1Describe the four stages of a VaR engine (risk factors, scenarios, revaluation, aggregation) and the choice each stage makes.
2Compare full revaluation with delta, delta-gamma and grid approximations on the same scenarios, and say where each breaks.
3Compute historical VaR and ES on different windows, including a stress window, and explain why they differ.
4Explain why scenarios are real-world moves while prices inside them are risk-neutral, and how little the drift matters at one day.
Why it matters
Take the sleeve the desk holds on 31 December 2025 and replay the day after the 2016 referendum. Every stage of the engine is the same, except one: how the options are repriced. In full, they make £3.66m. By their deltas, £5.11m. By delta and gamma together, £7.70m, more than twice the truth. A VaR number is the output of a machine with four stages, each with a choice in it. Knowing the machine is how you know what the number means.
Real market dataDerived proxyIllustrative positionsIllustrative volThe rates and FX sleeve, 31 December 2025
The sleeve is a small book RiskAtoms Bank plc's Markets desk runs for its rates and FX hedging: US Treasuries, euro AAA government bonds, FX forwards, a sterling OIS and a few GBP/USD options, worth £293.7m. Its sizes, coupons and strikes are set for the course (illustrative); the market it is priced in, and the history its scenarios come from, are real. Daily data, 2006 to 2025, via the RiskAtoms Risk Engine, downloaded on 11 November 2026:
US Treasury zero rates at ten tenors from 3 months to 30 years, bootstrapped by the RiskAtoms Risk Engine from the US Treasury's daily par yield curve (zero rates, not the published par yields). The 30-year series starts on 9 February 2006; its first 26 days are filled by regression on the 20-year.
Euro area AAA government spot rates at the same ten tenors, from the European Central Bank's fitted (Svensson) yield curve. Every tenor was negative for part of 2011 to 2022, so moves are absolute changes in basis points.
Sterling exchange rates (GBP/USD, EUR/GBP, GBP/JPY, GBP/CHF) from the ECB's euro reference rates: one fixing a day, around 14:10 CET, so intraday extremes are missed.
No daily sterling curve. The course holds no reliable daily sterling rate curve, so the sleeve's sterling rate factor is a proxy: 0.536 × the US 2-year move + 0.509 × the euro 2-year move, the slopes of a month-end regression of the Bank's 2-year OIS rate on them (L6.2.02: Risk factor mapping and proxies).
The GBP/USD implied volatility is illustrative: the course has no implied-vol history, so it is built from a GARCH forecast of real GBP/USD volatility plus seeded noise (seed 62001). Its premium over realised volatility and its link to spot are built in, not evidence.
RiskAtoms Bank plc's Markets desk keeps its own seeded 750-day history from M3.2: Market risk, and its booked figures stand.
The book
L3.2.09: Parametric VaR for a portfolio and L3.2.10: Historical simulation in practice computed VaR for RiskAtoms Bank plc's Markets desk from its seeded history. This module needs real market history, so it adds a second book: the rates and FX sleeve, worth £293.7m on 31 December 2025. It holds 4 US Treasuries, 2 Euro AAA bonds, 4 FX forwards, 1 sterling OIS, 4 GBP/USD options. The forwards sell most of the bonds' dollars and euros forward; the options form a small GBP/USD position that is long a put spread, short a 1-month call and long a 6-month call. Its first-order risks, computed by the engine, in pounds:
US rates (PV01, all tenors)−£88,369
Euro rates (PV01)−£57,836
Sterling OIS (PV01)−£34,733
GBP/USD up 1%−£496,194
EUR/GBP up 1%+£137,898
GBP/JPY up 1%−£142,025
GBP/CHF up 1%−£140,530
GBP/USD vol up 1 point+£198,934
P&L for a 1bp rise in every rate of a curve, a 1% rise in each exchange rate as quoted (a rise in GBP/USD means sterling gained; in EUR/GBP that sterling fell), and one vol point. The sleeve loses when rates rise and when sterling strengthens.
The four stages
Every VaR engineVaR engine: the system that turns positions and market data into a VaR or ES each day, in four stages: risk factors, scenarios, revaluation and aggregation., from a spreadsheet to a bank's overnight batch, does the same four things. Click each stage to see what this one does with the sleeve.
ExploreInside the engine
Click a box to see what happens there.
↓→
↓→
↓→
1. Risk factors
The engine prices everything from 26 risk factorsRisk factor: a market quantity the book is priced from and the engine moves: a zero rate at one tenor, an exchange rate, an implied volatility.: US Treasury and euro AAA zero rates at ten verticesVertex (tenor): one fixed maturity on a curve, used as a risk factor: 3 months, 2 years, 10 years and so on. each, one sterling OIS factor, four sterling exchange rates and one GBP/USD implied vol. 24 have real daily histories; the sterling rate is a proxy and the vol is illustrative. Choosing the factors decides what risks the engine can see at all: L6.2.02: Risk factor mapping and proxies is about that choice.
Stage 3: when sensitivities stand in for prices
Revaluing in full is exact but costs a pricing call per position per scenario. Sensitivities cost one set of bumps a day. For the bonds and forwards the difference is small; for the options it is not. The chart holds vol fixed and moves GBP/USD; the dashed red line adds the 19.5-point vol rise of 24 June 2016.
Full revaluation, vol unchangedFull, vol up as on 24 Jun 2016DeltaDelta-gamma
Near zero the three agree. Far out they part: on the way down the true P&L climbs faster than the parabola as the long 1.30 put comes into the money; on the way up the parabola flattens while the true P&L keeps falling with the short 1.38 call. Over every historical day:
Approximate minus full P&L on all 4,939 days, 2006 to 2025.
Approximation
RMS error a day
Largest error
Whole sleeve, delta
£23,803
+£1,008,407
Whole sleeve, delta-gamma
£62,443
+£4,027,896
Options, delta
£26,585
+£1,445,393
Options, delta-gamma
£62,649
+£4,042,711
Options, spot-vol grid (51 × 27 nodes)
£2,128
+£15,492
Delta-gamma does worse than delta over the whole history. Most days it is the better of the two; on a handful of huge days (the referendum, and jumps in the illustrative vol) its second-order terms overshoot by millions, and an RMS error is ruled by those days. The options are short a risk reversal (long the 1.30 put, short the 1.38 call), so terms beyond the second order are large. The grid, which reprices the options in full at each node of spot and vol, is close everywhere: on their worst day, 14 Oct 2008, it gives −£649,478 against −£641,980 in full, where delta says −£428,498. Making these approximations fast and safe is L6.2.06: Making Monte Carlo VaR fast; the vol factor itself is L6.2.07: Options and volatility in VaR.
Stage 4: which days count
The same book, the same revaluation, four sets of scenarios. The stress window is the 250-day stretch since 2007 in which the sleeve's 99% VaR would have been highest:
One-day historical simulation, full revaluation, the sleeve as held on 31 December 2025.
Scenarios
Days
VaR 99%
ES 99%
ES 97.5%
Last 250 days
23 Dec 2024 to 31 Dec 2025 (250)
£1,811,177
£2,224,578
£1,844,424
Last 500 days
18 Dec 2023 to 31 Dec 2025 (500)
£2,460,369
£2,653,131
£2,248,211
2006–2025
4 Jan 2006 to 31 Dec 2025 (4,939)
£2,409,666
£2,979,376
£2,475,070
Stress window
12 Feb 2008 to 13 Feb 2009 (250)
£3,978,433
£5,083,663
£3,967,147
Twenty years of days give a VaR close to the last 500; the last 250, a calm year for this book, give much less; the stress window gives 2.2 times as much. The sleeve's worst day in twenty years comes from 14 Oct 2008: −£7,034,720. Rescaling old days to today's volatility, rather than choosing among them, is L6.2.04: Volatility-weighted and filtered historical simulation.
Real-world scenarios, risk-neutral prices
L5.1.01: Replication and risk-neutral pricing priced options by replication, under risk-neutral probabilities, and promised that risk engines would use real-world moves. Here are both at once. Inside each scenario, every position is priced risk-neutrally: forwards from interest rate differentials, options with Garman-Kohlhagen. The scenarios themselves are real-worldReal-world measure: the probabilities with which market moves actually happen, which historical changes sample. Prices use the risk-neutral measure instead.: changes that actually happened, with whatever drift the markets had. Over 2006 to 2025 the US 10-year zero rate fell on average by 0.0026bp a day, against a daily standard deviation of 5.3bp in the last 500 days. Removing each factor's average move moves the last-500 VaR from £2,460,369 to £2,467,085. At one day, drift is noise; over ten days or a year, in simulations, it is not (L6.2.05: Monte Carlo VaR with fat tails).
The desk through the same engine
The engine is not built for one book. Fed the Markets desk's eight factors and its 750 seeded days, with each position's own pricing from M3.2: Market risk, it gives a delta-normal standard deviation of £835,845, a 99% HS VaR of £2,274,006 and a worst day of −£4,978,809: the booked figures exactly, day by day (every one of the 750 P&Ls agrees). The desk's booked numbers stand; this module puts model figures beside them.
The formal version
Historical simulation, full and approximate
Positions with value V(f) in pounds, risk factors f with today's levels f0, and n past days with moves Δfs:
f̃s is scenario s applied to today's levels; Ps an exchange rate on day s; α the confidence level; k the tail count, as in L3.2.10: Historical simulation in practice; δ and Γ the vector of first derivatives and matrix of second derivatives of V at f0, by central differences with bumps hi of 1bp for rates, 0.1% for FX and 0.1 vol point (ei is the i-th unit vector). The diagonal of Γ uses the three-point formula. The engine is instantaneous: nothing ages over the day, so there is no carry or theta. Pricing follows Garman and Kohlhagen (1983) for the options, with an illustrative SABR smile (Hagan et al. 2002) at the strike, as M5.1: Option pricing and volatility built them.
Try it
The engine below runs in your browser on the last 500 days and the stress window. Start with the defaults: the 99% VaR on 500 days is £2,460,369. Switch to the stress window and delta: the VaR is £4,025,335 against £3,978,433 in full. Then switch every group off except the options and compare full revaluation (£155,215 on 500 days) with delta-gamma (£148,785). Finally, remove the drift: watch how little moves.
Try itRun the engine on the sleeve
Stage 2: scenarios (window)
Stage 3: revaluation
Stage 4: confidence level
Scenarios as they happened, or with each factor's average move removed
Stage 1: the book (switch groups of positions off and on)
500 scenarios (last 500 days). Dashed lines: VaR and ES. Red: the tail.
VaR 99%, one day£2,460,369
ES 99% (mean of the 5 worst)£2,653,131
k, the tail count5 of 500
The 5 worst scenarios, worst first: the VaR is the last row, the ES the average of all of them.
Day the moves come from
P&L (full)
Largest group (full)
4 Oct 2024
−£2,792,575
US Treasuries −£2,466,890
10 Apr 2024
−£2,736,432
US Treasuries −£2,033,272
5 Mar 2025
−£2,728,520
US Treasuries −£1,336,074
2 Feb 2024
−£2,547,761
US Treasuries −£2,898,104
13 Feb 2024
−£2,460,369
US Treasuries −£2,384,536
Worked example
One day through the four stages, then five hundred
24 June 2016, replayed on the sleeve of 31 December 2025
Step 1 of 5
1
Stage 1: today's factors
On 31 December 2025 GBP/USD is 1.34655 (the ECB fixing), the US 10-year zero rate 4.200%, the illustrative 3-month GBP/USD vol 7.81%. Twenty-three more factors complete the vector.
Regulatory lens
The Basel internal models approach sets the stages' standards without choosing the engine: “No particular type of ES model is prescribed”, and supervisors may accept historical simulation, Monte Carlo or other appropriate analytical methods (MAR33.9). Stage 3 is constrained: models “must capture the non-linear price characteristics of options positions” and the volatility surface “across both strike price and vertex” (MAR33.11). This sleeve's single parallel vol factor would not meet the second. Stage 1 is too: material exposures in major currencies need a yield curve with “a minimum of six risk factors” (MAR31.6), which the US and euro curves have and the sterling proxy does not (L6.2.02: Risk factor mapping and proxies). Stage 2 has two clocks: the current measure's data must be updated at least every three months (MAR33.8), and a stressed calibration uses the most severe 12 months, with an observation horizon going back at least to 2007 and equally weighted observations (MAR33.7). In the UK the FRTB internal models approach is due to start on 1 January 2028 (PRA, PS1/26 and CP9/26), so every capital figure built on this engine in the module is illustrative.
Spark Atoms
✦Spark AtomAn engine given away+
In October 1994 J.P. Morgan's risk management group published its internal methodology as a technical document, with a free data set of volatilities and correlations for roughly twenty markets. RiskMetrics became the benchmark for measuring financial risk, and its fourth edition (1996) is still cited for cash-flow mapping and the 0.94 decay factor (Mina and Xiao, Return to RiskMetrics, 2001).
✦Spark AtomOne formula, two interest rates+
The sleeve's options are priced with the formula Mark Garman and Steven Kohlhagen published in 1983 for foreign currency options: Black-Scholes with the foreign interest rate playing the part of a dividend yield. In every one of the 4,939 scenarios the engine calls it again, with both the dollar and the sterling rate shocked.
Key takeaways
A VaR engine has four stages: risk factors, scenarios, revaluation and aggregation. Each holds a modelling choice.
Historical simulation applies each past day's factor moves to today's levels and revalues today's positions: 4,939 scenarios for the sleeve, 2006 to 2025.
Full revaluation is exact; delta and delta-gamma are local. On the options over 20 years delta-gamma's RMS error (£62,649) is worse than delta's, a spot-vol grid's (£2,128) far better.
The window is a choice: the sleeve's 99% VaR is £1.81m on 250 days, £2.46m on 500 and £3.98m in its stress window.
Scenarios are real-world moves; prices inside them are risk-neutral. At one day, removing the drift barely changes the VaR.
The same engine reproduces the Markets desk's booked figures exactly; model figures sit beside booked ones.
Reaction
Check what stuck
Question 1 · objective 1
A historical-simulation engine has 500 scenarios. How many of the worst P&Ls does it average for the 97.5% expected shortfall?
scenarios
Question 2 · objective 2
The option sleeve has a delta of −£149,236 per 1% rise in GBP/USD and a gamma of £14,464 per 1%². What P&L does delta-gamma give for a -3% move, with vol unchanged? (In pounds.)
£
Question 3 · objective 2
Over all 4,939 historical days, delta-gamma reprices the whole sleeve with a larger RMS error (£62,443) than delta alone (£23,803). Why?
Question 4 · objective 3
The sleeve's 99% HS VaR is £1,811,177 on the last 250 days, £2,460,369 on the last 500 and £3,978,433 on its worst 250-day window. Which statement is right?
Question 5 · objective 4
Removing each factor's average daily move from the scenarios moves the last-500 VaR from £2,460,369 to £2,467,085. What does that show?
Glossary
VaR engine
VaR engine: the system that turns positions and market data into a VaR or ES each day, in four stages: risk factors, scenarios, revaluation and aggregation.
Risk factor
Risk factor: a market quantity the book is priced from and the engine moves: a zero rate at one tenor, an exchange rate, an implied volatility.
Vertex
Vertex (tenor): one fixed maturity on a curve, used as a risk factor: 3 months, 2 years, 10 years and so on.
Scenario
Scenario: one vector of moves in every risk factor, here one past day's changes applied to today's levels.
Full revaluation
Full revaluation: pricing every position again with its own pricing model at the shocked factor levels.
Delta-gamma
Delta-gamma: a second-order Taylor approximation of the P&L, δ′Δf + ½Δf′ΓΔf, from first and second derivatives at today's levels.
Pricing grid
Pricing grid: the book's P&L computed in full on a grid of moves in a few factors (here spot and vol), and interpolated between nodes.
Real-world measure
Real-world measure: the probabilities with which market moves actually happen, which historical changes sample. Prices use the risk-neutral measure instead.
Stress window
Stress window: a 12-month period of history in which the book would have lost most; regulators require a stress-calibrated measure beside the current one.
J. Mina and J. Y. Xiao, Return to RiskMetrics: The Evolution of a Standard, RiskMetrics Group, April 2001, chapters 2 and 5 (simulation, parametric methods, mapping). Intermediate.
J.P. Morgan/Reuters, RiskMetrics Technical Document, fourth edition, December 1996.
J. Mina and J. Y. Xiao, Return to RiskMetrics: The Evolution of a Standard, RiskMetrics Group, April 2001.
M. B. Garman and S. W. Kohlhagen, “Foreign currency option values”, Journal of International Money and Finance 2(3), 231–237, 1983.
Basel Committee on Banking Supervision, Basel Framework, MAR31 and MAR33 (version effective 1 January 2023).
Prudential Regulation Authority, CP9/26 Basel 3.1: Adjustments to the internal model approach for market risk, 19 June 2026 (citing PS1/26 on the 1 January 2028 start).
U.S. Department of the Treasury, daily par yield curve rates; European Central Bank, euro area AAA yield curve and euro foreign exchange reference rates; Bank of England, OIS spot curve: via the RiskAtoms Risk Engine, downloaded 11 November 2026.
Watch this lesson
Video version
Prefer to watch? The same lesson as a 14-minute video will sit here, made from this page's script.