The Iran war began on 28 February 2026. The Strait of Hormuz closed. One-fifth of global oil and LNG supply went offline. Henry Hub jumped to $4.88 per mmbtu. TTF — Europe's gas benchmark — spiked toward €50 per MWh, the equivalent of $15.80 per mmbtu in energy terms. Every energy buyer, trader, and risk manager now faces the same question: should you buy gas futures to protect your position, or wait it out?
Most people answer that question with a forecast — a single number, a gut feeling dressed up as analysis. That approach fails precisely when it matters most. When the situation is genuinely uncertain, a point estimate does not quantify your risk. It hides it.
This article builds a different kind of answer. It uses a scenario probability tree to decompose the uncertainty into its structural components, attaches those components to a calibrated BMMRJD time series model in @RISK, and computes a probability-weighted expected value for the trade. More importantly, it builds in a Bayesian updating mechanism — so as news breaks and the situation evolves, the framework tells you when to change your mind and by how much. The result is not a prediction. It is a decision instrument: transparent about its assumptions, explicit about its uncertainty, and designed to be revised.
The structure of the uncertainty
Every complex decision under geopolitical uncertainty benefits from decomposing what you do not know into its components. In this case, the dominant uncertainty is not price volatility in the statistical sense. It is scenario uncertainty: which structural regime are we in, and how long will it persist? Duration is the key variable — because duration determines how much supply is removed from the market, for how long, and therefore what gas prices do.
I identified three conflict duration scenarios as the primary branches of the decision tree. Each carries a different supply outcome: a short resolution allows LNG flows to resume within weeks, limiting the price impact; a medium scenario sustains supply constraints long enough to drain European storage and force demand rationing; a protracted conflict embeds the supply loss structurally, removing Qatar's LNG from global markets for years and driving prices to levels not seen since the 2022 Ukraine crisis. Alongside these, a secondary discrete event — a US-Russia oil and gas agreement — could materially increase supply regardless of how the conflict on the ground evolves, putting downward pressure on prices across all three scenarios.
The conflict duration scenarios carry probabilities of 0.20 (short resolution, under one month), 0.45 (medium, one to six months), and 0.35 (protracted, over six months). The US-Russia deal probability is not unconditional. It is conditional on the conflict duration: 10% if the conflict resolves quickly, 25% under a medium scenario, and 40% under a protracted one. The longer the disruption, the greater the political and economic incentive on both sides to find an alternative supply arrangement — and the greater the supply impact if a deal materialises.
This conditional structure matters more than most analysts realise. A model that treats the deal probability as a flat 25% unconditional figure will overestimate it in the short scenario and underestimate it where it is most consequential — the protracted case. The expected value calculation changes materially depending on which approach you use.
When to change your mind: Bayesian updating in practice
Here is the most important thing I can say about the scenario probabilities: the numbers 0.20, 0.45 and 0.35 are starting estimates, not convictions. They are what an analyst could assign in April 2026, given available information. The moment new information arrives — a ceasefire announcement, a diplomatic breakthrough, a second strike on Qatari infrastructure — those numbers should move. The framework is designed for that. The question is how.
This is where Bayesian reasoning becomes a practical tool rather than a statistical concept. Bayes tells you how to update a prior belief when new evidence arrives. You do not discard your original estimate. You adjust it in proportion to how strongly the new evidence favours one scenario over the others. In plain terms: your new probability is your old probability, scaled by how likely that scenario would have been to produce the news you just saw.
The practical workflow is straightforward. Keep the scenario tree open in PrecisionTree. As news breaks, ask yourself which scenario it makes more likely and by roughly how much. Move the relevant probability. Watch the expected values cascade through the tree. If the EV remains positive and the position sizing is consistent with your updated risk profile, hold. If the EV turns marginal or negative, you now have an explicit analytical basis for exiting — not a feeling, not a headline, not a stop-loss triggered by noise.
Subjective probabilities are not a weakness of this framework. They are the honest admission that markets operate under genuine uncertainty. The discipline is not in pretending you know the probabilities precisely — it is in updating them systematically when the evidence changes.
This is also the direct answer to the most common objection to probabilistic models: that sophisticated mathematics applied to guessed inputs produces nothing but a false sense of precision. The objection is valid against models that treat their inputs as fixed facts. It does not apply here. The inputs are explicitly labelled as estimates. The updating mechanism is built into the structure. A model that tells you when to change your mind is is the only honest way to make decisions under genuine uncertainty.
Why BMMRJD — and what the data shows
Once the scenario structure is in place, the question becomes: within each scenario, which time series model best captures how natural gas prices actually behave?
I applied @RISK's Time Series Fitting tool to the Henry Hub monthly data from October 2019 to March 2026. The tool tests all available models against the historical series and ranks them by Akaike Information Criterion (AIC) — a measure that balances goodness of fit against model complexity. Lower AIC is better. The same approach can be applied to any date range — a shorter post-2021 window captures the more volatile recent regime, while the full 2009–2026 dataset reflects the complete post-fracking era. For European exposures, the identical fitting process can be run on TTF monthly data instead of Henry Hub, producing a separate set of calibrated parameters that reflects European supply dynamics, storage sensitivity, and the more extreme volatility of that market.
The result is unambiguous. BMMRJD wins by a wide margin. Its AIC of 186.25 compares to 219.60 for AR(1), 219.60 for BMMR (mean reversion without jumps), and 221.05 for ARMA(1,1). GBM and GBMJD returned N/A — they could not be fitted to this dataset, which itself tells you something about the nature of the price series.
The fitted parameters from @RISK are: long-run mean μ = 3.97/mmbtu, mean reversion speed α = 0.39, diffusion volatility σ = 0.089, jump frequency λ = 0.068 per month (approximately one jump every 15 months), and jump size standard deviation σ_j = 2.28. These are not assumed — they are calibrated from the actual historical data covering five distinct market regimes from COVID through the Iran war.
Three properties of the data explain why BMMRJD wins. First, mean reversion is real and strong. Gas prices spike and return. The 2022 Ukraine war pushed Henry Hub to $8.81, then gravity pulled it back toward $2–$3 within 18 months. The α of 0.39 quantifies this reversion force. Second, the data contains discrete jump events that continuous diffusion models cannot capture: February 2021 (Winter Storm Uri, +164% month-on-month), August 2022 (Ukraine war peak), January 2026 (Iran war onset). These are not statistical noise. They are structural shocks that require an explicit jump component. Third, BMMRJD models prices in log-space, which means simulated prices can never go negative — a minimum technical requirement for any commodity price model.
μ = 3.97 α = 0.394 σ = 0.089 λ = 0.068 μ_j = 0.088 σ_j = 2.28 Y0 = 3.31
For a scenario-driven model, these calibrated parameters serve as the baseline — the neutral, no-war equilibrium. Each conflict scenario then modifies μ upward (higher long-run mean under sustained disruption) and reduces α (slower mean reversion when supply removal is structural rather than temporary), while λ increases to reflect higher jump risk under geopolitical stress.
What the framework delivers — and what it does not
It is worth being precise about what this model is for. It is not a trading recommendation. It does not tell you to buy or sell. What it does is more valuable: it gives any decision-maker — an energy procurement manager, a CFO reviewing hedging strategy, a risk analyst preparing a board briefing — a structured way to think through a genuinely uncertain situation and understand the consequences of each possible outcome before committing to a course of action.
The framework produces three things that a single-point forecast cannot.
First, it makes the range of outcomes visible. The scenario tree has six terminal nodes. Each one carries a probability and a set of BMMRJD parameters that generate a distribution of simulated price paths, not a single price. A decision-maker looking at the output does not see "gas will average $X." They see: under these six scenarios, weighted by their probabilities, prices range from approximately $3.50 to $10.50 per mmbtu for Henry Hub, and from roughly €18 to €100 per MWh for TTF. That range is the honest answer to the question of what gas costs over the next twelve months. The deterministic alternative — a single number — is not a more confident answer. It is a hidden assumption.
Second, it separates the structural drivers. The model makes explicit what is actually driving prices in each scenario: conflict duration determines supply removal; the US-Russia deal determines whether that removal is partially offset; seasonality determines when European prices are most exposed within any scenario. A decision-maker can ask "what if the deal happens sooner?" or "what if the conflict escalates?" and see immediately which parameter to change and what the downstream effect on the price distribution is. That is not possible with a forecast. It is only possible with a model whose structure mirrors the causal logic of the market.
Third, and most importantly, it tells you when to change your position. This is the Bayesian layer. Every organisation facing energy cost exposure — whether they are procuring gas, pricing contracts, or managing a balance sheet — needs to know not just what the numbers say today, but under what conditions those numbers change enough to require a different decision. The framework answers that directly. A ceasefire announcement shifts probability mass toward the short-resolution branch. An escalation event shifts it toward the protracted branch. The decision-maker does not have to reinterpret the headlines in isolation — the model translates them into updated probabilities and updated expected outcomes automatically.
The goal is not to predict the price. The goal is to build a framework robust enough that when reality diverges from your starting assumptions — as it always does — you know exactly what to update and why.
This is the case for probabilistic modelling in volatile markets. Not because it produces the right answer, but because it produces an answer that can be wrong in a structured, auditable, correctable way — which is far more useful than an answer that is simply wrong with no mechanism for revision.
Where AI fits into this analysis — and where it does not
Artificial intelligence plays a genuine supporting role in this type of analysis, and it is worth being precise about what that role is — because it is not what most people assume.
AI did not build the scenario tree. Scenario structure requires domain knowledge: understanding why conflict duration is the right primary branching variable, why the US-Russia deal probability should be conditional rather than marginal, and why TTF and Henry Hub behave as structurally different markets. That judgment belongs to the analyst with energy market expertise.
What AI contributes is the surrounding infrastructure. Large language models can scan real-time news feeds, IEA reports, shipping insurance data, and tanker tracking information to detect signals of regime change — a Force Majeure notice, rising insurance premiums, storage drawdown rates — before they appear in price data. An AI agent running daily can monitor the Henry Hub price against the fitted BMMRJD distribution and flag when the price has moved outside the 5%–95% confidence band for two consecutive months, prompting the analyst to review the scenario probabilities. After each @RISK simulation run, an LLM can translate the P10/P50/P90 output into board-level language — something that currently takes an analyst several hours per reporting cycle. And AI can stress-test the model's assumptions systematically: "under what conditions does the US-Russia deal probability rise above 50%?" is a structured question that an LLM can explore across multiple geopolitical scenarios in minutes.
The boundary that does not move: AI does not run the simulation. The computation stays in @RISK. The accountability stays with the practitioner. Adding AI to this workflow does not reduce the analyst's responsibility for the model's assumptions — it reduces the time required to monitor them and communicate their implications. That is what an accelerator does.
The caveats that belong in every model
George Box, the British statistician, wrote that all models are wrong but some are useful. This model is wrong. Every parameter is an estimate. A sophisticated reader might reasonably ask whether complex mathematics applied to expert-guessed probabilities produces anything beyond a false sense of precision. That is a legitimate challenge, and it deserves a direct answer.
The answer is the Bayesian updating section above. A model whose inputs are treated as fixed facts deserves that criticism. A model built to be revised — where every probability is explicitly labelled a starting estimate, where three worked examples show how new information changes the output, and where the expected value shifts transparently in response to evidence — is something different. It is not a prediction. It is a decision instrument. The scenario probabilities are expert judgments, the BMMRJD parameters are calibrated on historical data blending COVID, Ukraine, and Iran shocks, and the conditional deal probabilities are informed estimates. A genuinely novel event falls outside the distribution entirely. None of that disqualifies the model. It is what makes it honest.
My position on this is direct: a wrong model with explicit, auditable assumptions is far more useful than a single-point deterministic forecast, and both are vastly more useful than doing nothing. A deterministic model that says "Henry Hub will average $6.00 in 2026" cannot tell you the probability you are wrong, the magnitude of loss if you are wrong, or what assumption needs to change if reality diverges. A probabilistic model with a scenario tree tells you all three — and gives you a starting point for updating the model when reality surprises you.
Additionally, several factors that a more complete model would incorporate are absent here, and I am naming them rather than hiding them.
Weather and temperature. The BMMRJD model captures historical seasonal patterns through the fitted parameters, but extreme weather events — a severe Northern Hemisphere winter, an unexpectedly warm Q4, a polar vortex equivalent to February 2021 — can move Henry Hub by 50–80% in a single month independently of geopolitical dynamics. A model of gas prices over 12–24 months should incorporate a stochastic temperature variable, particularly for European TTF where winter demand sensitivity is acute and storage levels entering winter 2025–2026 were already critically low at 30% capacity.
Storage levels. Gas storage is the primary buffer between supply disruption and price spike. When European storage is 90% full heading into winter, the market can absorb a supply shock for several months before prices react sharply. When storage is at 30% — as it was in late 2025 — the same supply shock transmits to prices almost immediately. The scenario probabilities in this model do not explicitly account for storage trajectories, which means the model may underestimate price volatility in the early months of a protracted conflict scenario.
LNG infrastructure constraints. The model assumes that supply can eventually restore through LNG rerouting once the Hormuz closure ends. But the Ras Laffan damage in Qatar is estimated at a 3–5 year repair timeline, independent of when hostilities cease. This means the structural supply removal from Qatar is not captured in any of the three conflict duration scenarios — it persists regardless of resolution timing. A fully specified model would include a fourth uncertainty branch for Qatari infrastructure recovery, separate from the conflict duration variable.
Subjective probabilities. The scenario probabilities (0.20/0.45/0.35) and the conditional deal probabilities (0.10/0.25/0.40) are expert judgments made in April 2026. They are not derived from prediction markets, historical base rates for comparable conflicts, or structured probabilistic elicitation. Two different analysts with comparable expertise and access to the same information could reasonably assign probabilities that differ by 10 percentage points in either direction. Before using this model for any real financial decision, run a formal sensitivity analysis on the scenario probabilities across their plausible range, and ideally conduct a structured expert elicitation with at least two independent domain experts.
Currency risk for TTF. The TTF analysis converts €/MWh to $/mmbtu at an indicative EUR/USD rate of 1.08. The actual rate has ranged from 0.96 to 1.22 over the past five years. A 10% move in the exchange rate changes the converted TTF payoff by approximately 10%. A European company hedging in euros faces a different analysis than a US fund trading a dollar-denominated synthetic TTF exposure.
A European industrial gas consumer — April 2026
A mid-sized European chemical producer consumes approximately 2 million MWh of gas annually, procured on rolling quarterly forward contracts. Their procurement team is reviewing Q3 and Q4 2026 positions as TTF sits at €50/MWh.
Without a probabilistic framework, the procurement manager faces a binary judgment: lock in at €50 now, or wait for prices to fall. History suggests waiting — but history did not include the Ras Laffan infrastructure damage with a five-year repair timeline.
Applying the scenario tree, the analyst computes an expected TTF price of approximately €60/MWh over a 12-month horizon under base scenario probabilities. The P10 outcome (short resolution) implies €28/MWh — a €22/MWh saving relative to locking in today. The P90 outcome (protracted, no deal) implies €100/MWh — a €50/MWh additional cost. The asymmetry is stark: the maximum downside from locking in today is €22 saved. The maximum upside from locking in today is €50 protected against.
The model does not make the decision. But it reframes it: the question is no longer "will prices fall?" The question is "at what probability of short resolution does waiting become the higher-expected-value choice?" The break-even requires p(short) > 55%. Given the Ras Laffan structural damage, that belief is very difficult to sustain. The procurement team locks in 60% of Q3–Q4 volume at €50 and leaves 40% on the market with explicit review triggers — a structured, documented position rather than a reactive guess.
The bottom line
I built this model because doing nothing — or using a single-point forecast — is worse than using an imperfect probabilistic framework. The scenario tree makes assumptions explicit. The BMMRJD calibration replaces guesswork with data. The expected value calculation shows that the buy decision is robust across a wide range of scenario probabilities. The caveats make clear where the model can break.
That is the standard I hold models to: not correctness, which is impossible, but usefulness relative to the alternatives. A model that shows you a positive expected value under 80% of plausible probability assignments, that tells you exactly which assumption to update if the short scenario materialises, and that scales the position sizing to the volatility of each scenario — that model is useful. It is better than a forecast. It is far better than nothing.
Build the scenario tree. Fit the time series to the data. Run the simulation. Act on the distribution — not on a number.