Fixed Income Factor Models: Level, Slope, Curvature, Credit Spread, and Liquidity Factors

Why Factor Models for Fixed Income Risk

A fixed income portfolio's interest rate risk can be measured at every point on the yield curve — a full-curve representation might include 10 or more tenor key rate durations (KRDs) per currency. Factor models compress this multidimensional exposure into a small number of orthogonal risk factors that explain most of the variance, enabling:

• Risk reporting: express portfolio risk in terms of "long level, short slope, neutral curvature" rather than 10 KRD numbers
• Hedging efficiency: hedge three factor exposures rather than 10 tenor points
• Stress testing: apply named scenarios (bear steepener = positive slope factor shock) in factor space
• Return attribution: decompose realized P&L into level contribution, slope contribution, credit spread contribution, and idiosyncratic

The Quant Finance templates use factor decomposition as the core risk framework for fixed income portfolio analysis, consistent with how Bloomberg's BFIF (Bloomberg Fixed Income Factor Model) and BlackRock's Aladdin structure fixed income risk.

The Litterman-Scheinkman Three-Factor Model

Litterman and Scheinkman (1991) applied PCA to US Treasury yield changes and found that three factors explain approximately 99% of total yield variance. This finding has been replicated across G10 government bond markets. The three factors are:

• Level (PC1): parallel shift. Factor loading approximately equal across all maturities (+1 normalized). Explains ~90% of variance. Duration risk is essentially the loading on this factor.
• Slope (PC2): 2s30s. Negative loading at short maturities, positive at long maturities. Explains ~7% of variance.
• Curvature (PC3): butterfly. Positive loading at short and long maturities, negative in the belly (5Y-10Y). Explains ~2% of variance.

• "Litterman-Scheinkman factor decomposition: I have the following weekly yield changes over 13 weeks (a quarter) for 4 Treasury tenors. Approximate historical data (bps): Week 1: 2Y +8, 5Y +7, 10Y +5, 30Y +3. Week 2: 2Y −5, 5Y −4, 10Y −4, 30Y −3. Week 3: 2Y +12, 5Y +11, 10Y +9, 30Y +7. Week 4: 2Y −3, 5Y −2, 10Y −3, 30Y −3. Week 5: 2Y +6, 5Y +7, 10Y +8, 30Y +8. Week 6: 2Y −10, 5Y −9, 10Y −7, 30Y −5. Week 7: 2Y +4, 5Y +5, 10Y +6, 30Y +7. Week 8: 2Y +15, 5Y +10, 10Y +5, 30Y +2. Week 9: 2Y −8, 5Y −5, 10Y −3, 30Y −1. Week 10: 2Y +2, 5Y +3, 10Y +5, 30Y +6. Week 11: 2Y −12, 5Y −10, 10Y −8, 30Y −5. Week 12: 2Y +9, 5Y +8, 10Y +7, 30Y +6. Week 13: 2Y +3, 5Y +4, 10Y +5, 30Y +6. Perform PCA: (1) construct the 13×4 matrix of weekly changes, (2) compute the 4×4 covariance matrix Σ, (3) compute eigenvalues and eigenvectors, (4) what percentage of variance does each factor explain?, (5) interpret the eigenvectors: which is the level factor, slope factor, curvature factor?, (6) compute factor scores for each week (the 'level change' and 'slope change' on each date)."
• "Factor betas for a bond portfolio: I want to express a fixed income portfolio's sensitivity to the three Litterman-Scheinkman factors. Portfolio KRDs: 2Y = $22,000/bp, 5Y = $38,500/bp, 10Y = $31,200/bp, 30Y = $8,800/bp. PCA eigenvectors (normalized): Level: [0.48, 0.51, 0.52, 0.49]. Slope: [−0.58, −0.32, 0.22, 0.72]. Curvature: [0.55, −0.43, −0.43, 0.55]. Factor beta in $/pc = Σ(KRD_i × eigenvector_i): (1) compute level beta = KRD vector dot product with level eigenvector, (2) compute slope beta and curvature beta similarly, (3) interpret: a positive slope beta means the portfolio gains when the curve steepens (long end rises more than short end). Is this portfolio long or short slope risk?, (4) a bear steepener scenario has PC2 (slope) = +50bps equivalent change. Estimate the P&L impact from the slope factor: slope beta × slope change, (5) what KRD adjustment would make the portfolio slope-neutral (slope beta = 0) while keeping the same total duration?"

Nelson-Siegel Factor Model

Nelson-Siegel is an alternative factor parameterization that gives factors direct economic interpretation and ensures a smooth, arbitrage-free curve shape. Rather than extracting factors statistically from historical data (PCA), Nelson-Siegel prescribes factor loadings as functions of maturity.

• "Nelson-Siegel curve fitting: Observed Treasury yields: 3M = 5.20%, 6M = 5.12%, 1Y = 4.95%, 2Y = 4.72%, 3Y = 4.62%, 5Y = 4.55%, 7Y = 4.51%, 10Y = 4.50%, 20Y = 4.55%, 30Y = 4.58%. Fit the Nelson-Siegel model: y(τ) = β0 + β1 × f1(τ, λ) + β2 × f2(τ, λ), where f1(τ, λ) = (1 − e^(−λτ)) / (λτ) and f2(τ, λ) = f1(τ, λ) − e^(−λτ). Use λ = 0.50 (a common starting value where the hump occurs around 7 years). (1) Compute f1 and f2 for each maturity point, (2) use ordinary least squares to estimate β0, β1, β2 that minimize the sum of squared fitting errors, (3) report β0 (level), β1 (slope — negative since short rates exceed long), β2 (curvature — sign indicates hump shape), (4) compute fitted yields at each tenor and the residuals, (5) what does a β1 of −0.60% tell you about the current slope of the curve? (Short rate = β0 + β1 = level − 0.60% = 4.50% − 0.60% = 3.90% if β0 = 4.50%)."
• "Diebold-Li dynamic Nelson-Siegel — factor forecasting: Quarterly observations of Nelson-Siegel factors (estimated from Treasury data): Q1 2023: β0 = 4.08, β1 = −1.45, β2 = −0.62. Q2 2023: β0 = 4.22, β1 = −1.28, β2 = −0.58. Q3 2023: β0 = 4.51, β1 = −1.05, β2 = −0.45. Q4 2023: β0 = 4.62, β1 = −0.85, β2 = −0.38. Q1 2024: β0 = 4.48, β1 = −0.72, β2 = −0.32. Q2 2024: β0 = 4.35, β1 = −0.68, β2 = −0.28. Each factor follows an AR(1) process: β_t = μ + φ(β_{t-1} − μ) + ε_t. (1) Estimate the AR(1) parameters for β0 (level factor) using OLS on the 6 quarterly observations: estimate μ (long-run mean) and φ (persistence), (2) forecast β0 one quarter ahead: what is the expected level factor in Q3 2024?, (3) if the level factor rises (β0 increases), what happens to the yield curve? What is the duration-weighted P&L impact on a $100M portfolio with level beta $95,000/unit-of-β0?, (4) what do the time series of β1 (slope) tell you about the yield curve shape from Q1 2023 to Q2 2024?"

Credit Spread Factors

Beyond rate factors, credit portfolios need credit spread factors. The most common are: (1) IG spread factor — parallel move in investment-grade OAS; (2) HY spread factor — parallel move in high-yield OAS; (3) duration-times-spread (DTS) factor — accounts for the empirical finding that spread volatility scales with the current spread level (higher-spread credits are more volatile in basis point terms). DTS = spread duration × OAS.

• "Credit factor model — IG and HY spread attribution: Portfolio positions: $120M IG financials (OASD 4.2, current OAS 115bps, DTS = 4.2 × 115 = 483), $80M IG industrials (OASD 5.8, OAS 108bps, DTS = 626), $40M HY energy (OASD 3.5, OAS 480bps, DTS = 1,680), $20M HY retail (OASD 2.8, OAS 650bps, DTS = 1,820). Factor definitions: IG Spread Factor = parallel 1bp move in all IG OAS. HY Spread Factor = parallel 1bp move in all HY OAS. IG-HY correlation = 0.68. Factor volatilities: σ_IG = 3.8bps/day, σ_HY = 14.5bps/day. Calculate: (1) portfolio IG spread DV01 = Σ(OASD_i × Notional_i × 0.0001) for IG positions, (2) portfolio HY spread DV01 similarly, (3) factor VaR: IG spread VaR (1-day 99%) and HY spread VaR, (4) total credit spread VaR with correlation, (5) which factor dominates total credit VaR despite the smaller HY allocation?"
• "Duration-times-spread (DTS) factor model: DTS is the most predictive variable for spread returns. Empirically, the percentage change in OAS is more stable than the basis-point change. Return_spread ≈ −OASD × ΔOAS ≈ −OASD × OAS × (ΔOAS/OAS) = −DTS × (ΔOAS/OAS). For the portfolio above: (1) compute DTS for each position (OASD × OAS), (2) portfolio weighted average DTS = Σ(DTS_i × Notional_i) / total notional, (3) if the daily proportional spread change (ΔOAS/OAS) has σ = 3.2% for IG and 5.5% for HY, what is the daily return vol from credit spread risk for each segment? (Return vol ≈ DTS × proportional vol), (4) compare the DTS-based spread VaR to the flat spread DV01 VaR computed earlier — when do they differ most? (Answer: when credits with very different OAS levels are mixed — DTS captures the higher volatility of wide-spread names), (5) how would you construct a DTS-neutral portfolio (zero net spread duration-times-spread) while maintaining positive carry?"

Multi-Factor Return Decomposition

The most useful application of factor models is after-the-fact: decomposing realized portfolio return into factor contributions for attribution reporting to investment committees.

• "Weekly portfolio return attribution using factor model: In the past week: Level factor change = +12bps (rates rose in parallel). Slope factor change = +4bps (curve steepened). Curvature factor change = −2bps (belly rallied relative to wings). IG spread factor = +8bps (IG OAS widened). HY spread factor = +25bps (HY OAS widened more). Portfolio factor exposures: Level beta = −$140,000/bp (long duration, hurts from rising rates). Slope beta = +$28,000/bp (long slope, benefits from steepening). Curvature beta = −$15,000/bp (short belly, benefits from belly rally). IG spread beta = −$85,000/bp (long IG credit, hurts from widening). HY spread beta = −$18,000/bp (long HY credit, hurts from HY widening). Compute: (1) P&L from each factor: factor beta × factor change, (2) total model-attributed P&L, (3) residual P&L (total actual P&L was −$1.42M — what is the unexplained/idiosyncratic component?), (4) express each factor's contribution as a percentage of total P&L, (5) what does a large idiosyncratic component (residual) tell you about the model's explanatory power this week?"

Commercial Factor Models: Bloomberg BFIF

Bloomberg's Fixed Income Factor Model (BFIF) is widely used by asset managers and risk teams for standardized factor risk decomposition. BFIF factors include: government yield factors by currency and tenor, IG spread factors by sector and rating, HY spread factors, EMD (Emerging Market Debt) factors, securitized product factors, and currency factors for global portfolios. BFIF also reports factor volatilities, factor correlations, and predicted tracking error against standard indices (Bloomberg Agg, Bloomberg IG, etc.).

• "Bloomberg BFIF-style factor risk report interpretation: Suppose your risk system reports the following factor exposures and contributions to predicted 1-year tracking error (vs. Bloomberg US Corporate IG Index) for a $500M IG portfolio: (1) Government rate risk: factor exposure = +8bps duration mismatch (longer than index), contribution to TE = 45bps/year; (2) Credit spread risk — financials: −5bps OAS duration vs. index (underweight financials), contribution = 28bps/year; (3) Credit spread risk — industrials: +12bps OAS duration vs. index (overweight industrials), contribution = 62bps/year; (4) Credit spread risk — utilities: +2bps OAS duration vs. index, contribution = 11bps/year; (5) Issuer-specific risk (idiosyncratic): contribution = 32bps/year; (6) Total predicted TE: √(45² + 28² + 62² + 11² + 32²) ≈ 86bps/year. Analyze: (a) which factor is the largest contributor to tracking error? (b) if you want to reduce total TE from 86bps to under 60bps without changing the government rate position, what is the maximum allowable industrials OAS duration mismatch? (c) how many bps of industrials OAS duration mismatch contributes 1bp of TE? (d) what does a TE of 86bps per year mean in dollar terms for a $500M portfolio?"

Where to Start

Fixed income factor models underpin risk reporting, portfolio construction, and return attribution at every major asset manager. The Quant Finance templates include tools for PCA-based factor extraction from yield time series, Nelson-Siegel curve fitting, and multi-factor return attribution. The underlying duration and KRD framework is covered in Bond Duration and Convexity with AI. For applying factor models to VaR calculation, see Fixed Income VaR. For factor-based stress scenarios, see Interest Rate Stress Testing. The full Fixed Income Analysis AI guide provides the workflow integration.

Frequently Asked Questions

How many factors are needed to explain bond portfolio risk?

For US Treasury or G10 government bond risk, three factors (level, slope, curvature) explain approximately 99% of yield variance — so three factors are sufficient for government bond portfolios. For corporate bond portfolios, you need the three rate factors plus at least 2-4 credit spread factors (IG spread, HY spread, sector-specific spread). For global multi-currency fixed income, you need rate factors per currency (3 per G10 currency = 30+ factors) plus cross-currency and credit spread factors. Commercial models like BFIF and Aladdin use 50-100+ factors for comprehensive multi-sector global fixed income portfolios. The marginal benefit of additional factors diminishes quickly — the 4th rate factor typically explains less than 0.5% of variance — so the question is whether the additional complexity of modeling it separately justifies the incremental accuracy for your specific portfolio.

What is the difference between PCA factors and Nelson-Siegel factors for fixed income?

Both are three-factor models of the yield curve, but they differ in how factors are extracted: PCA is purely data-driven — factors are computed from the covariance structure of historical yield changes and are orthogonal by construction. Nelson-Siegel is model-driven — factors are specified as mathematical functions of maturity (1, loading function 1, loading function 2) and are not orthogonal. PCA factors are harder to interpret (the first eigenvector of Treasury yields is "mostly parallel but slightly tilted") while Nelson-Siegel factors have clear economic labels (β0 = long-run rate, β1 = short-long spread, β2 = hump magnitude). PCA gives a better statistical fit to historical data; Nelson-Siegel gives a smoother curve that always has sensible shape. For risk management, PCA is preferred because orthogonality enables clean factor VaR aggregation. For curve fitting and central bank reporting, Nelson-Siegel is preferred because it always produces realistic curve shapes without requiring historical estimation.

How do fixed income factor models handle negative rates?

Standard factor models (Litterman-Scheinkman PCA, Nelson-Siegel) are parameter models that work on yield levels — they have no constraint preventing negative factor scores, and so they naturally accommodate negative rates without modification. The issue arises in term-structure models used for derivatives pricing (Vasicek, CIR): the CIR model prevents negative rates by construction (rates are mean-reverting with a square-root diffusion term that prevents rates from going below zero), while Vasicek allows negative rates but treats them as a rare large deviation. For factor model risk analysis, the relevant question is whether negative-rate bond pricing formulas need adjustment — they do not, since bond pricing is just present value of cash flows at the market yield, which can be negative (zero-coupon bond prices can exceed par when yields are negative). European government bond markets (German Bunds, Swiss Confederation bonds) provided extensive negative-yield data from 2015-2022 that validated factor model behavior at negative rate levels.

Can factor models be used for bond portfolio optimization?

Yes — factor models enable mean-variance optimization where the covariance matrix is structured by factors rather than estimated directly from bond return history (which is too short for most individual bonds). The procedure: (1) estimate factor covariance matrix Σ_F (from yield history — much longer than bond price history), (2) compute factor loadings B for each bond (from KRD-to-eigenvector mapping), (3) portfolio return covariance = B Σ_F B^T + D (specific risk). Then optimize: minimize portfolio variance (B^T w)^T Σ_F (B^T w) subject to target factor tilts (e.g., minimum slope exposure, target level beta consistent with benchmark duration ±0.5 years). Factor-constrained optimization is preferred over unconstrained bond-level optimization because it prevents factor concentration while allowing intentional active bets on credit spread or curve positioning.