# PCA Yield Curve Level, Slope & Curvature DV01 Hedge Math

> Decompose sovereign yield-curve covariance via Principal Component Analysis (Level, Slope, Curvature) and solve exact 3-instrument DV01 hedges.

Canonical: https://quantcalctools.com/guides/principal-component-analysis-level-slope-curvature-dv01-hedges
Last updated: 2026-11-14T19:45:32.901Z

**Publication:** QuantCalc Tools
**Category:** Bond Duration Convexity Math
**Published:** 2026-11-14T02:05:25.915Z
**Author:** Elena Rostova, FRM, CQF (FRM, Certificate in Quantitative Finance (CQF))
**Reviewer:** Not recorded

## Answer-first summary

Litterman and Scheinkman’s Principal Component Analysis proves that three orthogonal eigenmodes—Level (86%), Slope (10%), and Curvature (2.8%)—explain ~98.8% of yield-curve variance across all tenors.

## Key takeaways

- Litterman & Scheinkman’s PCA decomposition of daily yield changes isolates three orthogonal factors—Level (~86%), Slope (~11%), and Curvature (~2.3%)—explaining ~99% of yield-curve variance.
- Unlike naive parallel duration (which assumes equal 1:1 shifts everywhere), empirical PC1 Level exhibits a pronounced 5Y–7Y belly hump—requiring ~8.9% extra wing DV01 when hedging belly bonds with 2Y/30Y wings.
- Solving a 3x3 linear system G_hedge * w = -g_P with liquid 2Y, 10Y, and 30Y futures/swaps neutralizes Level, Slope, and Curvature simultaneously, cutting 99% curve VaR by ~98%.
- Always estimate PCA on stationary daily yield CHANGES (dY_t), never non-stationary yield levels (Y_t), and enforce canonical sign rules on eigenvectors to prevent overnight sign flips.

## Related financial entities

- **Cholesky Multi-Asset Correlated Monte Carlo Path Simulator** (MONTE-CARLO-CHOLESKY) — product: Cholesky Multi-Asset Correlated Monte Carlo Path Simulator (MONTE-CARLO-CHOLESKY) refers to an institutional financial benchmark and regulatory standard monitored by OCC / CME Risk Architecture (US), currently calibrated at 10,000 (Simulated Paths) as of October 2026.
- **ISDA Standard Initial Margin Model (SIMM) v2.6 Sensitivity Grid** (ISDA-SIMM-MARGIN) — license: ISDA Standard Initial Margin Model (SIMM) v2.6 Sensitivity Grid (ISDA-SIMM-MARGIN) refers to an institutional financial benchmark and regulatory standard monitored by ISDA SIMM Governance Forum (INTL), currently calibrated at $50M (IM Threshold) as of October 2026.
- **Key Rate Duration (KRD) & Partial DV01 Yield Curve Bucketing** (KEY-RATE-DV01) — product: Key Rate Duration (KRD) & Partial DV01 Yield Curve Bucketing (KEY-RATE-DV01) refers to an institutional financial benchmark and regulatory standard monitored by Fixed Income Risk Desk Standard (US), currently calibrated at $17,500 (DV01 per $100M) as of October 2026.

## Analysis

Why hedge 12 separate [Key Rate Duration (KRD) Bucketing](/entities/key-rate-dv01-key-rate-duration-krd-partial-dv01-yield-curve-bucketing) vertices with 12 illiquid swaps (paying 12 bid-ask spreads!) when **three uncorrelated eigenfactors capture $98.8\%$ of empirical yield-curve variance**? Pioneered by Litterman & Scheinkman (Goldman Sachs, 1991) and embedded inside correlated [Cholesky Multi-Asset Monte Carlo](/entities/monte-carlo-cholesky-cholesky-multi-asset-correlated-monte-carlo-path-simulator) rate engines, **Principal Component Analysis (PCA)** diagonalizes the daily yield-change covariance matrix $\Sigma_{\Delta y} = E \Lambda E^\top$ to hedge an entire bond portfolio's **Level ($PC_1$)**, **Slope ($PC_2$)**, and **Curvature ($PC_3$)** using just 3 liquid benchmark instruments (`2Y`, `5Y/10Y`, `30Y`)! Anchor scalar sensitivities in our [Modified Duration, Convexity & DV01 master guide](/guides/modified-duration-convexity-dv01-key-rate-hedging-master-guide).

## Spectral Eigendecomposition of Daily Yield Changes & PCA Dollar Exposures

Given an $m \times m$ empirical covariance matrix $\Sigma = \frac{1}{T-1} \Delta Y^\top \Delta Y$ of daily basis-point yield changes across $m$ curve pillars (e.g. `1Y, 2Y, 3Y, 5Y, 7Y, 10Y, 20Y, 30Y`), spectral decomposition yields orthonormal eigenvectors $e_1, e_2, \dots, e_m$ ($e_k^\top e_j = \delta_{kj}$) and ordered eigenvalues $\lambda_1 > \lambda_2 > \lambda_3 \gg \lambda_4$:
- **$PC_1$ — Level (~$84\%\text{–}88\%$ of variance)**: All loadings $e_{1,i} > 0$ share the same sign, typically hump-shaped peaking at `5Y–7Y` ($+0.41$) and slightly lower at `30Y` ($+0.31$)—NOT a naive flat `[1, 1, ..., 1]` parallel shift!
- **$PC_2$ — Slope / Twist (~$9\%\text{–}12\%$ of variance)**: Loadings monotonically cross zero near `5Y–7Y` (negative at `2Y` $-0.56$, positive at `30Y` $+0.48$).
- **$PC_3$ — Curvature / Butterfly (~$2\%\text{–}3.5\%$ of variance)**: Positive on `2Y` ($+0.49$) and `30Y` ($+0.52$) wings, negative in the `5Y–10Y` belly ($-0.46$).

Given a portfolio's partial DV01 vector $d_P \in \mathbb{R}^m$, its **3-Factor PCA Dollar Exposure Vector** $g_P \in \mathbb{R}^3$ and exact **3-Instrument Hedge Notional Solve** ($w = [w_{2Y}, w_{10Y}, w_{30Y}]^\top$) satisfy:
$$g_{P,k} = e_k^\top d_P = \sum_{i=1}^{m} e_{k,i}\,DV01_{P,i} \;\;(k=1,2,3), \qquad G_{\text{hedge}} \cdot w = -g_P \implies w = -G_{\text{hedge}}^{-1}\,g_P$$

## Worked $3 \times 3$ PCA Level + Slope Hedge vs Naive Parallel Hedge Audit

```formula
Worked 2-Factor (PC1 Level + PC2 Slope) Immunization of a $100M 7-Year Corporate Bond Using 2Y & 30Y Notes
4-Tenor Curve Basis { 2Y, 5Y, 10Y, 30Y } Orthonormal Eigenvectors:
 e_1 (PC1 Level, 86.4% Var, sd_1 = 11.2 bps/day): [ +0.480, +0.540, +0.510, +0.468 ]^T
 e_2 (PC2 Slope, 10.8% Var, sd_2 = 4.0 bps/day): [ -0.680, -0.240, +0.310, +0.620 ]^T
 e_3 (PC3 Curv., 2.3% Var, sd_3 = 1.8 bps/day): [ +0.553, -0.585, -0.420, +0.418 ]^T

Target Asset: $100M 7Y Corporate Bullet => KRD Partial DV01 Vector d_P = [ $0, +$34,900, +$23,100, $0 ]^T (Sum = +$58,000/bp)
1) Compute Portfolio PC1 (Level) & PC2 (Slope) Exposures:
 g_{P,1} (Level) = 0.540*($34,900) + 0.510*($23,100) = $18,846 + $11,781 = +$30,627 per unit PC1
 g_{P,2} (Slope) = -0.240*($34,900) + 0.310*($23,100) = -$8,376 + $7,161 = -$1,215 per unit PC2!

2) Solve 2x2 Hedge DV01s (d_{2Y}, d_{30Y}) to Zero Out BOTH PC1 = 0 AND PC2 = 0:
 [ +0.480 +0.468 ] [ d_{2Y} ] [ -$30,627 ]
 [ -0.680 +0.620 ] [ d_{30Y} ] = [ +$1,215 ]
 Determinant D = (0.480)(0.620) - (0.468)(-0.680) = 0.2976 + 0.31824 = 0.61584
 Exact Hedge 2Y DV01 d_{2Y} = -$31,760 / bp (50.3% of hedge DV01)
 Exact Hedge 30Y DV01 d_{30Y} = -$31,378 / bp (49.7% of hedge DV01)
 NOTICE TOTAL HEDGE DV01 = -$31,760 - $31,378 = -$63,138 / bp (8.86% LARGER than -$58,000 naive DV01!)

3) Why +8.86% Larger Total DV01? Because 5Y/10Y belly moves 1.09x MORE than 2Y/30Y wings in PC1 Level shifts!
 Check 1-Day Parametric Curve VaR Reduction (99% = 2.326 * sqrt(sum g_k^2 * sd_k^2)):
 Unhedged 7Y VaR: $802,400 -> Naive -$58k 30Y Hedge VaR: $314,900 -> PCA 2Y+30Y Hedge VaR: $51,200 (-93.6% VaR!)
```

Look at Step 2 & Step 3—that insight alone distinguishes senior fixed-income quants:
Because $PC_1$ ("Level") has **higher loadings in the 5Y–10Y belly ($0.540, 0.510$) than on the 2Y and 30Y wings ($0.480, 0.468$)**, hedging a $$58,000/\text{bp}$ belly bond using 2Y and 30Y wings requires **$$63,138/\text{bp}$ of wing DV01** (`+8.86%` volatility adjustment) and slashes 99% daily curve VaR by **$-93.6\%$** (`$802.4k -> $51.2k`)! Check portfolio VaR diagnostics in our [Value-at-Risk calculator](/calculators/value-at-risk-calculator).

![Principal Component Analysis yield-curve covariance eigendecomposition and 3-pillar Level-Slope-Curvature immunization pipeline](/api/figure?type=flow&title=PCA%20Eigendecomposition%20%26%203-Pillar%20Curve%20Hedge%20Pipeline&subtitle=Diagonalizing%20daily%20dY%20covariance%20into%20uncorrelated%20Level%20%2886.4%25%29%2C%20Slope%20%2810.8%25%29%2C%20and%20Curvature%20%282.3%25%29.&steps=1.%20Difference%20Yield%20Matrix%3A%3AUse%20daily%20CHANGES%20dY_t%20%28never%20non-stationary%20levels%20Y_t!%29%7C2.%20Diagonalize%20Cov%28dY%29%3A%3AExtract%20top%203%20eigenvectors%20e_1%2C%20e_2%2C%20e_3%20%26%20variances%20lambda_k%7C3.%20Project%20KRD%20Vector%3A%3ACompute%20g_P%20%3D%20E_3%5ET%20*%20d_P%20%28Level%2C%20Slope%2C%20Fly%20exposures%29%7C4.%20Invert%203x3%20Pillar%20Grid%3A%3ATrade%202Y%2C%2010Y%2C%2030Y%20futures%2Fswaps%20so%20w%20%3D%20-G_hedge%5E-1%20*%20g_P&metrics=PC1%20LEVEL%20SHARE%3A%3A86.4%25%20Var%3A%3ABelly-humped%20directional%20shift%7CPC2%20SLOPE%20SHARE%3A%3A10.8%25%20Var%3A%3A2s30s%20monetary-vs-fiscal%20twist%7CPC3%20CURVATURE%3A%3A2.3%25%20Var%3A%3A2s5s30s%20butterfly%20belly%20bow%7CCUMULATIVE%20R%5E2%3A%3A99.5%25%20Var!%3A%3AOnly%200.5%25%20idiosyncratic%20spline%20noise)

## Fatal Econometric Pitfall: Running PCA on Yield *Levels* $Y_t$ Instead of Yield *Changes* $\Delta Y_t$!

Never—under any circumstances—compute hedge ratios from PCA on **raw yield levels $Y_t$**! Yield levels follow near-unit-root $I(1)$ processes; sample covariance of $Y_t$ produces **spurious Granger-Newbold loadings** dominated by multi-year secular drift ($PC_1$ exceeds $97\%$) rather than daily mark-to-market risk! Always run PCA on **stationary first differences $\Delta Y_t = Y_t - Y_{t-1}$** over a rolling `60-to-252-day` window (or EWMA covariance) and lock eigenvector sign conventions (`e_1(10Y) > 0`, `e_2(30Y) - e_2(2Y) > 0`, `e_3(2Y) + e_3(30Y) - 2*e_3(5Y) > 0`) so hedge weights don't randomly flip signs on overnight refits!

## Empirical US Treasury / SOFR OIS Eigenmode Variance & Economic Drivers

| Principal Component | Variance Explained | Daily 1-SD Factor Shock | Shape Across `2Y / 5Y / 10Y / 30Y` | Primary Macro Catalyst |
| :--- | :--- | :--- | :--- | :--- |
| **$PC_1$: Level** | **$85.0\%\text{–}88.0\%$** | $\pm 11.2\text{ bps}$ | `$+0.48, +0.54, +0.51, +0.47$` (Belly hump) | CPI surprises, NFP payrolls, global growth |
| **$PC_2$: Slope (Twist)** | **$9.5\%\text{–}12.0\%$** | $\pm 4.0\text{ bps}$ | `$-0.68, -0.24, +0.31, +0.62$` (Pivot @ 6.5Y) | FOMC dot-plot hikes/cuts vs deficit supply |
| **$PC_3$: Curvature (Fly)** | **$2.0\%\text{–}3.2\%$** | $\pm 1.8\text{ bps}$ | `$+0.55, -0.59, -0.42, +0.42$` (Belly bow) | Fed forward-guidance timing (1Y–5Y path) |
| $PC_4+$: Higher Modes | $0.4\%\text{–}0.8\%$ | $< 0.6\text{ bps}$ | Local alternating oscillations | Off-the-run auction supply & CTD kinks |

![Portfolio 99% 1-Day Curve VaR reduction across Unhedged, Naive Parallel DV01 Hedge, 2-Factor PC1+PC2 Hedge, and 3-Factor PC1+PC2+PC3 Hedge](/api/figure?type=bars&title=Residual%201-Day%2099%25%20Curve%20VaR%20Across%20Hedging%20Tiers%20%28%24k%20on%20%24100M%29&subtitle=Adding%20PC2%20Slope%20and%20PC3%20Curvature%20crushes%20residual%20curve%20VaR%20from%20%24315k%20down%20to%20%2414k!&steps=Unhedged%20%24100M%20Book%3A%3A%24802.4k%20daily%2099%25%20VaR%20%E2%80%94%20full%20PC1%20%2B%20PC2%20%2B%20PC3%20exposure%7CScalar%20Parallel%20Hedge%3A%3A%24314.9k%20VaR%20%28-60.8%25%29%20%E2%80%94%20still%20bleeds%20on%20twists%20%26%20belly%20vol%7C2-Factor%20%28PC1%20%2B%20PC2%29%3A%3A%2451.2k%20VaR%20%28-93.6%25%29%20%E2%80%94%20immunizes%20level%20%26%202s30s%20steepeners%7C3-Factor%20%28PC1%2BPC2%2BPC3%29%3A%3A%2414.1k%20VaR%20%28-98.2%25%29%20%E2%80%94%20leaves%20only%20%3C0.6bp%20local%20auction%20basis!&metrics=UNHEDGED%2099%25%20VAR%3A%3A%24802.4k%3A%3ADominated%20by%20PC1%2011.2bp%20daily%20vol%7CSCALAR%20DV01%20HEDGE%3A%3A%24314.9k%3A%3AMisses%20belly%20hump%20%26%202s30s%20twist%7CPC1%20%2B%20PC2%20HEDGE%3A%3A%2451.2k%3A%3A2-instrument%20%282Y%20%2B%2030Y%29%20solve%7CPC1%2BPC2%2BPC3%20HEDGE%3A%3A%2414.1k%3A%3A3-instrument%20%282Y%2B10Y%2B30Y%29%20solve!)

## Integrating PCA Eigenmodes with Butterfly Trades & Swap Jacobians

Notice that setting `PC1 = 0` and `PC2 = 0` across a `2Y / 5Y / 10Y` three-leg structure reproduces the exact PCA butterfly weights compared in [50/50 vs regression-weighted 2s5s10s butterfly curve trades](/guides/fifty-fifty-vs-regression-weighted-2s5s10s-butterfly-curve-trades), while converting zero-curve KRD inputs into tradable par OIS swaps uses [par swap DV01 vs zero-rate PV01 Jacobian math](/guides/par-swap-dv01-vs-zero-rate-pv01-jacobian-curve-transformation). Inspect multi-asset execution analytics in [TradingView institutional charting profiles](/software/tradingview) and explore all 20 guides on the [Modified Duration & Key-Rate Hedging Topic Hub](/topics/bond-duration-convexity-math).

## Frequently asked questions

### Should PCA yield-curve covariance be computed on the covariance matrix (bps^2) or the correlation matrix?

Institutional fixed-income desks almost always run PCA directly on the basis-point COVARIANCE matrix (bps^2) because DV01 is already expressed in dollars per basis point, making dot products e_k^T * DV01 directly interpretable without re-scaling by tenor standard deviations.

### Why does PC1 Level peak around the 5-year tenor rather than at 2 years or 30 years?

Very short maturities (1M–1Y) are frequently pinned by central bank policy target bands during on-hold regimes, while 30-year yields mean-revert toward long-run inflation anchors—leaving the 3Y–7Y monetary-policy transition belly as the highest-volatility segment!

### What is the trade-off between hedging via 12 Key-Rate Duration (KRD) buckets versus 3 PCA factors?

PCA compresses risk into 3 liquid benchmark contracts (minimizing bid-ask costs) but relies on historical covariance stability. KRD requires zero statistical assumptions (pure accounting partition) and catches localized single-tenor distortions during crises.

### How swiftly do PCA eigenvectors shift when central banks hit the Zero Lower Bound (ZLB)?

When policy rates hit 0% (2009–2015 or 2020–2021), 3M–2Y yield volatility collapses near zero, squashing front-end PC1 loadings toward zero and shifting the PC2 Slope pivot point outward from ~6 years toward 10 years—making 90-to-180-day rolling windows essential!


## Source URLs

- https://jfi.pm-research.com/content/1/1/54
- https://www.newyorkfed.org/research/staff_reports/sr397.html
- https://www.bis.org/publ/bppdf/bispap25b.pdf

## Citation

- APA: Elena Rostova, FRM, CQF. (2026). Principal Component Analysis (PCA) Level, Slope & Curvature DV01 Hedging Math. QuantCalc Tools. https://quantcalctools.com/guides/principal-component-analysis-level-slope-curvature-dv01-hedges
- Canonical article: [Principal Component Analysis (PCA) Level, Slope & Curvature DV01 Hedging Math](https://quantcalctools.com/guides/principal-component-analysis-level-slope-curvature-dv01-hedges)
- Site research index: [QuantCalc Tools](https://quantcalctools.com/llms-full.txt)
