The Rhine Is Low. The Output Effect Is Still Missing.

The Rhine is low enough to make the causal story feel obvious. On 22 August 2026, the partial-day mean at Kaub was 38 centimeters. During the month, the daily mean had fallen as low as 6 centimeters. Both readings sit below the official 77-centimeter equivalent water level used as a navigation benchmark.

The economic inference appears to follow: barges carry less, freight becomes expensive, factories wait for inputs, and production falls. Yet an appealing mechanism is not an estimate.

I built a fast monthly prototype to test the first link from river conditions to industrial output. It combines 15-minute Kaub water levels with seasonally adjusted production for nine German industry families and a pre-shock freight-exposure proxy. The result is less dramatic than the river. Before 2020, the estimated interaction between low water and sector exposure is almost exactly zero: +0.06 percentage points, with a 95% interval from -0.11 to +0.23.

That is not evidence that the Rhine does not matter. It is evidence that national sector data are too aggregated, and the exposure proxy too coarse, to recover a stable production effect from water levels alone. The null result changes the next question.

For the Rhine-Neckar region, this is not a distant macroeconomic curiosity. Ludwigshafen is built around the Rhine as both industrial artery and constraint. BASF reported that roughly 40% of incoming volumes at the site normally arrived by ship; after the 2018 episode, it invested in early warning, low-water-capable vessels, more flexible loading stations, and purpose-built ships. Two of those optimized vessels entered service in 2023. This is the first economic counterpoint to the near-zero national coefficient: a shock can remain costly while adaptation prevents it from appearing as lost aggregate output.

The same mechanism is visible farther downstream. In a public post during the current low-water episode, Covestro CEO Markus Steilemann wrote that the recent rise in water was welcome but no all-clear. At Covestro's Lower Rhine sites, more than 30% of produced materials and roughly 75% of raw materials move via the Rhine. The company can deploy the low-water vessels Courage and Curiosity, coordinate a cross-functional Rhine Taskforce, increase sailing frequency and inventories, use truck or rail where possible, and shift production between sites. Those measures absorb part of the constraint. They do not replace the river.

Together, BASF and Covestro reveal two observations the national panel lacks: plant-specific exposure and endogenous adaptation. BASF shows how the 2018 shock changed capital and logistics. Steilemann shows that the same physical constraint still reaches operations in 2026 despite those defenses. This is not a contradiction to the null coefficient. It is the missing middle.

I also want to acknowledge Arne Jonas Warnke, who has worked on this topic at BASF. The harder analytical question is no longer whether low water creates pressure. It is where that pressure appears after firms have learned to absorb it.

The shock is real, and nonlinear

Kaub is the relevant bottleneck because it sits on the Middle Rhine, where shallow water constrains vessel draft. The gauge level is not the navigable depth, and the 77-centimeter equivalent water level is not a switch that closes the river. It is a reference level. Every centimeter below it reduces the loading options available to operators, but the exact capacity response depends on the vessel, cargo, and channel.

For that reason, the prototype does not treat one day at 76 centimeters like one day at 20 centimeters. It calculates the proportional shortfall below 77 centimeters, squares it, and adds the daily values within each month. The measure is zero above the benchmark and rises disproportionately as the river falls.

The chronology is unmistakable. November 2011 recorded 20 days below 77 centimeters. November 2015 recorded another 20. In November 2018, all 30 observed days were below the benchmark and the nonlinear stress index reached its historical maximum in this sample. August 2022 produced 21 low-water days. July and August 2026 have now added another severe episode, with 19 low-water days in July and 22 through 22 August.

Hydrology is not the weak part of the design. Exposure is.

Exposure can be measured, but only approximately

The identification strategy needs industries that depend differently on inland shipping. Eurostat reports German inland-waterway tonnage by goods category, while structural business statistics report turnover by industry. I map nine NST 2007 goods families to their corresponding two-digit NACE industries, average freight over 2017-2019, and divide by 2019 turnover.

The resulting ranking is economically plausible. Refining sits at the top with about 355 tonnes of inland-waterway freight per EUR 1 million of turnover. Chemicals, pharmaceuticals, and plastics follow at 63 tonnes, almost level with non-metallic minerals at 62. Machinery and electronics sit near the bottom at 1.3 tonnes; transport equipment records 2.2.

This is useful variation. It is not a direct measure of Rhine dependence. German freight statistics cover the inland-waterway system, not only Kaub. Transported goods do not map perfectly to the industry that ultimately uses them. Turnover is a denominator, not physical output. A factory beside a Rhine port and a factory supplied by rail receive the same national sector score.

The proxy therefore captures bulk-logistics intensity, not plant-level exposure. That distinction becomes decisive in the regression.

The attractive coefficient is approximately zero

The baseline model compares annual production growth across sectors within the same month. Sector fixed effects absorb persistent differences between industries. Month fixed effects absorb shocks common to German manufacturing. The remaining question is whether output weakens more in a low-water month when an industry has greater inland-waterway freight intensity.

The primary sample runs from January 2011 through December 2019. This avoids both the pandemic and the 2022 energy shock. The water variable averages current and prior-month nonlinear stress because an input interruption need not appear in production on the day a barge sheds cargo.

The coefficient is +0.06 percentage points for a one-standard-deviation increase in both water stress and exposure. Its confidence interval spans zero comfortably, and the interaction explains almost none of the within-panel variation. A contemporaneous specification turns slightly negative at -0.05. A simple count of days below 77 centimeters turns positive at +0.12. None is precise.

At the November 2018 peak, the primary estimate implies that a high-exposure sector would outperform a low-exposure sector by 1.08 percentage points. The 95% interval runs from a 1.92-point shortfall to a 4.08-point advantage. A model that permits both signs and explains almost no variation has not measured the effect we came to find.

This is the central result. The fast prototype fails its own first test.

2022 looks different because 2022 was different

Event windows make the contrast visible. Around the 2011, 2015, and 2018 low-water peaks, the average output path of the three most exposed industry families does not fall systematically behind the three least exposed families. At the event month, the high-exposure group is 0.53 percentage points behind the low-exposure group relative to the common six-to-four-month pre-event baseline.

The 2022 path is different. At the August low-water peak, the high-exposure group is 9.38 points behind. Three months later, the gap reaches 15.01 points; after six months, 20.73 points.

The 2026 line is different in another way: it ends before the event. Figure 4 shows February through June, with the last observation two months before the August hydrological peak. The revised visual now separates the completed historical episodes from the unfinished 2026 window more explicitly. It establishes the pre-shock position of the two groups. It cannot yet show an effect.

It would be tempting to call this the Rhine effect. It would also be wrong. The same period contains the European gas shock, extraordinary electricity prices, supply-chain disruption, and sector-specific shutdowns. Chemicals, refining, minerals, and metals are exposed to both bulk logistics and energy. Extending the regression through 2022 turns the coefficient negative, to -0.14, but the interval widens from -0.57 to +0.29.

The sign change is informative. It shows that 2022 carries the result. It does not tell us which constraint did the carrying.

A mechanism needs intermediate evidence

The proposed chain has five links: low water reduces barge capacity; lower capacity raises freight costs; expensive or delayed transport creates input shortages; shortages constrain production; scarcity may then reach producer prices.

This prototype observes the first and last parts of that chain. It measures water levels, assigns a sector exposure proxy, and observes national production. It does not yet observe monthly barge load factors, route-specific freight rates, plant inventories, input delivery failures, or order backlogs for all nine industries.

That missing middle matters. Firms can respond before production falls. They can charter more vessels, carry smaller loads, switch to rail, hold larger inventories, change suppliers, or accept higher transport costs. If those margins absorb the shock, low water will appear in freight prices and working capital rather than output. A production-only regression would then conclude that nothing happened precisely because adaptation worked.

The correct next version should therefore estimate the mechanism, not merely repeat the outcome regression with more decimal places.

2026 is a test with a missing dependent variable

The current episode offers a genuine out-of-sample test, but not yet a completed one. German monthly production in the dataset ends in June 2026. Figure 6 adds a fixed-weight output proxy for the nine industry families, using 2019 turnover weights and Eurostat's native 2021=100 scale. The index rises from 92.76 in January to 94.72 in June.

Through June, the output path is a baseline, not an impact estimate. The severe decline at Kaub begins in July and intensifies in August, after the production series ends. In other words, the shock has arrived and the relevant outcome has not.

That release lag is analytically inconvenient and methodologically useful. A model estimated on historical data can be frozen before the relevant production observations appear. When July and August output are released, the test will be clean in one narrow sense: the prediction cannot be adjusted after seeing the result.

The current model predicts no reliable national exposure penalty. Given its wide historical interval, that forecast should be read as a benchmark to beat, not as reassurance. A visible decline concentrated in refining, chemicals, minerals, and metals would reject the prototype and support a richer exposure design. No differential decline would shift attention toward freight costs, inventories, and substitution.

What this prototype changes

The original hypothesis remains plausible: low Rhine levels can act as a nonlinear supply shock concentrated in bulk-dependent industry. The fast national panel does not establish it.

Three improvements now have clear value. First, replace national goods-sector matching with port and route-level freight exposure. Second, connect establishments to Rhine-dependent ports and alternative rail capacity. Third, separate logistics exposure from energy intensity, especially in 2022. Those additions create sector-by-region-by-time variation and reveal whether the shock reaches production or is absorbed earlier in the chain.

There is also a practical lesson for short data projects. A good weekly analysis need not confirm its opening intuition. The useful product is a reproducible test that narrows the next question. Here, the river shock is visible, the proposed national output coefficient is not, and the difference is the result.

Methods and caveats

Water levels are unvalidated raw measurements from PEGELONLINE at Kaub, aggregated from 15-minute observations to daily and monthly frequency. The production index is Eurostat's German monthly series, seasonally and calendar adjusted, with 2021 equal to 100. Two-digit NACE industries are combined using fixed 2019 turnover weights.

Exposure is the average 2017-2019 German inland-waterway tonnage of the matched NST 2007 goods family divided by mapped 2019 sector turnover. The regression uses annual log production growth, sector and month fixed effects, and Newey-West standard errors with six monthly lags on period-level score sums.

The design does not yet control for sector-specific energy intensity, foreign demand, temperature, maintenance shutdowns, inventories, or regional production shares. The nine-industry exposure proxy is estimated, not observed at establishment level. Results should therefore be read as a prototype and a falsification exercise, not as a final causal estimate.

Sources

• PEGELONLINE Kaub station and historical raw-data service, General Directorate for Waterways and Shipping, DL-DE Zero 2.0. Retrieved 22 August 2026.
• BASF SE Financial Statements 2019 and BASF Report 2023, resilience measures and low-water-optimized vessels at Ludwigshafen.
• Markus Steilemann: Covestro's operational response to low water on the Rhine.
• Covestro: Markus Steilemann's current term as Chief Executive Officer.
• Eurostat STS_INPR_M, monthly production in industry. Data through June 2026.
• Eurostat IWW_GO_ATYGO, inland-waterway transport by type of good, 2017-2019 exposure window.
• Eurostat SBS_NA_IND_R2, industry turnover in 2019.
• Derived data and publication manifest.