XVA and CVA: Counterparty Credit Risk, Monte Carlo Exposure Simulation and GPU Architecture (2026)

A single interest rate swap with a single counterparty looks harmless on a trade ticket. Its value is a clean mark-to-market number, and the day-one price is a par rate. Yet the moment you ask “what do I lose if this counterparty defaults while the swap is in my favour?”, the single number becomes a probability distribution over thousands of future dates, and the bank needs a computer farm to price it. That adjustment layer is what the industry calls XVA, and it is now one of the most computationally expensive workloads in any investment bank.

XVA matters right now because the cost of ignoring it has been written into accounting standards, regulatory capital rules and trading desk P&L. At the same time, the arithmetic is brutal: a bank with tens of thousands of netting sets must simulate market scenarios over decades, reprice every trade on every path, and still produce risk sensitivities before the morning meeting. This post is a systems-level walk-through of how that works.

You will leave with a clear definition of each valuation adjustment, a worked CVA calculation you can reproduce, a runnable NumPy Monte Carlo sketch, and a reference architecture for nightly and intraday XVA on GPUs with adjoint algorithmic differentiation (AAD).

What this covers: the XVA family and why it exists, exposure metrics and netting, the Monte Carlo pipeline, GPU and AAD economics, wrong-way risk, nightly versus intraday architecture, and the controls that make results reproducible.

Systems and architecture analysis only. Nothing here is investment, trading or regulatory advice.

Context and Background

Before 2007, most dealers priced an over-the-counter (OTC) derivative off a risk-free discount curve and treated the counterparty as if it could not fail. The credit crisis broke that assumption in two ways. First, defaults and near-defaults of dealers showed that counterparty credit risk is real. Second, and less obviously, the market value of that risk moved violently even for counterparties that never defaulted. The Basel Committee observed that during the crisis a large share of counterparty credit risk losses came from the deterioration of counterparties’ creditworthiness rather than from outright defaults (I recall the figure as roughly two thirds of losses, but I could not re-verify the exact number this run, so treat it as unverified). That mark-to-market volatility is exactly what a credit valuation adjustment captures.

Three forces then turned a niche quant topic into an industry. Accounting standards, notably IFRS 13 on fair value measurement and its US GAAP counterpart ASC 820, require that the fair value of a derivative reflect the credit risk of the counterparty and, for liabilities, the entity’s own non-performance risk. Prudential regulation under Basel III added an explicit capital charge for the volatility of CVA. And derivatives market reform pushed standardised products to central counterparties while raising the cost of the remaining bilateral trades through mandatory margin rules.

The result is a family of adjustments, collectively written XVA, that sit on top of the “clean” risk-free price. A useful mental model: the clean price is what the derivative is worth in a frictionless world, and XVA is the sum of the frictions: credit, funding, collateral and capital. Researchers still argue about which pieces belong in an accounting fair value and which belong only in a desk’s pricing. John Hull and Alan White, for example, argue that funding value adjustment should not appear in reported mark-to-market values, while accepting that it can be a real cost for pricing decisions; they also point out that a dealer’s own funding spread partly compensates lenders for default risk, which creates an offsetting benefit (Hull and White, XVAs: Theory and Practice).

For the engineer, the debate matters less than the consequence. Each adjustment is an expectation over future exposure, so all of them are computed from the same expensive object: a cube of simulated portfolio values indexed by scenario, time and netting set. That shared cube is the architectural pivot of this entire post. If you already run a risk engine, the neighbouring problems are covered in our posts on Value-at-Risk and Expected Shortfall engine architecture and on collateral and margin management.

XVA Architecture: From Exposure Cube to Every Adjustment

Direct answer: XVA is a set of valuation adjustments (CVA, DVA, FVA, MVA, KVA) added to the risk-free price of a derivatives portfolio to account for counterparty default, own default, funding, initial margin and capital costs. Every adjustment is an expectation over simulated future exposure, so a bank computes one shared exposure cube and then integrates it against different spreads, funding curves and capital rules.

Figure 1: One exposure engine, many adjustments. Each XVA is a different integral over the same simulated exposure profile.

The diagram shows why XVA is organised as a platform rather than as five separate models. The exposure engine takes netted portfolios as input and emits expected exposure (EE), expected positive exposure (EPE), expected negative exposure (ENE) and potential future exposure (PFE). The five adjustments then differ only in which exposure they integrate and what price of risk they multiply it by. CVA and DVA use credit spreads, FVA uses the bank’s funding spread, MVA uses the cost of funding initial margin, and KVA uses a cost of capital applied to projected regulatory capital.

The family, defined precisely

Credit valuation adjustment (CVA) is the market value of counterparty default risk. In its standard unilateral form it is the risk-neutral expectation of the loss given default applied to the positive exposure at the default time:

CVA = (1 − R) × ∫ EE(t) × D(t) × dPD(t)

where R is the recovery rate, EE(t) is the expected positive exposure at time t (discounted exposure is often written with the discount factor D(t) inside), and dPD(t) is the marginal default probability density of the counterparty. In practice you replace the integral with a sum over a time grid. Note the independence assumption hidden in this formula: it multiplies exposure and default probability as if they were uncorrelated. Section on wrong-way risk below shows what happens when they are not.

Debit valuation adjustment (DVA) is the mirror image: the benefit to the bank from the possibility that it defaults while owing money, computed from the expected negative exposure and the bank’s own credit spread. DVA is controversial. It produces profit when a bank’s credit worsens, which regulators dislike. Basel III capital rules remove DVA gains from the capital base, and desks typically hedge or ignore it for decision-making.

Funding valuation adjustment (FVA) captures the cost or benefit of funding uncollateralised derivative positions at the bank’s own funding spread rather than at the risk-free rate. It is often decomposed into a funding cost adjustment on positive-value positions and a funding benefit adjustment on negative-value positions. As noted, whether FVA belongs in accounting fair value is disputed; many banks nonetheless record it, so you must be able to compute it.

Margin valuation adjustment (MVA) is the cost of funding initial margin posted over the life of a trade, the gap between the return earned on posted margin and the bank’s cost of funding it. Because initial margin is itself computed by a risk model (commonly the ISDA Standard Initial Margin Model, SIMM), MVA requires projecting a margin model forward along every simulated path. That makes it the most expensive adjustment to compute.

Capital valuation adjustment (KVA) is the cost of holding regulatory capital against the trade over its life, discounted at a hurdle rate. It also requires forward-projecting a regulatory calculation, here the capital rules for counterparty credit risk and CVA.

Regulatory capital runs on a different track

Accounting CVA is a price; regulatory CVA is a capital charge for the risk that CVA changes. Under the Basel framework, banks using the standardised approach compute CVA risk capital in one of two ways. The basic approach (BA-CVA) is a formula over counterparty-level stand-alone charges; the standardised approach (SA-CVA) adapts the sensitivity-based market risk method and requires supervisory approval. The following paragraphs rely on the Canadian regulator’s published rules text, which mirrors the Basel text, because it was the most complete primary source I could retrieve this run (OSFI Capital Adequacy Requirements, Chapter 8).

For the reduced BA-CVA, capital is 0.65 times K_reduced, where K_reduced combines the counterparty stand-alone charges using a supervisory correlation of 50%. The stand-alone charge for each counterparty is (1/1.4) × RW × the sum over netting sets of M × EAD × DF, where RW is a supervisory risk weight set by sector and credit quality, M is effective maturity, EAD is exposure at default and DF is a discount factor equal to 1 for banks using the internal model method for counterparty credit risk, and (1 − e^(−0.05M)) / (0.05M) otherwise. The full version of BA-CVA recognises eligible hedges and blends the reduced and hedged results with a parameter of 0.25, so hedging cannot remove most of the charge. SA-CVA sums delta and vega charges across six and five risk classes respectively, excludes default risk and curvature, and requires monthly calculation and a dedicated CVA desk. The CVA multiplier m_CVA is set at 1 in that text, with the supervisor free to demand a higher value. The effective date in the OSFI document header reads November 2026 and January 2027, and the footers carry a September 2025 date, so confirm timing against OSFI’s own notice.

Jurisdictions differ. The UK Prudential Regulation Authority rulebook, in the version dated 1 January 2027, offers SA-CVA with permission, BA-CVA otherwise and a simplified alternative approach in prescribed cases (PRA Rulebook, Credit Valuation Adjustment Risk). The EU’s CRR3 carries the revised framework, which I recall applying from 2025 but did not verify. US rules were unfinalised in the reporting I found, so check current status before relying on them.

The engineering takeaway is that BA-CVA needs only EAD and maturity per netting set, so it is cheap, while SA-CVA needs full CVA sensitivities to curves, FX and volatility, which means a pathwise Monte Carlo engine and AAD. That is the commercial reason banks invest in the machinery described below.

Exposure Profiles, Netting Sets and the Margin Period of Risk

Exposure is the quantity everything else multiplies, so it deserves careful definition. For a portfolio with value V(t) on a given path, the positive exposure is max(V(t), 0): you lose only when the counterparty owes you. Averaging that across paths gives expected exposure EE(t), and averaging EE over time gives expected positive exposure (EPE), the single number most people quote. The negative side, min(V(t), 0) averaged, is the expected negative exposure that feeds DVA. A high quantile of the same distribution, typically 95% or 97.5%, is the potential future exposure (PFE), used for credit limits rather than for pricing.

The exposure hump

For a single interest rate swap, the profile has a characteristic shape. At inception the swap is worth zero. As time passes, rate uncertainty grows with the square root of time, which pushes the distribution of V wider, but the number of remaining cash flows shrinks, which pulls it back. Exposure therefore rises, peaks somewhere around a third to a half of the swap’s life, and decays toward zero at maturity.

Netting sets and why they dominate the number

If a legal netting agreement such as an ISDA Master Agreement is enforceable in the counterparty’s jurisdiction, then on default all trades in the netting set collapse into one net claim. Exposure is then max(sum of V, 0) rather than the sum of max(V, 0). The gap between those two quantities can be an order of magnitude. This means CVA is not additive across trades: adding a trade with a counterparty can reduce the total if it offsets existing positions. It also means data quality at the netting-set level matters more than almost anything else, because a trade assigned to the wrong set changes the answer by a lot.

Figure 2: From trades to collateralised exposure. Enforceable netting collapses trades into one set; the CSA and the margin period of risk then determine how much of the netted value remains unsecured.

Collateral and the margin period of risk

A credit support annex (CSA) to the master agreement governs variation margin (VM) and sometimes initial margin (IM). Its terms (threshold, minimum transfer amount, independent amount, margin call frequency and eligible collateral) determine how much of the exposure is covered. If exposure were fully collateralised in real time, counterparty risk would vanish. It is not, because of the margin period of risk (MPOR): the time between the last successful margin exchange and the moment the defaulted counterparty’s positions are closed out and hedged.

Over the MPOR the portfolio value keeps moving while no collateral arrives. The conventional regulatory floor for bilateral margined derivatives is ten business days, with shorter periods for centrally cleared trades and longer ones when disputes or illiquid positions are present; these are the Basel and SA-CCR conventions as I recall them, and you should confirm exact floors in the rule text for your jurisdiction. In a simulation, this means the engine must compute the portfolio value at both the exposure date and the date a lag earlier, then apply the collateral held as of the earlier date. For a ten-day MPOR on a monthly grid, that doubles the valuation work, which is why many production engines add a second pricing pass or approximate with a diffusion-based “classical plus” method.

The post-crisis uncleared margin rules added a mandatory IM requirement for large participants in phases that, to my recollection, completed in 2022. IM is held in a segregated account and is sized to cover a ten-day 99% move; it shrinks exposure dramatically but creates MVA, as the bank must fund the margin it posts. Our collateral and margin management architecture post covers the operational side of those margin flows.

The Monte Carlo Pipeline, Step by Step

Monte Carlo exposure simulation answers one question: across thousands of plausible futures, what will this portfolio be worth at each future date? The answer must be consistent with today’s market prices (risk-neutral calibration for pricing XVA) and must capture the dependence between rates, FX, credit and other risk factors. Figure 3 shows the stages.

Figure 3: The Monte Carlo exposure simulation pipeline. Each box is a separable service with a defined input and output.

Stage 1: scenario generation

The simulation grid is a set of exposure dates, dense early (daily or weekly for the first month, to capture MPOR effects) and sparse later (quarterly or annually for decades). The model choice is a trade-off between realism and calibration burden. For interest rates, one-factor Hull-White (the extended Vasicek model of Hull and White, 1990) is the workhorse because it fits today’s yield curve exactly, has closed-form zero-coupon bond prices and allows exact simulation of the state variable. Multi-factor versions and the Libor market model appear when the product mix demands richer correlation structure. For FX, a lognormal or local-volatility spot process with interest rate differentials drives the drift, and its correlation to the two rate processes matters. For credit spreads, a reduced-form intensity model, often Cox-Ingersoll-Ross or a lognormal hazard rate, is used when wrong-way risk is in scope.

Under Hull-White the short rate is r(t) = x(t) + φ(t), where x follows an Ornstein-Uhlenbeck process and φ(t) is a deterministic function that fits the curve. The OU process can be stepped exactly over any interval, so there is no discretisation error. This is the model implemented in the snippet below; the parameters (mean reversion 5%, volatility 1%) are illustrative round numbers, not calibrated values.

A design decision with large downstream consequences is the measure. Pricing XVA under the risk-neutral measure is consistent with CVA as a price. Regulatory exposure metrics such as PFE and the internal model method for counterparty credit risk instead use real-world calibrations, with drifts and volatilities estimated from history. Many banks therefore maintain two calibrated models, which doubles the model validation burden; see our post on model risk management under SR 11-7 for how validators approach that.

Stage 2: pathwise pricing, and when it fails

For each path and date the engine needs the value of every trade. Vanilla products such as swaps, forwards and FX forwards have closed-form or cheap analytic values given the simulated state, and can be priced “pathwise” directly. The following listing implements the whole chain for one swap: scenario generation, pathwise pricing, exposure and CVA. It is a teaching sketch, not production code; it ignores day counts, calendars, stub periods and multi-curve basis.

import numpy as np

rng = np.random.default_rng(42)
a, sigma, r0 = 0.05, 0.01, 0.04            # Hull-White 1F, flat 4% curve (illustrative)
dt, T, n_paths = 1 / 12, 5.0, 50_000
n = int(T / dt)
t = np.arange(n + 1) * dt
phi = r0 + sigma**2 / (2 * a**2) * (1 - np.exp(-a * t)) ** 2

# 1) Scenario generation: exact Ornstein-Uhlenbeck step for the x factor
decay = np.exp(-a * dt)
step_sd = sigma * np.sqrt((1 - np.exp(-2 * a * dt)) / (2 * a))
x = np.zeros((n_paths, n + 1))
for i in range(n):
    x[:, i + 1] = x[:, i] * decay + step_sd * rng.standard_normal(n_paths)
r = x + phi                                  # short rate paths

def zcb(i, maturity):
    """Hull-White zero-coupon bond P(t_i, maturity) on every path."""
    B = (1 - np.exp(-a * (maturity - t[i]))) / a
    ln_A = -r0 * (maturity - t[i]) + B * r0 \
           - sigma**2 / (4 * a) * (1 - np.exp(-2 * a * t[i])) * B**2
    return np.exp(ln_A - B * r[:, i])

# 2) Pathwise pricing: 5y annual receive-fixed swap, notional 100m
pay_dates = np.arange(1, 6)
fixed = np.exp(r0) - 1                       # par rate for a 4% flat curve
notional = 100e6
V = np.zeros((n_paths, n + 1))
for i in range(n + 1):
    remaining = pay_dates[pay_dates > t[i] + 1e-9]
    if remaining.size == 0:
        continue
    annuity = sum(zcb(i, p) for p in remaining)
    prev = np.floor(t[i] + 1e-9); nxt = prev + 1
    float_leg = zcb(i, nxt) / zcb(int(round(prev / dt)), nxt) - zcb(i, remaining[-1])
    V[:, i] = notional * (fixed * annuity - float_leg)

# 3) Exposure profile and CVA
EE = np.maximum(V, 0).mean(axis=0)           # expected positive exposure at each date
hazard, recovery = 0.02, 0.40                # flat credit curve (illustrative)
surv = np.exp(-hazard * t)
disc = np.exp(-r0 * t)
dPD = surv[:-1] - surv[1:]
cva = (1 - recovery) * np.sum(0.5 * (EE[1:] * disc[1:] + EE[:-1] * disc[:-1]) * dPD)
print(f"CVA = {cva:,.0f}  ({cva / notional * 1e4:.1f} bp of notional)")

I ran this exact listing with the stated seed on 50,000 paths and it printed a CVA of about 59,206, or 5.9 basis points of notional. That figure is a property of these illustrative inputs, not a market estimate. Changing the seed moves the result by a few hundred units; splitting the paths into five batches gave a standard error of roughly 360, or about 0.6% of the estimate.

Pathwise pricing breaks down for products whose value depends on a decision or on a path-dependent feature: Bermudan swaptions, callable bonds, American-style options and anything with an early exercise right. At each exposure date, the value of such a trade is itself a conditional expectation over the future, so a naive implementation needs a nested simulation: a simulation inside every outer path and date. With 10,000 outer paths, 100 dates and 1,000 inner paths, that is a billion inner paths per trade, which is unworkable.

Stage 3: regression and American Monte Carlo

The standard escape is regression-based pricing, descended from the Longstaff-Schwartz least-squares Monte Carlo (LSMC) method of 2001. The idea is to approximate the conditional expectation at each exposure date by regressing future discounted cash flows onto a small basis of functions of the current state, for example polynomials in the short rate. One regression per date replaces thousands of inner simulations. The same machinery lets a single set of outer paths price American and Bermudan features, which is why practitioners call the approach “American Monte Carlo” in the exposure context.

Stage 4: collateral, netting and aggregation

Once values are in the cube, the engine applies netting by summing values within each set, then applies CSA logic: threshold, minimum transfer amount, rounding, and the lagged collateral balance implied by the MPOR. Only after collateralisation is the positive part taken. The order matters: applying max(·, 0) before netting or before collateral gives a completely different, and wrong, answer. This is the single most common logic bug in home-built exposure engines, and a unit test that compares a hand-computed two-trade example against the engine output is cheap insurance.

A Worked Example: CVA for One Swap

Take the swap from the listing: five-year, annual fixed payments, 100 million notional, receive fixed at the par rate. Assume a flat 2% annual hazard rate and 40% recovery for the counterparty, an uncollateralised position, and a flat 4% continuously compounded discount rate. The loss given default is therefore 60% of exposure. These are assumptions chosen for transparency, not market data.

The Monte Carlo run gives expected positive exposure at the middle of each year. Multiplying by the discount factor, the probability of default in that year and the loss given default produces the CVA contribution of each bucket.

Year bucket (mid-point)	EE at mid-point	Discount factor	Default probability in year	CVA contribution
0.5	908,968	0.9802	1.980%	10,585
1.5	1,523,922	0.9418	1.941%	16,713
2.5	1,506,417	0.9048	1.902%	15,559
3.5	1,196,161	0.8694	1.865%	11,635
4.5	679,450	0.8353	1.828%	6,224
Total				about 60,700

The coarse annual calculation gives about 60,700, within roughly 2.5% of the 59,206 from the monthly grid, and the shape tells the story: the contribution peaks in years two and three where the exposure hump and still-high survival probability overlap. Two diagnostic checks follow. First, the average of EE over the whole life (the EPE) was about 1.14 million, about 1.1% of notional, while the 95th percentile PFE peaked near 6.4 million, so the 95th percentile is roughly four times the expected exposure at the hump. Second, a rule of thumb sanity check: CVA is approximately credit spread times EPE times maturity. A 2% hazard with 40% recovery implies a spread near 120 basis points; 1.2% × 1.14 million × 5 years is about 68,000, the right order of magnitude relative to 59,000 once discounting and the shape of the profile are considered.

A regulatory counterpart, also illustrative

For comparison, compute the BA-CVA reduced capital for the same counterparty using the OSFI-published formula. Assume the counterparty is an investment grade financial with a 5% risk weight, a single netting set with an EAD of 2.0 million (an assumption standing in for an SA-CCR result), effective maturity of 5 years and a non-IMM discount factor. The discount factor is (1 − e^(−0.25)) / 0.25, about 0.885. The stand-alone charge is (1/1.4) × 0.05 × 5 × 2.0 million × 0.885, about 316,000. With a single counterparty, K_reduced equals that figure, and the capital is 0.65 × 316,000, about 205,000, equivalent to roughly 2.6 million of risk-weighted assets after the conventional 12.5 multiplier. Notice that the regulatory charge is more than three times the accounting CVA of the same trade. That gap is not an error: the capital charge covers the volatility of CVA, not its level, and uses supervisory parameters rather than the desk’s. It is precisely this gap that makes hedging CVA with credit default swaps attractive, and why full BA-CVA and SA-CVA recognise hedges.

Computational Cost, GPUs and Adjoint Differentiation

The cost of XVA is the product of four numbers: paths, dates, trades and risk factors to bump. A realistic desk might use 5,000 to 20,000 paths, 100 to 300 dates, tens of thousands of netting sets holding hundreds of thousands of trades, and, for sensitivities, hundreds or thousands of risk factors (every curve tenor, every vol point, every credit spread). I have no benchmark to cite for a specific bank, so treat these as order-of-magnitude planning figures rather than measurements. The point is the multiplication: 10,000 paths × 200 dates × 300,000 trades is 600 billion trade valuations before a single sensitivity is computed.

Why GPUs fit

Monte Carlo exposure is close to the ideal GPU workload. Paths are independent, so parallelism is embarrassing. Pricing a vanilla swap is arithmetic over a vector of state variables with little branching, so GPU lanes stay busy. The working set (path states for one date) fits in device memory, and the arithmetic intensity is high enough to hide memory latency. Where a CPU farm needs thousands of cores and hours, a rack of GPUs can finish the same cube in minutes, and vendors and banks have published speed-ups in the range of tens of times for specific kernels. I have not verified any specific published multiplier this run, so I will not quote one; benchmark your own trade mix.

The bump-and-revalue wall

Sensitivities are where costs explode. The naive approach, bump-and-revalue, perturbs one input at a time and reruns the whole calculation. If CVA needs 500 sensitivities (say, 30 interest-rate tenors across three currencies, 30 credit spread tenors across a hundred counterparties collapsed to buckets, FX deltas, vegas), you run the engine 501 times. A one-hour run becomes three weeks. Even with GPU speed-ups, bump-and-revalue cannot deliver a full risk ladder overnight for a large book.

Adjoint algorithmic differentiation (AAD) removes the wall. Instead of re-running the calculation, AAD records the computational graph of the forward pass and then propagates derivatives backwards from the output (CVA) to every input in one reverse sweep. The headline theoretical result is that the cost of computing the full gradient is bounded by a small constant multiple of the cost of computing the value once, independent of the number of inputs. The constant is often quoted as in the range of roughly three to five; the actual figure depends on implementation, memory traffic and the tape. The idea entered finance through the work of Mike Giles and Paul Glasserman on “Smoking Adjoints: fast Monte Carlo Greeks”, published in Risk in 2006 (Risk.net landmark article), and Giles has a review of the technique in his AAD review notes.

For a CVA with 500 inputs, the arithmetic is stark. Bump-and-revalue costs roughly 501 forward passes. AAD costs one forward pass plus one reverse pass, perhaps four to six forward-pass equivalents in total. That is a gain on the order of one hundred, before any hardware advantage, and the gain grows with the number of risk factors. It is why serious XVA desks pair GPUs for the forward Monte Carlo with AAD for the Greeks.

Sensitivities that matter: CS01, IR delta and friends

Two sensitivities dominate the CVA desk’s risk report. CS01 is the change in CVA for a one basis point move in the counterparty’s credit spread, usually reported per tenor bucket so that the desk can hedge with single-name or index credit default swaps. Interest rate delta is the change in CVA for a one basis point parallel or bucketed shift in the discount and forward curves, hedged with swaps. Beyond those come FX delta, vega to the volatilities of the processes that drive exposure, and cross-gamma. Regulatory SA-CVA needs sensitivities aligned to its prescribed risk classes and buckets, so the sensitivity engine must be able to map the desk’s native risk factors to the regulatory grid; the Basel text supplies that grid, and the mapping layer is a source of bugs worth auditing.

Wrong-Way Risk: Where the Independence Assumption Fails

The basic CVA formula treats exposure and default probability as independent. Wrong-way risk (WWR) is the situation where they are positively dependent: the counterparty is more likely to default precisely when your exposure to it is largest. Right-way risk is the opposite.

Modelling general WWR requires joint simulation of exposure drivers and credit. Two common routes are a stochastic intensity linked to the market factors (for example, a hazard rate that rises with a Brownian component correlated to FX), and a conditional-default approach in which the loss given default and the exposure are conditioned on the default event. Either increases cost, because credit becomes a simulated risk factor with its own volatility and correlation, and the calibration data for correlations are thin. The sharpest effect is in emerging-market FX forwards with local counterparties: if you receive the weakened local currency and the sovereign’s credit deteriorates together, the exposure is largest exactly when the counterparty is least likely to pay.

The architectural implication is that WWR forces the scenario generator and the credit model to share a state vector. In a clean separation of services, that means the credit curve cannot be a static input; it becomes a stochastic path array in the same cube. Plan for it up front, because bolting it on later means rewriting the data model.

Nightly Versus Intraday XVA: Reference Architecture

Traders want XVA at quote time because it changes the price: a client trade that adds 40 basis points of CVA and 20 of FVA should not be quoted at mid. Risk managers want the full cube overnight. These two needs push toward two engines sharing one data foundation.

Figure 4: A two-speed XVA platform. The nightly batch builds and versions the exposure cube; the intraday engine reuses it for incremental and pre-deal pricing.

The nightly batch

The nightly engine consumes a frozen market data snapshot, the end-of-day trade store and CSA data, runs the full Monte Carlo on a GPU grid, and writes a versioned exposure cube. It then computes the official XVA numbers, the AAD sensitivities and the regulatory CVA capital inputs. Three properties are non-negotiable. It must be deterministic, so the same inputs produce the same outputs bit for bit. It must be restartable, because a 90-minute job should not fail at minute 85 and lose everything. And it must finish inside the window, which is why you design for a budget with headroom, say 60% of the window, rather than for the average run.

The intraday engine

The intraday problem is incremental: a new trade arrives or a counterparty’s spread moves, and the desk wants an updated XVA in seconds, not hours. Three techniques make this tractable. First, cube reuse: for a new trade, simulate only the new trade on the cached scenarios and combine with the stored netting-set values, which is exact for CVA given the same scenarios because the netting set value is a sum. Second, sensitivity-based approximation: reprice the book’s XVA for a small market move by applying stored CS01 and delta, accepting second-order error. Third, proxy exposure profiles for pre-deal quotes: use a pre-computed exposure profile library keyed by trade type, tenor and moneyness, scaled by notional. The trade-off is accuracy versus latency, and a good platform lets the desk see which approximation produced each number.

Data lineage and reproducibility

XVA numbers move P&L and capital, so every one must be reproducible months later. That demands more than logging. The minimum contract is: a content hash of the market snapshot, trade store and CSA terms; the code version and model parameters; the random seed and the generator algorithm (and, for GPUs, the kernel and driver versions, since floating-point reduction order can change the last digits); and the cube version produced. With those, an auditor can regenerate yesterday’s CVA and explain a P&L difference between two dates by attributing it to market moves, new trades, model changes or data changes. Without them, the P&L explain gap becomes an unexplained residual that finance and risk will not sign off.

Trade-offs, Gotchas, and What Goes Wrong

Monte Carlo noise in sensitivities. A CVA estimate with 0.6% standard error can have sensitivities with far larger relative noise, because they are differences or derivatives of noisy quantities. Common random numbers (reusing the same random draws across bumped runs) and AAD both reduce the problem, but unexplained sensitivity flicker from night to night usually traces back to a changed random sequence or a path count that varies with cluster availability.

Netting-set and CSA data quality. A wrong threshold, a missing independent amount or a trade mapped to the wrong set changes the answer more than any modelling refinement. Treat legal terms as versioned reference data with the same controls as market data, and reconcile them against the operations system that actually calls margin.

Practical Recommendations

Start from the data model, not the models. Define the exposure cube schema (netting set, date, scenario, value, collateral) and its versioning before choosing a pricing library; everything else is a service that reads or writes that schema. Insist on a netting-set identifier that is stable across the trade store, the legal system and the margin system.

A short checklist to close:

Cube schema and versioning agreed before model selection.
Stable netting-set and CSA identifiers across systems.
Risk-neutral and real-world calibrations separated and documented.
CPU reference implementation retained for regression tests.
Common random numbers and fixed seeds for sensitivity runs.
AAD outputs validated against finite differences on a sample.
Refresh triggers for the intraday cube defined and logged.
Reproducibility standard (bitwise or tolerance) documented and tested.

Frequently Asked Questions

What does XVA stand for and what does it include?

XVA is a collective label for valuation adjustments applied to derivative prices. The “X” is a placeholder for the specific adjustment: credit (CVA), debit (DVA), funding (FVA), margin (MVA) and capital (KVA). Each adjusts the risk-free price for a friction such as counterparty default or the cost of funding collateral. Which adjustments appear in reported accounts versus trading prices varies by firm, and the treatment of FVA and DVA in particular is still debated in the literature.

How is CVA calculated?

CVA is the loss given default multiplied by the discounted expected positive exposure and the marginal default probability, summed across future dates. In practice, a Monte Carlo engine simulates the portfolio value on many paths and dates, applies netting and collateral, averages the positive part to get expected exposure, and combines it with a credit curve from market spreads. Our worked example above gives roughly 59,000 for a 100 million five-year swap under stated assumptions.

What is the difference between CVA, EPE and PFE?

Expected positive exposure (EPE) is the average of positive portfolio values over time and is an input to pricing CVA. Potential future exposure (PFE) is a high quantile, such as the 95th percentile, of the exposure distribution and is used for credit limits and stress. CVA is a price, equal to the expected loss from counterparty default, so it combines exposure with default probability and recovery rather than being an exposure measure itself.

Why do banks use GPUs and AAD for XVA?

Monte Carlo exposure simulation is highly parallel, so GPUs process many paths at once and shorten the nightly run. Sensitivities are the harder problem: bump-and-revalue needs one full run per risk factor, while adjoint algorithmic differentiation (AAD) computes all sensitivities in one reverse sweep at a small constant multiple of the cost of the forward run. Together they make a full risk ladder feasible inside a batch window.

What is the difference between BA-CVA and SA-CVA?

BA-CVA is the basic approach: a formula using supervisory risk weights, exposure at default, maturity and a discount factor, with a reduced version and a full version that recognises hedges. SA-CVA is the standardised approach, which sums delta and vega charges across risk classes using CVA sensitivities and requires supervisory approval and a dedicated CVA desk. SA-CVA is more risk sensitive but needs a sensitivity engine.

What is wrong-way risk in counterparty credit risk?

Wrong-way risk is the positive dependence between exposure to a counterparty and its probability of default, so losses are largest when default is most likely. General wrong-way risk comes from macro links such as a currency crisis that raises the value of a forward and weakens the sovereign. Specific wrong-way risk comes from direct links such as a counterparty selling protection on itself. Standard CVA formulas that assume independence understate it.