When Winter Becomes Knowable
European gas storage cannot deliver its nameplate withdrawal rate at low inventory. We built the daily index that measures the gap, backfilled it eleven winters, and asked a question the index can answer about itself: on what date does it start telling you how the winter ends?
Fill is a volume. Sendout is a rate. The gap between them is not published anywhere, so we compute it.
A storage site cannot deliver its declared maximum withdrawal rate when it is nearly empty. Deliverability is pressure-driven and decays as working gas drains, which means a headline fill percentage overstates what the system can actually send out on a cold day, and overstates it by more the emptier the system gets.
The index applies ENTSOG and GSE deliverability curves to each gas day's fill and declared technical maximum, for Germany, France, Italy, the Netherlands and Austria. It reads as a percentage: available withdrawal capability against declared technical maximum. On 8 September 2026 it stood at 92.4%.
What follows is not a market call. It is the instrument being tested against its own record, including where it fails.
What the curves are, precisely
ENTSOG's Annex A publishes withdrawal curves per storage area, and its own methodology note states each row is already "a weighted average of the facilities (salt caverns, aquifers or depleted fields) of each area". We average those areas unweighted within each country, because the annex publishes working gas volume per country but not per area.
That averaging has a consequence worth stating plainly. One identical curve appears sixteen times across eight countries: it is the generic curve ENTSOG applies where no area-specific curve exists. Ten of the eighteen areas across our five countries carry it, including seven of Germany's eight and Italy's only area. France is the clean case, with five France-specific curves, and its aquifer row falls to 0.48 of nameplate at 40% fill, the earliest decay of the five.
So the German blend is softer evidence than a single number suggests. On Germany's own area curve the shed between 40% and 30% fill is about a tenth. Blended across all eight German areas it reads as a fifth. Both are in the data; only the second is in the index.
Half of what the index will ever know about March arrives in fifteen days.
For each of eleven winters we took the index reading on a set of autumn dates and compared it with that winter's eventual March trough. The measure is r²: the share of the variation in March outcomes that the reading on that date explains. Zero means the date tells you nothing. One would mean it tells you everything.
Both lines use the same eleven-winter record. The cyan line drops 2022/23 and 2023/24, which carry AGSI backfill holes of about 72 days each and are covered by 86 and 87 days against roughly 151 for every other winter. Both were comfortable winters with high December readings, so including them adds two easy points and flatters the fit. The nine-winter line is the honest one.
The finding, stated carefully
A September reading explains almost nothing about March. By 10 December the index explains about 43% of the variation, and by late December about 50%. The gain between 25 November and 10 December is +0.235 of r², about half of the 0.456 the index gains between 15 September and 28 December. A desk that watches this series has a datable reason to start paying attention, and it is not the start of the heating season.
Every winter in the record starts inside a four-point band. They finish inside a thirty-six-point one.
The most tempting claim about a low start is that it implies a hard finish. The record does not support it. On 15 September the eleven winters span 93.7% to 97.8%. Their March troughs span 52.9% to 89.4%. There is a faint positive relationship and it is not worth trading.
r² = 0.13 across eleven winters, and 0.05 across the nine with full coverage.
Read the two extremes
The winter with the lowest September reading in the record, 2021/22 at 93.7, troughed at 63.2. The winter with the highest, 2019/20 at 97.8, troughed at 87.3. That looks like a relationship until you add 2016/17, which started at 97.5 and troughed at 66.3, and 2020/21, which started at 97.7 and troughed at 66.0.
2017/18, the season that set the record low of 52.9%, started at 96.2, squarely mid-pack.
This winter's 92.4 on 8 September is the lowest September reading in the series, by 1.2 points. On the record above, that fact carries far less than it appears to.
Three winters read within a tenth of a point of each other in late December, and finished eighteen points apart.
The obvious next question is whether the December reading gives a threshold: a level below which the winter is in trouble. We tested it, and it does not.
Highest December reading among winters that troughed below 75: 94.6. Lowest among those that did not: 94.5. No threshold separates them.
The limitation, not buried
2020/21 read 94.6 on 28 December and troughed at 66.0. 2018/19 read 94.6 and troughed at 83.9. Same instrument, same date, same value, and outcomes seventeen points apart. The index narrows the distribution of March outcomes materially between late November and mid December. It does not produce a signal, and a desk that treats a December reading as a trigger will be wrong roughly as often as it is right.
March is flat in the comfortable winters and moves hard in the bad ones.
A fair objection to any late-season measure is that the tail is flat, so the exact date of a minimum is noise. That holds here for the easy winters and fails for the ones that matter.
In 2017/18 the index stood at 65.4 on 1 March and fell to 52.9 by 30 March, a drop of 12.6 points inside the month. In 2018/19 it moved 0.91 points across the whole of March. The month is only quiet when the winter was never a problem.
All eleven troughs fall between 19 March and 1 April. That is close to arithmetic rather than discovery, because the index is a monotone function of fill and European storage bottoms at the end of the withdrawal season by definition. It is worth stating because it tells you where to look, not because it is surprising.
Eleven winters, every reading behind every number above.
| Winter | 15 Sep | 1 Nov | 25 Nov | 28 Dec | 31 Jan | March low | On | Days covered |
|---|---|---|---|---|---|---|---|---|
| 2015/16 | 95.9 | 96.7 | 96.3 | 94.5 | 90.6 | 80.1 | 1 Apr 2016 | 152 |
| 2016/17 | 97.5 | 97.7 | 96.8 | 93.0 | 82.6 | 66.3 | 24 Mar 2017 | 151 |
| 2017/18 | 96.2 | 97.0 | 96.4 | 92.2 | 87.2 | 52.9 | 30 Mar 2018 | 150 |
| 2018/19 | 95.8 | 97.4 | 97.0 | 94.6 | 89.8 | 83.9 | 24 Mar 2019 | 151 |
| 2019/20 | 97.8 | 97.8 | 97.7 | 96.9 | 93.7 | 87.3 | 31 Mar 2020 | 152 |
| 2020/21 | 97.7 | 97.6 | 97.6 | 94.6 | 87.2 | 66.0 | 25 Mar 2021 | 150 |
| 2021/22 | 93.7 | 95.0 | 93.7 | 88.7 | 78.0 | 63.2 | 19 Mar 2022 | 151 |
| 2022/23 Partial | 97.2 | 97.7 | 97.7 | 96.7 | — | 87.7 | 29 Mar 2023 | 86 |
| 2023/24 Partial | 97.7 | 97.7 | 97.7 | 96.9 | — | 89.4 | 31 Mar 2024 | 87 |
| 2024/25 | 97.7 | 97.7 | 97.0 | 94.2 | 87.9 | 70.6 | 20 Mar 2025 | 151 |
| 2025/26 | 95.9 | 96.0 | 95.0 | 91.9 | 76.1 | 64.4 | 31 Mar 2026 | 151 |
March low is the minimum daily reading between 1 March and 15 April. Days covered counts the gas days present in our record between 1 November and 31 March. The two partial winters have AGSI backfill holes of roughly 72 days each, which is why they are excluded from the headline fit and greyed here. Their true minima could sit lower than shown, and their trough dates cannot be tested, because their records begin in March.
Where this instrument is weaker than it looks.
- Eleven observations. Nine of them clean. Every statistic here rests on a sample a single unusual winter would move.
- One curve vintage across eleven years. The whole history is scored on the Summer 2026 edition. Comparing that edition against Summer 2025 on the same five countries, the blended curve moves by up to 7.2 points at low fill and 5.2 points at 50% fill. Our own two alert thresholds sit 5 points apart. Cross-season comparisons are therefore softer than three significant figures imply, and a record low of 52.9% means 52.9% as scored today.
- Ten of eighteen areas carry a generic curve. Detailed in section 01. Italy contributes no Italy-specific curve at all.
- Coverage is five countries, about 80% of EU technical capacity. Not an EU total, and the missing fifth is not randomly distributed.
- Availability is not scarcity. OIES records that EU daily withdrawal has never exceeded roughly 60% of daily withdrawal capacity since 2017, and rarely 50%. This index measures what the system could deliver, not what it needed to. A low reading is a narrowing margin, not a shortage.
- The index is a monotone transform of fill. A December reading predicting a March one is autocorrelation in a slowly draining series, and is expected. The finding here is not that the correlation exists; it is where along the calendar it appears.
- Strategic and non-market volumes are in the fill figure. Italy reserves roughly a quarter of working volume to ministerial call and Austria holds a floor near 21%, so at very low headline fill the curve is being read at a point the market cannot actually reach.
Where the instrument runs
The index updates every gas day on the Voltstack platform, with the backfilled record behind it and per-country curves charted against the current position. Two thresholds sit on it as live alerts, at 80% and 75%, and the record above is what puts them in context rather than the other way round.
This winter's window opens on 25 November. The instrument will not tell you how the winter ends, and this note has been specific about why not. What it will tell you, earlier than the fill number will, is how much of the answer is still outstanding.