Regularized Pricing & Risk Models¶
https://www.youtube.com/watch?v=aga-Tak3c3M&list=PLUl4u3cNGP63ctJIEC1UnZ0btsphnnoHR
Bond Duration¶
Sensitivity of bond price \((\ln P)\) to bond yield \(y\)
- Duration gives the “weighted time”
- Duration of zero coupon bond = maturity
- Duration of regular coupon bond < maturity
- As there is only a fixed \(y\) for all payment dates, the duration is a sensitivity to “parallel” move
Good measure for price changes for small variation in yield
Second derivative required for large changes in yield
$$ \begin{aligned} P &= \sum \limits_{i=1}^n e^{-y t_i} C_i \ P_y &= \dfrac{\partial P}{\partial y}= - \sum \limits_{i=1}^n t_i e^{-y t_i} C_i \ \implies d &= \dfrac{P_y}{P} \ c &= \dfrac{\partial^2 P}{\partial y^2} = \sum \limits_{i=1}^n t_i^2 e^{-y t_i} C_i \ \end{aligned} $$ where
- \(d=\) Bond Duration
- \(c =\) Bond convexity: Always positive
- \(P=\) price of bond
- \(y=\) yield of bond
- \(C_i = i\)th cashflow
Swaps¶
Valuing fixed and float legs of the swap
Swap can be hedged with bond $$ \begin{aligned} \text{PV}\text{fixed} &= \sum_i C \delta_i \Alpha_i = C \sum_i w_i \ \text{PV}\text{float} &= \sum_i C r_i \delta_i \Alpha_i = \sum_i r_i w_i \ \text{PV}\text{fixed} &= \text{PV}\text{float} \ \implies C &= \dfrac{\sum_i r_i w_i}{\sum_i w_i} \end{aligned} $$ where
- \(c =\) swap rate (fixed leg coupon)
- Weighted sum of forward rates (assuming same frequency of payments of fixed & floating legs)
- \(\Alpha_i =\) discount factor for payment date \(i\)
- \(\delta_i =\) day count fraction
- \(r_i =\) forward rate (floating rate of future payment)
Yield Curve¶
- Select input instruments
- Choose interpolation
- Interpolation space (daily forward rates, zero rates, etc)
- Spline (piece-wise constant, linear, tension spline, etc)
- Knot points and model parameters
- Calibrate
- Solve for spline parameters such that input instruments are re-priced at par
Bond Spread¶
where
- \(\Alpha_i =\) discount factor for payment date \(i\) computed from curve
- \(s=\) bond spread
- \(t_i =\) future time of payment in years
- \(C_i = i\)th cashflow
If the model is available for typical movements of the curve embedded in \(\Alpha_i\) we can build more effective risk model for bond, rather than using single “parallel” shift mode (bond duration)
Hedging¶
where
- \(r =\) portfolio risk
- \(H =\) hedging portfolio risks
- \(x =\) weights of hedging instruments
- \(F =\) market scenarios (factors)
PCA¶
Use SVD to decompose market movements data \(D\) into principal comments \(P\) and corresponding uncorrelated market dynamics \(U\) with weights \(S\) $$ D = U \cdot S \cdot P^T $$ Use few SVD components with largest singular values - low rank approximation of market data $$ P^T (r + Hx) = 0 $$
PCA Risk Model¶
“Formally” tuned to historical data
Hedge coefficients are unstable, especially if historical window is short
Costly to re-hedge when PC factors change
Instability is coming from PCs corresponding to small singular values
Over-fitting to historical data
NO assumptions of shape of yield curve
Regularized Risk Models¶
Assumption: Forward rates move smoothly $$ H^T R = I \ {\vert \vert L \cdot J \cdot R \vert \vert}^2 \to \min \ R \sim \Big(HH^T + \lambda^2 (L \cdot J)^T \cdot L \cdot J \Big)^{-1} $$ where
- \(J =\) Jacobean matrix translating shifts of yield curve inputs to movements of forward rates
- \(L=\) Smoothness regularity matrix
- \(\lambda =\) regularization parameter
Pricing Model¶
HJM Heath-Jarrow-Morton Model¶
Evolution of forward rates $$ {df}{t, s} = \mu^\beta V(t, s) \rho (t, s) \cdot dB_t^Q $$ where} dt + f_{t, s
- \(f =\) forward rate
- \(\mu=\) drift
- \(\beta=\) model skew factor
- \(\rho=\) Correlation/factor structure
- \(V(t, s)=\) parametric volatility surface
- \(d B_t^Q =\) Brownian motion
Regularized Volatility Surface¶
Challenges¶
- High dimensionality
- Need to calibrate many elements
- Large memory requirement to store matrix
- Relatively small number of calibration instruments
- Under-determined problem
- Sensitivity areas of calibration instruments overlap significantly
- Ill-posed inverse problem
- Unstable, noisy solution
- Need regularity conrtaints
- Has to be smooth to produce realistic prices for similar instruments
IDK¶
Represent volatility surface as linear combination of \(n\) basis functions $$ v = v_0 + \beta x $$ where
- \(v =\) vector of volatility grid elements
- \(\beta=\) matrix corresponding to basis functions
- \(x=\) vector of weights
Make \(n\) equivalent to number of calibration instruments \(M\)
“Formally” unambiguous
Make basis functions piecewise constant matching sensitivity of calibration instruments, 0 otherwise
Sensitivities¶
where
- \(q_{mdl} =\) model price
- \(q_\text{market} =\) market price
- \(q_0 =\) base price
- \(x=\) vector basis functions coefficients