Unit 9, lesson 72

Collateral discounting and multiple curves

Why full cash collateral makes the collateral rate the discount rate, how OIS discounting, projection curves and the tenor basis enter multi-curve swap valuation, and how a compounded overnight rate accrues.

  • Video 72
  • Rate options, smiles and collateral curves

A two-year swap on a term rate index is fully collateralized with cash that earns the overnight rate. Which curve discounts its payments, which curve projects its floating rate, and how much goes wrong if one curve is asked to do both jobs? This lesson derives, under stated idealizations, why full collateral makes the collateral rate the discount rate, builds that discount curve from overnight indexed swaps, separates it from the projection curve of the index, and computes a compounded overnight rate from daily fixings.

Key terms

Collateral discounting

Definition Work in one currency with a derivative whose value to its holder is Vt=V(t,Xt)V_t = V(t, X_t), where XX is a Markov state with generator L\mathcal{L} under the pricing measure Q\mathbb{Q}, and L\mathcal{L} already contains the financing of the hedge instruments, for example a stock financed by repo drifts at its repo rate less its dividend yield. Assume continuous, symmetric variation margin in cash, with collateral balance CtC_t that the holder has received when Ct>0C_t > 0 and posted when Ct<0C_t < 0, interest on it at the collateral rate ctc_t, paid by the party that holds the collateral, unsecured borrowing and lending at one rate ftf_t, reuse of received collateral, and no thresholds, minimum transfer amounts, margin gaps or defaults. In this simplified replication ledger the part V−CV - C of the value that collateral does not finance is funded at ff, and the collateral itself costs cc, so

∂tV+LV−f(V−C)−cC=0,\begin{aligned} &\partial_t V + \mathcal{L}V \\[0.4em] &\quad - f(V - C) - cC = 0, \end{aligned}

which is the same as

∂tV+LV−fV+(f−c)C=0.\begin{aligned} &\partial_t V + \mathcal{L}V - fV \\[0.4em] &\quad + (f - c)C = 0. \end{aligned}

With full collateral, C=VC = V, the two terms with ff cancel and

∂tV+LV−cV=0.\partial_t V + \mathcal{L}V - cV = 0.

Under the regularity and growth conditions of the Feynman-Kac theorem, the solution with terminal value VT=HV_T = H is

Vt=EQ[e−∫tTcs dsH∣Ft],V_t = \mathbb{E}^{\mathbb{Q}}\Bigl[e^{-\int_t^T c_s\, ds} H \Bigm| \mathcal{F}_t\Bigr],

so a fully collateralized payoff is discounted at its collateral rate. Write Btc=exp⁡(∫0tcs ds)B^c_t = \exp(\int_0^t c_s\, ds) for the collateral account. Today’s collateral discount factor Dc(0,T)D^c(0,T) is the value of a fully collateralized payment of 11 at TT,

Dc(0,T)=EQ[1/BTc],D^c(0,T) = \mathbb{E}^{\mathbb{Q}}\bigl[1/B^c_T\bigr],

which equals exp⁡(−∫0Tc(s) ds)\exp(-\int_0^T c(s)\, ds) when the collateral rate is deterministic. The cancellation removes ff from this equation only. It does not change the drifts in L\mathcal{L}, which still need their own financing assumptions, and it fails as soon as one of the assumptions above fails.

In plain words Holding a trade worth VV ties up VV in cash, or frees that much when VV is negative. When collateral covers the whole value and costs the collateral rate, nothing is left to fund at the unsecured rate, so the trade’s value grows at the collateral rate and its cash flows are discounted at that rate. Calling a contract collateralized is not enough, because the collateral has to work exactly this way.

Example Take a constructed payment of 100100 at T=2T = 2 with constant rates c=0.03c = 0.03 and f=0.05f = 0.05. The state plays no role, so LV=0\mathcal{L}V = 0 and the value equation is an ordinary differential equation in tt. With full collateral it reads ∂tV=0.03V\partial_t V = 0.03V, and

V0=100e−0.06≈94.176453.V_0 = 100e^{-0.06} \approx 94.176453.

Without collateral, C=0C = 0, and this symmetric funding model, which ignores default, gives ∂tV=0.05V\partial_t V = 0.05V and V0=100e−0.10≈90.483742V_0 = 100e^{-0.10} \approx 90.483742. When f≠cf \ne c, the bracket −fV+(f−c)C-fV + (f - c)C equals −cV-cV only when C=VC = V. For instance, with f=0.06f = 0.06, c=0.03c = 0.03, V=10V = 10 and C=VC = V it is −0.6+0.3=−0.3=−cV-0.6 + 0.3 = -0.3 = -cV.

Compounded overnight rate

Definition Let an accrual period contain the business days j=1,…,nj = 1, \dots, n, let rjr_j be the overnight rate published for day jj as an annual simple rate, and let dj≥1d_j \ge 1 be the number of calendar days that rjr_j applies to, more than one over weekends and holidays. Under the actual/360 day count the accrual factor and the compounded rate are

G=∏j=1n(1+rjdj360),R=G−1D/360,D=∑j=1ndj.\begin{gathered} G = \prod_{j=1}^{n} \Bigl(1 + \frac{r_j d_j}{360}\Bigr), \\[0.4em] R = \frac{G - 1}{D/360}, \\[0.4em] D = \sum_{j=1}^{n} d_j. \end{gathered}

A loan or swap leg with notional NN then pays N(G−1)=NRD/360N(G - 1) = NRD/360 for the period. The rate is backward-looking. It is known only after the last fixing of the period, while a term rate is fixed at the start. Lookbacks, observation shifts, lockouts, payment lags and the day-count basis, such as 360360 or 365365, come from the contract. Expanding the product shows that G−1G - 1 is the sum ∑jrjdj/360\sum_j r_j d_j/360 plus products of two or more of the terms rjdj/360r_j d_j/360. So if every rj≥0r_j \ge 0, the compounded rate is at least the day-weighted arithmetic rate ∑jrjdj/D\sum_j r_j d_j / D, with equality exactly when at most one rjr_j is positive. With fixings of mixed sign the inequality can fail. An example of an overnight benchmark is SOFR, a secured rate based on overnight repurchase agreements collateralized by Treasury securities. The benchmark is the daily rate, and the compounding of its observations is defined by the contract.

In plain words Each day’s interest is added to the balance and earns interest itself from the next fixing on, while a rate that covers a weekend earns simple interest over all of its days. The compounded rate is the single simple rate over the whole period that pays the same interest.

Example In a constructed four-day period on actual/360, the rate is 0.030.03 for one day, 0.040.04 for two days and 0.050.05 for one day. The three factors are 1+0.03/3601 + 0.03/360, 1+0.08/3601 + 0.08/360 and 1+0.05/3601 + 0.05/360, and

G−1≈0.000444505404,R=G−14/360≈0.0400054863,\begin{aligned} G - 1 &\approx 0.000444505404, \\[0.4em] R &= \frac{G - 1}{4/360} \\[0.4em] &\approx 0.0400054863, \end{aligned}

about 4.0005494.000549 percent. The day-weighted arithmetic rate is exactly (0.03+0.08+0.05)/4=0.04(0.03 + 0.08 + 0.05)/4 = 0.04, and the difference of about 0.05490.0549 bp comes from the cross products, not from any new rate. Treating the two-day fixing as two daily factors would give about 0.04000660.0400066 instead, which is not how a rate that covers two days accrues.

OIS discounting

Definition An overnight indexed swap, or OIS, on the schedule 0=T0<T1<⋯<Tn0 = T_0 < T_1 < \dots < T_n with accrual fractions δk\delta_k exchanges a fixed rate KK for the compounded overnight rate. At each TkT_k the fixed payer pays δkK\delta_k K and receives δkRk\delta_k R_k per unit of notional, where 1+δkRk1 + \delta_k R_k is the factor GG of the period [Tk−1,Tk][T_{k-1}, T_k]. OIS discounting is collateral discounting for trades whose cash collateral earns that same overnight rate, with the discount curve built from OIS quotes. Assume the hypotheses of collateral discounting with this collateral rate, and in addition that the collateral account grows over each period by exactly the factor that the floating leg compounds, BTkc/BTk−1c=1+δkRkB^c_{T_k}/B^c_{T_{k-1}} = 1 + \delta_k R_k, with no payment lag. Since B0c=1B^c_0 = 1,

EQ[δkRkBTkc]=EQ[1BTk−1c−1BTkc]=Dc(0,Tk−1)−Dc(0,Tk),\begin{aligned} &\mathbb{E}^{\mathbb{Q}}\Bigl[\frac{\delta_k R_k}{B^c_{T_k}}\Bigr] \\[0.4em] &= \mathbb{E}^{\mathbb{Q}}\Bigl[\frac{1}{B^c_{T_{k-1}}} - \frac{1}{B^c_{T_k}}\Bigr] \\[0.4em] &= D^c(0,T_{k-1}) - D^c(0,T_k), \end{aligned}

so the floating leg telescopes to 1−Dc(0,Tn)1 - D^c(0,T_n), and the par OIS rate is

Kn=1−Dc(0,Tn)∑k=1nδkDc(0,Tk).K_n = \frac{1 - D^c(0,T_n)}{\sum_{k=1}^{n} \delta_k D^c(0,T_k)}.

Solving these equations one maturity at a time,

Dc(0,Tn)=1−Kn∑k<nδkDc(0,Tk)1+Knδn,\begin{aligned} &D^c(0,T_n) \\[0.4em] &= \frac{1 - K_n \sum_{k<n} \delta_k D^c(0,T_k)}{1 + K_n \delta_n}, \end{aligned}

bootstraps the collateral curve from quoted OIS rates, as in Bootstrapping discount curves. The phrase OIS discounting names this matched setup and nothing more. A trade with other collateral terms, such as collateral in another currency, securities as collateral, a spread on the collateral rate or a threshold, needs its own discount curve or further valuation terms.

In plain words When collateral earns the overnight rate and the swap pays that same compounded rate, each floating payment is worth the difference of two collateral discount factors. So OIS quotes pin down the discount curve for trades collateralized at the overnight rate.

Example With annual periods, constructed OIS quotes of K1=0.03/0.97≈0.030928K_1 = 0.03/0.97 \approx 0.030928 for one year and K2=0.06/1.91≈0.031414K_2 = 0.06/1.91 \approx 0.031414 for two years bootstrap to

Dc(0,1)=11+K1=0.97,Dc(0,2)=1−0.97K21+K2=0.94.\begin{aligned} D^c(0,1) &= \frac{1}{1 + K_1} = 0.97, \\[0.4em] D^c(0,2) &= \frac{1 - 0.97 K_2}{1 + K_2} = 0.94. \end{aligned}

The overnight forward rates read from this curve are 1/0.97−1≈0.0309281/0.97 - 1 \approx 0.030928 for the first year and 0.97/0.94−1≈0.0319150.97/0.94 - 1 \approx 0.031915 for the second.

Projection curve

Definition Fix a floating index xx whose fixing LkxL^x_k for the period [Tk−1,Tk][T_{k-1}, T_k] is a simple rate, paid as δkLkx\delta_k L^x_k at TkT_k under the collateral terms of the discount curve DcD^c, with EQ[∣Lkx∣/BTkc]<∞\mathbb{E}^{\mathbb{Q}}[\lvert L^x_k \rvert / B^c_{T_k}] < \infty. The projected forward FkxF^x_k is the number that makes the payment’s value equal to its accrual times its discount factor times the forward,

EQ[δkLkxBTkc]=δk Dc(0,Tk) Fkx.\begin{aligned} &\mathbb{E}^{\mathbb{Q}}\Bigl[\frac{\delta_k L^x_k}{B^c_{T_k}}\Bigr] \\[0.4em] &= \delta_k\, D^c(0,T_k)\, F^x_k. \end{aligned}

Equivalently Fkx=EQc,Tk[Lkx]F^x_k = \mathbb{E}^{\mathbb{Q}^{c,T_k}}[L^x_k], the expectation under the collateralized TkT_k-forward measure, whose density with respect to Q\mathbb{Q} is (1/BTkc)/Dc(0,Tk)(1/B^c_{T_k})/D^c(0,T_k). The projection curve of xx is the collection of these forwards, often stored as pseudo discount factors with Px(0,T0)=1P^x(0,T_0) = 1 and

Px(0,Tk−1)Px(0,Tk)=1+δkFkx.\frac{P^x(0,T_{k-1})}{P^x(0,T_k)} = 1 + \delta_k F^x_k.

These pseudo discount factors are not prices of traded bonds, and absence of arbitrage does not require them to be. Projection curves are usually bootstrapped from swaps on the index after the discount curve is known. Under the OIS hypotheses the compounded overnight index has FkON=(Dc(0,Tk−1)/Dc(0,Tk)−1)/δkF^{\mathrm{ON}}_k = (D^c(0,T_{k-1})/D^c(0,T_k) - 1)/\delta_k, so its projection curve is the discount curve itself, while a term index generally has its own.

In plain words The discount curve says what one currency unit on a given date is worth today. The projection curve says what the index is expected to fix at, in the pricing sense that makes each payment’s value come out right. They answer different questions, and a swap on a term index needs both.

Example With Dc(0,1)=0.97D^c(0,1) = 0.97 and Dc(0,2)=0.94D^c(0,2) = 0.94, suppose the constructed par swap rates on index xx, with annual payments on both legs, are S1=0.04S_1 = 0.04 for one year and S2=0.0811/1.91≈0.042461S_2 = 0.0811/1.91 \approx 0.042461 for two years. The one-year swap gives F1x=0.04F^x_1 = 0.04. The two-year swap requires

0.97⋅0.04+0.94 F2x=S2⋅1.91=0.0811,\begin{aligned} &0.97 \cdot 0.04 + 0.94\, F^x_2 \\[0.4em] &= S_2 \cdot 1.91 = 0.0811, \end{aligned}

so F2x=0.0423/0.94=0.045F^x_2 = 0.0423/0.94 = 0.045. The pseudo discount factors are 1/1.04≈0.9615381/1.04 \approx 0.961538 and 0.961538/1.045≈0.9201320.961538/1.045 \approx 0.920132, different from the collateral discount factors 0.970.97 and 0.940.94.

Multi-curve valuation

Definition Consider a vanilla swap with notional NN on the schedule above, fixed rate KK and floating index xx, with both legs paying at the period ends under the collateral terms of DcD^c and with the projected forwards FkxF^x_k taken as given. Then

PVfloat=N∑k=1nδkDc(0,Tk)Fkx,\begin{aligned} &\mathrm{PV}_{\text{float}} \\[0.4em] &= N \sum_{k=1}^{n} \delta_k D^c(0,T_k) F^x_k, \end{aligned}PVfixed=NKAc,Ac=∑k=1nδkDc(0,Tk),\begin{gathered} \mathrm{PV}_{\text{fixed}} = N K A^c, \\[0.4em] A^c = \sum_{k=1}^{n} \delta_k D^c(0,T_k), \end{gathered}

and the par rate, at which the two legs have equal value, is

Kpar=∑kδkDc(0,Tk)FkxAc.K_{\text{par}} = \frac{\sum_{k} \delta_k D^c(0,T_k) F^x_k}{A^c}.

A payer swap with fixed rate KK is worth NAc(Kpar−K)N A^c (K_{\text{par}} - K) and a receiver swap the negative of that. The floating leg reduces to N(1−Dc(0,Tn))N(1 - D^c(0,T_n)) when every FkxF^x_k equals the forward read from DcD^c itself, as for the overnight index under the OIS hypotheses, and not in general. This is cash-flow accounting with a declared projection convention. A random basis, and the convexity adjustments needed when a payment date differs from the end of its accrual period, are further model ingredients, and giving two curves two names does not supply them.

In plain words Project each floating payment with the index’s own curve, discount every payment with the collateral curve, and add. The single-curve shortcut is the special case in which the index is the very overnight rate that the collateral earns.

Example On the constructed curves the discounted accrual-rate products are 0.97⋅0.04=0.03880.97 \cdot 0.04 = 0.0388 and 0.94⋅0.045=0.04230.94 \cdot 0.045 = 0.0423, which add to 0.08110.0811, so with N=1,000,000N = 1{,}000{,}000 the floating leg is worth 81,10081{,}100. The annuity is Ac=1.91A^c = 1.91, and

Kpar=0.08111.91≈0.042460733,\begin{aligned} K_{\text{par}} &= \frac{0.0811}{1.91} \\[0.4em] &\approx 0.042460733, \end{aligned}

about 4.2460734.246073 percent. Forcing each index forward to be read from adjacent collateral discount factors, as Dc(0,Tk−1)/Dc(0,Tk)−1D^c(0,T_{k-1})/D^c(0,T_k) - 1, gives the telescoped leg 1,000,000 (1−0.94)=60,0001{,}000{,}000\,(1 - 0.94) = 60{,}000 and the par rate 0.06/1.91≈0.0314136130.06/1.91 \approx 0.031413613, which is just the two-year OIS rate of the previous cards. The difference of 21,10021{,}100 in the floating leg shows that discount factors alone do not determine what a term index is expected to fix at. Compare Interest rate swaps and forward rate agreements, where one curve does both jobs.

Tenor basis

Definition The tenor basis of index xx over index yy for the period [Tk−1,Tk][T_{k-1}, T_k] is the difference of their projected forwards,

bk=Fkx−Fky,b_k = F^x_k - F^y_k,

in rate units and usually quoted in basis points, with 11 bp =0.0001= 0.0001. Most often yy is the compounded overnight index, whose forwards come from DcD^c under the OIS hypotheses, or a term index of another tenor in the same currency. A basis swap exchanges the two floating legs, one of them with a spread ss added. When both legs share one schedule and the discount curve DcD^c, the par spread on the yy leg is

s=∑kδkDc(0,Tk) bkAc,s = \frac{\sum_k \delta_k D^c(0,T_k)\, b_k}{A^c},

a weighted average of the period bases with the positive weights δkDc(0,Tk)/Ac\delta_k D^c(0,T_k)/A^c, which add up to one. Legs with different payment frequencies need their own schedules in each sum. In this lesson the basis is an input read from quotes. Its economic sources and its random dynamics lie outside the model.

In plain words Two floating rates for the same period in the same currency need not have the same forward. The basis measures that gap, and a single curve cannot express it.

Example In the constructed example the overnight forwards are about 0.0309280.030928 and 0.0319150.031915, so the basis of index xx over the overnight index is about 0.0090720.009072 and 0.0130850.013085, or 90.7290.72 and 130.85130.85 bp. The discounted bases 0.97⋅0.009072≈0.00880.97 \cdot 0.009072 \approx 0.0088 and 0.94⋅0.013085≈0.01230.94 \cdot 0.013085 \approx 0.0123 add up to 0.02110.0211, so the par spread over the overnight leg is

s=0.02111.91≈0.011047,s = \frac{0.0211}{1.91} \approx 0.011047,

about 110.47110.47 bp. On the notional 1,000,0001{,}000{,}000 the spread leg is worth 1,000,000⋅1.91⋅s1{,}000{,}000 \cdot 1.91 \cdot s, which is exactly 21,10021{,}100, the amount by which the single-curve shortcut undervalues the floating leg on index xx. The basis here is constructed, not an observed market quote.

Check yourself

  1. A loan of 10,000,00010{,}000{,}000 accrues at the overnight rate over a Thursday, a Friday and the following Monday on actual/360. The fixings are 0.050.05 for Thursday, which covers one day, 0.0520.052 for Friday, which covers three days through the weekend, and 0.0510.051 for Monday, which covers one day. Find the interest, the compounded rate and the day-weighted arithmetic rate.
Show the answer

The factors are 1+0.05/360≈1.0001388891 + 0.05/360 \approx 1.000138889, 1+0.156/360≈1.0004333331 + 0.156/360 \approx 1.000433333 and 1+0.051/360≈1.0001416671 + 0.051/360 \approx 1.000141667. Their product minus one is

G−1≈0.000714030147,G - 1 \approx 0.000714030147,

so the interest is about 10,000,000⋅0.000714030≈7140.3010{,}000{,}000 \cdot 0.000714030 \approx 7140.30. The period has D=5D = 5 days, and the compounded rate is (G−1)⋅360/5≈0.0514102(G - 1) \cdot 360/5 \approx 0.0514102. The day-weighted arithmetic rate is (0.05+0.156+0.051)/5=0.0514(0.05 + 0.156 + 0.051)/5 = 0.0514, which would give simple interest of about 7138.897138.89. Compounding adds about 1.411.41, because each day’s interest earns interest over the later days. The Friday rate enters once, as 1+0.052⋅3/3601 + 0.052 \cdot 3/360, not as three daily factors.

  1. On the constructed curves Dc(0,1)=0.97D^c(0,1) = 0.97 and Dc(0,2)=0.94D^c(0,2) = 0.94 with index forwards F1x=0.04F^x_1 = 0.04 and F2x=0.045F^x_2 = 0.045, a payer swap on index xx entered earlier at K=0.04K = 0.04 has two annual payments left on the notional 1,000,0001{,}000{,}000. Value it. A colleague values the floating leg as 1,000,000 (1−0.94)1{,}000{,}000\,(1 - 0.94). What value does the colleague get, and which assumption is wrong?
Show the answer

The fixed leg is worth 1,000,000⋅0.04⋅1.91=76,4001{,}000{,}000 \cdot 0.04 \cdot 1.91 = 76{,}400 and the floating leg 81,10081{,}100, so the payer swap is worth

81,100−76,400=4,700=NAc(Kpar−K),\begin{aligned} &81{,}100 - 76{,}400 = 4{,}700 \\[0.4em] &= N A^c (K_{\text{par}} - K), \end{aligned}

with Kpar≈0.042461K_{\text{par}} \approx 0.042461. The colleague gets 60,000−76,400=−16,40060{,}000 - 76{,}400 = -16{,}400, a value of the wrong sign. The shortcut assumes that the index forwards equal the overnight forwards read from the collateral curve, that is, a zero tenor basis, while here the basis is about 90.7290.72 and 130.85130.85 bp. The colleague’s number is the value of a payer OIS at the fixed rate 0.040.04, a different contract.

  1. A payment of 1,000,0001{,}000{,}000 at T=2T = 2 satisfies the idealized collateral assumptions except that the collateral covers only half of the value, C=12VC = \tfrac12 V. The rates are constant, c=0.03c = 0.03 and f=0.05f = 0.05. Find its value and compare it with full collateral and with no collateral in the symmetric funding model. Which condition makes the unsecured rate drop out?
Show the answer

With LV=0\mathcal{L}V = 0 and C=12VC = \tfrac12 V the value equation is

∂tV=fV−(f−c)12V=0.04 V,\begin{aligned} \partial_t V &= fV - (f - c)\tfrac12 V \\[0.4em] &= 0.04\,V, \end{aligned}

so the effective discount rate is the average 12c+12f=0.04\tfrac12 c + \tfrac12 f = 0.04 and V0=1,000,000 e−0.08≈923,116.35V_0 = 1{,}000{,}000\,e^{-0.08} \approx 923{,}116.35. Full collateral gives 1,000,000 e−0.06≈941,764.531{,}000{,}000\,e^{-0.06} \approx 941{,}764.53, and no collateral gives 1,000,000 e−0.10≈904,837.421{,}000{,}000\,e^{-0.10} \approx 904{,}837.42 in this model, which ignores default and treats borrowing and lending alike. The unsecured rate drops out only when C=VC = V at every time, with the margin continuous, in cash in the same currency, paid interest at cc, free to reuse, and without thresholds, gaps or defaults. With a fraction θ\theta of the value collateralized the effective rate is θc+(1−θ)f\theta c + (1 - \theta) f, and ff disappears only at θ=1\theta = 1.

How this lesson connects