This runs on Henry Hub and one filed US pipeline tariff, because those are public. A different hub, your own forward marks, your facility’s own rate sheet, or a contract structure this does not cover — that is the work, and it starts with a conversation. [email protected]
Fourteen years of observed natural gas forward curves through a least-squares Monte Carlo storage valuation, charged at a filed pipeline tariff rather than at assumed costs. What follows is what that settled, what it refuted, and the figures behind both.
The valuation machinery for gas storage is not controversial. Intrinsic value, rolling intrinsic, spot-optimal by Longstaff–Schwartz: the hierarchy is textbook and every implementation of it agrees to the third decimal. What decides the answer is the cost deck — and the cost deck is where published storage economics is thinnest. This project replaced every invented cost with a filed rate sheet and re-ran fourteen years of curves. Four of its own conclusions did not survive that.
| $/MMBtu of working gas | gross | + NSS charges | + the tariff’s slower reservoir |
|---|---|---|---|
| Intrinsic — the seasonal schedule | 0.4604 | 0.0211 | −0.1002 |
| Spot-optimal — with the right to re-optimise | 1.0767 | 0.6367 | 0.3067 |
| Extrinsic — what flexibility adds | 0.6162 | 0.6156 | 0.4069 |
167 twelve-month storage years, 2012-09 to 2026-08, 20,000 paths each, policy fitted on half the paths and run on the other half so the valuation cannot peek. NGPL Rate Schedule NSS as filed.
On the real tariff the seasonal schedule is worth less than nothing. Over fourteen years of actual curves the summer–winter spread did not cover the rent: intrinsic value is −$0.1002 and negative in 76% of storage years. A party that pre-sold the optimal schedule at inception would on average have locked in a loss. The spot-optimal value is +$0.3067, essentially all of it extrinsic. What pays for the space is the right to re-optimise against spot, not the shape of the curve.
That is the mirror image of the same machinery applied to FX, where the extrinsic value of a staged hedge is zero — and the comparison is the point. Storage optionality comes from seasonality and mean reversion in the underlying, not from its volatility. Run the identical code on a random walk at the same volatility and it finds nothing.
One qualification belongs immediately beside that headline, because it is the difference between a finding and an artefact. Most of the gap between the second and third columns is arithmetic. This deck is 93% capacity rent, and a capacity rent cannot change extrinsic value at all — it is a constant subtracted from the intrinsic and the spot-optimal alike, so it cancels in their difference. Read the rent as a hurdle on the total, not as evidence that the option shrugs costs off.
Scaling every forward by a constant holds the curve’s shape fixed and moves only its level. Doing that at eight price levels separates the cost terms cleanly, because they are invariant in different things:
| charge | share of intrinsic | reaches the option? |
|---|---|---|
| Fuel — retained in gas | 7.56% at every price level | no — −0.28% of it |
| Capacity rent — fixed in dollars | 491% at $0.86, 15% at $27 | cannot, by construction |
| Per-unit cash — charged on gas moved | falls like 1/price | yes, ratio 1.1 |
Fuel’s share is flat to 4.2e−17 over a thirty-two-fold range of price, because it is charged in gas and the problem is homogeneous of degree one — the schedule genuinely does not move.
On this tariff the fuel share and the rent share cross at $56/MMBtu, sixteen times where US gas trades. So the rent dominates at every price this market has seen, and the ranking of cost line items is not a property of the facility: a deck calibrated at $3 gas ranks itself wrongly at $12, and a blended cents-per-unit figure cannot express the difference at all.
Every storage figure above values a reservoir with the gas already inside it. Gas has to reach the field and leave again, and in the US that is a separate pipeline contract with its own tariff. NGPL’s firm reservation converts to $1.2496 per Dth of working gas a year — 3.1× the storage reservation itself ($0.4064), and its fuel retention is 2.45% each way against storage’s 0.92% on one leg. A storage economics model that omits transport is not missing a detail.
| same physical service | taken from the schedule | taken from the option | ratio |
|---|---|---|---|
| Firm (FTS) — a rent | $1.3669 | $0.0402 | 34× |
| Interruptible (ITS) — per unit moved | $0.2587 | $0.2342 | 1.1× |
Two contracts for one pipe, landing in completely different places. Firm transport buys a schedule; interruptible transport prices an option and then charges for exercising it. A shipper choosing between them on headline cost alone is comparing the wrong number — the totals are close and what they do to the flexibility is not.
A pipeline quotes firm reservation per Dth of maximum daily quantity, and how much daily capacity a storage user needs is set by how fast their reservoir empties. A depleted field on NSS withdraws about 1/75 of its inventory a day; a salt cavern withdraws about 1/10. The identical $7.81/Dth-MDQ/month is therefore $1.2496 a year for the field and $9.3720 for the cavern — 7.5×, same pipe, same path.
That is the arithmetic behind siting. A fast-cycling facility — the cavern turns over 8.7 times a year against the field’s 1.0 — has a transport bill that scales with the very thing that makes it worth owning, so it has to be where the haul is short. It also refutes the obvious intuition, which this project wrote down before measuring: that per-unit charges would punish a cavern while a fixed rent would not. The rent scales with speed too, through the daily-quantity denominator, and it goes from 11% to 40% of the cavern’s option.
Europe’s 85%-fill obligation is not a fee. Nothing is invoiced; the obligation simply deletes the states the option was worth something in. On the same reservoir it takes $0.1498 of the extrinsic value — 37% of it.
The larger effect is one no cost deck expresses at all: under the mandate, only 39 of 156 inception months admit a feasible schedule. Filling to 85% from empty takes seven months and a storage year opened in September has one. That is why the European storage year starts in April — a convention this model derives rather than assumes.
And on the 39 years where it is defined, the same mandate is nearly free: measured against real forward curves with monthly re-optimisation it takes $0.000 of intrinsic value and $0.003 of rolling value, against a rolling uplift of $0.038, and it binds in only 6 of the 39. Against simulated spot paths the identical constraint takes 37% of the extrinsic. Both are right, and the gap between them is the finding: what a fill mandate costs is not a property of the mandate but a measure of how much flexibility the valuation being constrained actually contained. An operator marking storage on the forward curve will price the obligation at nothing, and that conclusion is an artefact of their valuation method rather than a fact about their obligation.
Each of these was written down, measured, and refuted by the measurement. They are listed because a method that never overturns its own conclusions is not being tested.
Every valuation here carries a correctness identity that costs nothing to evaluate and fails loudly when the accounting is wrong. Rolling intrinsic re-optimises a plan the operator already holds, so each month’s increment is non-negative by construction — at worst you keep the schedule you started with. Wiring that minimum into the report rather than into a test suite caught four separate defects in a single pass, none of which announced itself any other way:
The same discipline applies to the policy itself. Standard Longstaff–Schwartz values a policy on the paths its own regressions saw, which peeks; here the policy is fitted on one half of the paths and evaluated on the other. On an FX control that bias was the entire answer — +2.9 bp against a true value of zero. And the shock process is calibrated on the roll-adjusted return rather than the price series, because front-month gas is flat over fourteen years while a rolled position lost 96%: calibrating on the price series hands the model an apparent mean reversion that the roll has already sold, and it would then value an option that does not exist.
Try the model in your browser — the same valuation, solved live, with the tariff as sliders. Or see where a storage year opened this month stands.