hat\beta(\tau,\theta)} = \text{mean over grid} \]

Figure 9 — QQR Worst-Case β(0.05, 0.95): MVREMXTR vs. CSI, All 29 Firms

MVREMXTR: 8/29 negative. CSI: 14/29 negative. The shift in count and rank ordering when using a better commodity proxy reveals that the equity index is systematically concealing downside exposure.

Firm-Level Surface Explorer

SELECT FIRM

QQR Surface — Full Sample / Shock Days / Normal Days NVDA.O

β(worst) MV—

β(shock) MV—

β(worst) CSI—

β(shock) CSI—

δ̂ᴹᵛ—

δ̂ᶜˢᴵ—

δ̂ᴾᴾᴵ—

Exposure—

Extensions

This paper establishes the core finding: markets are systematically mispricing rare earth supply risk for U.S. manufacturers, and the standard equity-based proxy is not a valid measure of actual input cost exposure. Several natural extensions follow directly.

Macro Control Integration

Add VIX, geopolitical risk indices (Caldara & Iacoviello 2022), US 10Y, DXY, and WTI as controls in the PPI event regression. Data already downloaded. Expected to tighten standard errors without reversing sign patterns.

Hedging Strategy Backtest

Long-short portfolio: long δ̂ᴾᴾᴵ > 0 firms (natural hedges, like Seagate), short δ̂ᴾᴾᴵ < 0 firms. Backtest across three RE shock episodes: 2010 Chinese quota shock, 2019–20 trade war, 2023 gallium/germanium export controls.

EDGAR Supply Chain Integration

Scrape 10-K filings for RE-specific language to calibrate dollar exposure to actual procurement volumes rather than pure market cap scaling. Converts the reduced-form measure into a structural estimate grounded in firm-specific supply chain architecture.

Markov Regime-Switching Extension

Condition QQR on Markov regime probability rather than rolling z-score — a formally model-based approach statistically superior to the ad-hoc threshold. Two-state Markov model already fitted to the RE index.

Metal-Level Disaggregation

Run firm-level exposure estimates for neodymium, dysprosium, gallium, and germanium separately, matching metals to firms' specific product lines. The usage-weight structure already encodes the prior over which metals matter most.

DOE Grant — NOFO DE-FOA-0003589

This framework directly addresses "market acceptance barriers" to domestic RE supply investment identified by DOE ($69M total). Topic Area 2 (gallium/germanium for semiconductors) is closest to our sample composition. Letter of Intent in development.

References

Black, F. (1976). Studies of stock market volatility changes. Proceedings of the ASA, 177–181.

Caldara, D. & Iacoviello, M. (2022). Measuring geopolitical risk. American Economic Review, 112(4), 1194–1225.

Glosten, L. R., Jagannathan, R., & Runkle, D. E. (1993). On the relation between the expected value and the volatility of the nominal excess return on stocks. Journal of Finance, 48(5), 1779–1801.

Koenker, R. & Bassett, G. (1978). Regression quantiles. Econometrica, 46(1), 33–50.

Panazan, O. & Gheorghe, C. (2025). Quantile-on-quantile analysis of rare earth price impacts on energy transition sectors. IJFS, 13(1).

Sim, N. & Zhou, H. (2015). Oil prices, US stock return, and the dependence between their quantiles. Journal of Banking & Finance, 55, 1–8.

U.S. Department of Energy. (2025). Critical materials manufacturing and materials recycling (NOFO DE-FOA-0003589).