Quantitative risk in commodities: a methodology whitepaper

Risk numbers are only ever as good as the positions and market data behind them. This paper sets out a practical, self-contained methodology for measuring market and counterparty risk in energy and commodity portfolios, and argues a single connecting thesis: the methodology only delivers value when it runs against live positions on the same governed model as valuation. A risk number computed from a stale or separate copy of the book is not a conservative approximation of the truth; it is a different number about a different portfolio.

The paper walks the full chain of quantitative risk, from the pricing foundation that produces value and sensitivities, through the Greeks and their cross-effects, to Value at Risk in its three main forms, to Expected Shortfall and the coherent-measure critique, to scenario and stress testing, and on to counterparty credit risk and the valuation adjustments that price it. It closes with the governance that keeps a risk framework honest and the implementation argument that ties the whole methodology back to live positions and one model.

It is written to be read by risk managers, quantitative analysts, model validators, and the technologists who build risk systems, and to be complete enough that a reader could use it as a reference for how a modern commodity risk framework fits together.

The intended reader is anyone responsible for measuring, validating, or building the measurement of risk in a commodity trading operation. Risk managers will find a coherent account of how the measures relate and where each is appropriate. Quantitative analysts will find the methods stated with enough precision to implement, though the paper favours clarity of concept over exhaustive derivation. Model validators will find the assumptions and limitations of each method drawn out explicitly. And risk technologists will find the through-line that connects methodology to architecture: why the same numbers require the same live positions and the same model.

How the paper is organised

1. The pricing foundation. Where risk begins: value and sensitivities from market data.
2. Sensitivities and the Greeks. Delta, gamma, vega, theta, and the cross-effects that commodity portfolios cannot ignore.
3. Value at Risk. Historical, parametric, and Monte Carlo approaches, with their trade-offs.
4. Beyond VaR. Expected Shortfall, coherence, and what a single quantile cannot say.
5. Scenario and stress testing. Asking what specific severe moves would do.
6. Counterparty credit risk. Exposure, PFE, and the valuation adjustments.
7. Model governance and validation. Keeping the framework honest.
8. Why live positions and one model matter. The implementation thesis.

A note on notation: the paper keeps mathematics light and states results in words wherever possible, so that it is readable by risk managers as well as quants. Where a formula clarifies, it is given, but the argument never depends on the reader following a derivation.

Risk begins with valuation, and valuation begins with market data. Before any risk measure can be computed, the portfolio must be valued, and to value it the system must construct the market: the forward curves, volatility surfaces, correlations, and discount factors against which every position is priced. The quality and consistency of this market construction determines the quality of every risk number that follows, which is why the pricing foundation is the first module and not an assumed input.

Forward curves

The single most important valuation input in a commodity portfolio is the forward curve: the set of prices for delivery at each future date. Unlike a financial forward curve, a commodity curve carries the shape of physical delivery, seasonality in gas and power, the granularity of hourly settlement periods, the basis between locations. Constructing it is a craft: liquid market quotes anchor the curve at traded points, and the gaps between them are filled by interpolation and shaped to the granularity that positions require, so that an hourly power position values against an hourly curve rather than a smeared monthly average. Errors in curve construction do not stay contained; they propagate into value, into every sensitivity, and into every risk measure.

Volatility surfaces

Options require a volatility surface: implied volatilities across strike and maturity that capture the market’s view of future variability. The shape of the surface, its skew across strikes and its term structure across maturities, is itself risk-bearing, because an option’s value depends on the whole surface, not a single number. Commodity volatility surfaces carry their own peculiarities: seasonal volatility in gas and power, the pronounced skews of markets prone to spikes, and the interaction of volatility with the forward curve’s shape.

Correlations

Commodity portfolios are portfolios of spreads: a generator is long power and short gas and carbon; a storage book is long one delivery period and short another. The value and risk of such positions depend not only on individual prices but on how they move together, which is to say on correlations. A correlation matrix across commodities, locations, and tenors is therefore a first-class market input, and one of the most treacherous, because correlations are unstable, tend toward extremes in stress, and are estimated from limited data. A risk framework that treats correlation casually will misstate exactly the spread positions that dominate commodity books.

Discounting

Future cashflows must be discounted to present value, which requires a yield curve and the discount factors derived from it. In a multi-currency book, foreign cashflows must also be converted, bringing FX curves into the market construction. Discounting is less glamorous than curves and vols, but it is where a surprising amount of valuation error hides, particularly in long-dated physical contracts whose cashflows stretch years into the future.

Sensitivities as the raw material of risk

The pricing foundation produces two things: value, and the sensitivities of value to each market input. Those sensitivities, how value changes as a curve point moves, as a volatility shifts, as a correlation changes, are the raw material of every risk measure that follows. Risk is, at bottom, the study of how value responds to market moves, and the pricing engine is what tells us how value responds. This is the first appearance of the paper’s central thesis: because risk is computed from sensitivities, and sensitivities come from valuation, risk and valuation must be computed on the same positions against the same market, or they describe different portfolios.

The Greeks are the named sensitivities of value to the market, and they are the working vocabulary of a trading desk’s risk. Each answers a specific question about how the book will move, and together they give a local picture of the portfolio’s exposure that is immediate and actionable in a way that a single aggregate number is not.

Delta

Delta is the sensitivity of value to a change in the underlying price, the primary directional risk. For a linear position it is simply the size of the position; for an option it is the rate at which the option’s value tracks the underlying, varying from near zero for a deep out-of-the-money option to near one for a deep in-the-money one. Delta is the first thing a desk manages, because it is the exposure to the market simply going up or down. In a commodity book, delta is naturally expressed in delta-equivalent terms, restating options and linear positions alike in units of the underlying so that they aggregate into one comparable directional exposure.

Gamma

Gamma is the rate of change of delta as the underlying moves: the convexity of the position. It matters because a position with gamma sees its delta shift as the market moves, so a hedge that neutralises delta now will not neutralise it after a large move. Gamma is largest for options near their strike and near expiry, exactly where a small move in the underlying produces a large change in delta. A desk that ignores gamma will find its carefully delta-hedged book becoming directional precisely when the market moves most.

Vega

Vega is the sensitivity of value to implied volatility. Any book holding options has vega, and because commodity volatility is itself volatile, vega is a material risk. Vega is not a single number but a profile across the volatility surface: a book can be long vega at one maturity and short at another, exposed to the term structure of volatility as well as its level. Managing vega means managing exposure to the whole surface.

Theta and rho

Theta is the sensitivity of value to the passage of time, the time decay that erodes an option’s time value as expiry approaches. It is the price paid for holding optionality, and the income earned from selling it. Rho, the sensitivity to interest rates, is usually minor in commodity books but grows for long-dated positions whose cashflows are meaningfully discounted.

Cross-effects and bucketing

Two refinements are essential in commodity portfolios. The first is cross-gamma: the change in delta to one factor when a related factor moves. Because commodity books are built of spreads, the delta to power depends on the level of gas, and ignoring these cross-effects misstates the risk of exactly the positions that dominate. The second is bucketing: delta and the other Greeks are not single numbers but profiles across tenor buckets, showing where along the curve the risk sits. A book can be delta-neutral in aggregate while being long the front and short the back, a curve position that a single aggregate delta would hide entirely. Bucketed, cross-aware sensitivities are what make the Greeks a faithful picture of a commodity portfolio rather than a convenient oversimplification.

Value at Risk compresses the whole distribution of possible portfolio losses into a single number: the loss that will not be exceeded, over a chosen horizon, at a chosen confidence level. A one-day 99 percent VaR of a given amount says that, on 99 days out of 100, the loss over one day should not exceed that amount. VaR became the industry’s standard risk statistic because it is a single, intuitive number that aggregates across a whole portfolio, and it remains indispensable despite the well-known limitations discussed in the next module. There are three main ways to compute it, and the choice among them is one of the defining methodological decisions of a risk framework.

Historical simulation

Historical simulation applies the market moves actually observed over a past window, say the last one to two years, to the current portfolio, producing a distribution of hypothetical profit and loss from which the VaR quantile is read directly. Its great virtue is that it makes no assumption about the shape of the distribution: it inherits whatever fat tails, skew, and correlation structure the historical period contained. For commodities, whose returns are notoriously non-normal, this is a significant advantage. Its weaknesses are that it assumes the future resembles the chosen past, that it is blind to any move not in its window, and that it weights a quiet and a turbulent past equally unless deliberately adjusted.

Parametric VaR

Parametric, or variance-covariance, VaR assumes the portfolio’s returns follow a known distribution, usually normal, characterised by a covariance matrix of the risk factors. VaR is then computed analytically from the portfolio’s sensitivities and that covariance matrix. Its virtue is speed: once the covariance matrix is estimated, VaR is a quick calculation, which makes it attractive for real-time and pre-trade use. Its weakness is the normality assumption, which understates the fat tails that dominate commodity risk, and its reliance on sensitivities, which makes it a local approximation that degrades for portfolios with significant optionality and gamma.

Monte Carlo simulation

Monte Carlo VaR simulates a large number of possible future market states from a specified model of how the risk factors behave, revalues the portfolio under each, and reads the VaR from the resulting loss distribution. It is the most flexible approach: it can accommodate non-normal factor distributions, complex correlation structures, and the full non-linearity of an options book, because it fully revalues rather than approximating from sensitivities. Its cost is computational, a full revaluation across thousands of paths is heavy, which is precisely the bursty, parallelisable workload that the elastic compute of a cloud-native architecture is built to serve.

Choosing among them

No single method is correct; each trades off intuition, speed, and fidelity differently. A mature framework often uses more than one: a fast parametric or historical measure for real-time and pre-trade checks, and a full Monte Carlo measure for the thorough end-of-day and regulatory numbers, the two-speed pattern of the architecture appearing again in the risk methodology. What matters is that the choice is deliberate, its assumptions understood, and its limitations respected rather than forgotten.

Whatever the method, VaR is only as meaningful as the positions it is computed on. A VaR computed on last night’s position tells you about last night’s risk, not today’s. This is the implementation thesis again: the measure requires live positions to describe the live book.

pythonimport numpy as np

def historical_var(pnl_vectors, weights, alpha=0.99):
    """VaR from applying historical factor moves to the current book.
    pnl_vectors: array [scenarios x factors] of factor P&L; weights: sensitivities."""
    scenario_pnl = pnl_vectors @ weights        # full revaluation per historical day
    losses = -scenario_pnl
    var = np.quantile(losses, alpha)             # the VaR quantile
    es  = losses[losses >= var].mean()           # Expected Shortfall: mean of the tail
    return var, es

VaR is necessary but not sufficient, and understanding why is essential to using it responsibly. Its limitations are not reasons to discard it but reasons to complement it, and the complements, Expected Shortfall, scenario analysis, and stress testing, are as much a part of the methodology as VaR itself.

What VaR does not tell you

VaR is a quantile: it states the threshold that losses will not exceed at a given confidence, but it says nothing about how bad the loss is when that threshold is breached. Two portfolios can have the same VaR and utterly different tail risk, one losing a little beyond the threshold, the other losing catastrophically. For a commodity book exposed to price spikes, this silence about the tail is a serious deficiency, because the tail is exactly where commodities do their damage.

The coherence critique

There is a deeper, theoretical objection. A risk measure is called coherent if it satisfies a small set of reasonable properties, the most important being sub-additivity: the risk of a combined portfolio should not exceed the sum of the risks of its parts, because diversification should never increase risk. VaR is not, in general, sub-additive: there are portfolios for which combining two positions produces a VaR larger than the sum of their individual VaRs, which is a nonsensical result that can perversely penalise diversification. This is not a mere curiosity; it means VaR can mislead when used to allocate risk across a portfolio.

Expected Shortfall

Expected Shortfall, also called conditional VaR, answers the question VaR evades: given that the loss exceeds the VaR threshold, what is the average loss? By averaging over the tail rather than reading a single point on its edge, it describes the severity of bad outcomes, not just their frequency. It is also coherent, and in particular sub-additive, so it behaves sensibly under diversification and risk allocation. For these reasons regulators have increasingly favoured Expected Shortfall over VaR for capital purposes, and a modern framework computes both: VaR for continuity and intuition, Expected Shortfall for a faithful account of the tail.

Neither measure, though, tells the desk what specific event would hurt it, and for that the methodology turns from statistics to scenarios.

Statistical measures like VaR and Expected Shortfall summarise the distribution of outcomes, but they are backward-looking, inheriting their view of the future from history or from a fitted model. Scenario and stress testing complement them by asking a forward, specific question: what would a particular severe event do to this portfolio? Where VaR asks how bad is a bad day in general, stress testing asks what happens if this exact thing occurs.

Scenarios

A scenario is a defined set of simultaneous market moves applied to the portfolio to read off the profit-and-loss impact. Scenarios can be historical, replaying the market moves of a specific past episode, or hypothetical, constructed from a view about what might happen. The value of a scenario is its specificity: it does not ask about a statistical tail but about a concrete, nameable situation, a cold snap that spikes gas and power, an outage that widens a locational basis, a demand shock from an unexpected source, and it shows exactly how the book responds, position by position.

Stress testing

Stress testing applies severe but plausible shocks, often calibrated to historical crises or to regulatory prescriptions, to reveal vulnerabilities that normal-times statistics miss. A stress test deliberately reaches into the tail, applying moves far larger than a typical day, to answer the question that keeps risk managers awake: if the market moves violently, in the way it has before or in a way we can imagine, do we survive? For commodities, whose markets are prone to spikes, squeezes, and structural breaks, stress testing is not a regulatory afterthought but a core discipline.

Reverse stress testing

A particularly valuable variant works backward: instead of applying a shock and measuring the loss, reverse stress testing starts from an unacceptable loss, or the point of insolvency, and searches for the scenarios that would produce it. This reframes the question from what might we lose to what would break us, and it often surfaces vulnerabilities that forward scenarios miss, particularly combinations of moves that individually seem benign but together are fatal. For a book dominated by spread positions, whose danger lies in correlations breaking down, reverse stress testing is especially revealing.

Why scenarios need full revaluation

Scenario and stress results are only trustworthy if the portfolio is fully revalued under the scenario, not approximated from sensitivities. Sensitivities are local: they describe the portfolio’s response to small moves around the current market. A stress scenario, by construction, applies large moves, exactly where the local approximation breaks down and where gamma and higher-order effects dominate. A credible stress framework therefore fully reprices the book under each scenario, which is, once again, the heavy, parallel workload that elastic compute exists to serve. The methodology and the architecture keep pointing at each other.

Market risk asks what the portfolio might lose to price moves; counterparty credit risk asks what it might lose if a counterparty fails to perform. In derivatives and physical trading alike, a counterparty that defaults when trades are in the money represents a real loss, and measuring and pricing that risk is a discipline in its own right, closely connected to but distinct from market risk.

Exposure

The starting point is exposure: the amount that would be at risk if a counterparty defaulted. Current exposure is the present replacement cost of the trades with a counterparty, the positive mark-to-market that would be lost, after netting and collateral. But current exposure is only today’s snapshot, and a trade that is flat today may be deeply in the money next month. Potential future exposure, or PFE, addresses this by estimating, at a high confidence level, how large the exposure could become over the life of the trades as the market moves. PFE is computed by simulating market paths, revaluing the portfolio along each, and reading the exposure profile through time, again a simulation workload.

Netting and collateral

Exposure is not assessed trade by trade but across a netting set: the group of trades that legally offset one another under an enforceable master agreement. Within a netting set, positive and negative mark-to-market cancel, so the exposure is the net, not the gross, which is why the enforceability of netting is a first-order determinant of credit risk. Collateral held under a credit support annex reduces exposure further, subject to thresholds, minimum transfer amounts, and haircuts. A credit framework that ignores netting and collateral overstates exposure enormously; one that models them correctly reflects the true economic risk.

The valuation adjustments: xVA

Modern practice does not only measure counterparty risk but prices it, as an adjustment to the value of the trades themselves. The Credit Valuation Adjustment, or CVA, is the market value of the counterparty’s default risk: the expected loss from their potential default, computed over the exposure profile and their probability of default. Its mirror, the Debit Valuation Adjustment or DVA, reflects the firm’s own default risk. Beyond these sit funding, capital, and margin adjustments, FVA, KVA, MVA, collectively the xVA family, each pricing a different cost of doing the trade. Computing xVA requires simulating exposure across every counterparty over the life of the book, one of the heaviest calculations a trading firm performs, and a clear case for the scheduled, elastic batch path of the architecture.

Wrong-way risk

A subtle and dangerous refinement is wrong-way risk: the case where exposure to a counterparty rises just as their creditworthiness falls, because the two are correlated. A producer hedging with a counterparty whose fortunes move with the same commodity is a classic example: the hedge is most valuable, and the exposure largest, exactly when the counterparty is most stressed. Wrong-way risk defeats naive exposure measures that treat market and credit as independent, and a serious credit framework must model the correlation explicitly.

A risk framework is a collection of models, of curves, of volatility, of the distribution of returns, of default, and every model is a simplification that can be wrong. What keeps a framework honest is not the sophistication of its models but the governance around them: the discipline of validating them, monitoring them, and knowing their limits. This module is less mathematical than the others and no less important, because an ungoverned model is a liability dressed as a measurement.

Model validation

Model validation is the independent assessment of whether a model is fit for its purpose: whether its assumptions are reasonable, its implementation correct, its outputs accurate, and its limitations understood. Validation is independent by design, performed by people other than those who built the model, because a builder is poorly placed to see their own blind spots. It is not a one-time gate but an ongoing responsibility, revisited as markets change and as the model is used in ways its builders did not anticipate.

Backtesting

A VaR model makes a falsifiable claim: that losses will exceed the VaR only as often as the confidence level allows. Backtesting checks this claim against reality, counting how often actual losses breached the VaR and comparing that to the expected rate. Too many breaches means the model understates risk; too few may mean it overstates it. Backtesting is the empirical conscience of a VaR framework, and regulators require it precisely because it turns an abstract statistical claim into a testable one.

Reserves and adjustments

Where a model is known to be uncertain, prudent practice holds a reserve against that uncertainty, reducing recognised value until the uncertainty resolves. Valuation adjustments for bid-offer, liquidity, model risk, and credit bring a raw model value down to a prudent fair value. These adjustments are themselves governed, because they are a lever on reported profit, and a framework that lets desks set their own reserves invites exactly the optimism that reserves exist to restrain.

The audit trail and reproducibility

Governance finally depends on reproducibility: the ability to reconstruct exactly how any risk number was produced, from which positions, against which market data, using which model version and configuration. This is where risk methodology meets the architecture of the companion paper: the bitemporal, versioned, lineage-tracked governed model is what makes a risk number reproducible, and a risk framework built on a model that overwrites its history can never fully answer the question of how a past number was produced. Governance is not only a set of policies; it is a property the underlying architecture must support.

Every module of this paper has pointed, explicitly or implicitly, at a single implementation thesis: the methodology only delivers value when it runs against live positions on the same governed model as valuation. This closing analytical module states the thesis directly and draws out why it is not a technical detail but the difference between risk numbers that can be trusted and risk numbers that merely exist.

The tie-out problem

Consider what must be true for risk to tie out to P&L. The P&L a desk earns is the change in the value of its positions; the risk is the sensitivity of that same value to the market. If risk is computed on a different copy of the positions, captured at a different time, or against a different market snapshot than valuation used, then risk and P&L describe subtly different portfolios, and their attribution will never quite agree. The desk is left unable to explain why the day’s P&L differs from what the risk predicted, not because of a genuine effect but because of a bookkeeping discrepancy. On a shared model with live positions, this discrepancy cannot arise, because there is one position and one market, and risk and P&L are two views of the same thing.

Staleness as a hidden risk

A risk system fed by overnight extracts is always describing the book as it was, not as it is. A position taken this morning is invisible to it until tonight. In a calm market this is a nuisance; in a fast-moving one it is a genuine danger, because the desk is managing today’s risk with yesterday’s numbers precisely when the gap between them is largest. Live positions, updated by the event-driven architecture as trades book, remove this staleness, so that the risk on the screen is the risk in the book.

One model, many measures

The measures in this paper, sensitivities, VaR, Expected Shortfall, scenarios, PFE, xVA, are different questions asked of the same portfolio against the same market. When they are computed from one governed model, they are mutually consistent by construction: the delta that feeds parametric VaR is the same delta the desk hedges; the positions in the stress test are the positions in the book; the exposure in the CVA is the exposure in the credit report. When they are computed from separate copies, they drift apart, and the risk function spends its time reconciling its own numbers rather than managing risk. The argument for treating quantitative risk as a first-class part of an integrated platform is, at bottom, this: the methodology is only as coherent as the model it runs on.

This paper has set out a methodology for quantitative risk in commodities that runs from the pricing foundation, through the Greeks and their cross-effects, through Value at Risk and its three computational forms, to Expected Shortfall and the coherence critique, through scenario and stress testing, to counterparty credit risk and the xVA adjustments, and finally to the governance that keeps the whole framework honest. Throughout, a single thesis has recurred: the value of every measure depends on the positions and market data behind it, and those must be live and shared across valuation and risk, or the numbers describe different portfolios.

The practical implications are concrete. Use more than one VaR method, and understand what each assumes. Complement VaR with Expected Shortfall for the tail and with scenarios and stress for specificity. Model netting, collateral, and wrong-way risk properly in credit. Govern and validate every model, and hold reserves against what the models cannot capture. And insist that the whole framework runs on live positions against one governed model, because that is what turns a collection of measures into a coherent picture of risk. A modern risk function is not defined by the sophistication of any single number but by the consistency and timeliness of them all.

This paper draws on the established literature of quantitative risk management rather than on proprietary methods. Readers seeking depth on particular topics may find the following areas useful; specific regulatory figures and formulae should be checked against their current published versions.

• The foundational literature on Value at Risk and its computation, for the historical, parametric, and Monte Carlo methods of Module 3.
• The work on coherent risk measures, for the sub-additivity critique and Expected Shortfall of Module 4.
• The literature on counterparty credit risk and xVA, for the exposure, PFE, and valuation adjustments of Module 6.
• Regulatory frameworks relevant to the reader’s jurisdiction governing market risk capital, backtesting, and stress testing, for the governance discussion of Module 7.
• Companion Gravitas whitepapers: Cloud-native ETRM architecture, for the governed model and elastic compute this methodology depends on, and the reader-friendly VaR explained.

The Gravitas ETRM and CTRM data dictionary defines the risk, valuation, and credit terms used in this paper in full, and is a useful companion reference.