# Expected without the war: the counterfactual models

Every trade number on the map is a change against what a country would have bought from or sold to Russia and Ukraine if there had been no war. That level cannot be observed, so it is predicted, with models tested on past years. The short answer: for most flows the best prediction is 2021 held flat; only sales to Russia are predicted better by a model that uses the country's trade with everyone else.

In short: without the war, each country would have kept buying what it bought in 2021, at no-war prices (small and mid-size buyers: see below). Where something the buyer did not choose cut what Russia could deliver (Ukraine's strikes, Russia's sudden export bans, a pipeline hit in the war), the baseline is lowered: a buyer that wanted Russian fuel but could not get it did not choose to buy less. Buyers that had stopped buying by their own choice are unaffected.

One rule for what counts. A country's figure reflects its own choices and the foreseeable answers to them. Things forced on it that it did not choose — Ukraine's strikes on Russian refineries, Russia's sudden export bans — count neither for it nor against it.

- Counted (choices, or foreseeable answers to them): embargoes, sanctions and the price cap; switching to other suppliers; Russia's 2022 gas cut-off, an answer to the buyers' own sanctions; Russia's counter-sanctions and import bans on "unfriendly" countries; the end of gas transit through Ukraine on 1 January 2025 (Ukraine's decision not to renew its contract, announced more than a year ahead, so buyers could switch; labelled, only for gas still flowing in 2024); and the war-time prices paid by buyers who kept buying (the windfall), with no credit for lower prices that others' sanctions caused.
- Not counted (forced): fuel Russia could not deliver because of Ukraine's strikes, its 2023 ban on gasoline and diesel exports, and the 2026 halt of the Druzhba pipeline to Hungary and Slovakia after a Russian strike.
- The trade-off. For willing buyers hit by a forced loss, the figure shows what they chose to pay, not what they managed to buy: about EUR 13bn more in total than actually reached Russia (about EUR 4bn of it Russia's earnings).

Details for developers: [the full model tables, formulas and research files](https://github.com/intelligent-cream/war-balance/blob/main/docs/reference/method/counterfactual.md).

## 1. What "expected without the war" means

For each flow and year from 2022, the map compares the expected volume (the pre-war level, usually 2021, moved by the model's prediction) with the actual volume (the year's money divided by its price level). The difference is valued at that year's prices and summed. A good counterfactual must not use anything the war caused, such as Russia's own trade after 2022 or realised 2022 energy prices.

## 2. Four generations of models

The project tried four generations: a flat 2019–2021 average (biased low, because trade usually grows), each country's own damped trend (too narrow, and fitted on the rebound after the 2014 crash), nine statistical methods (only "2021 held flat" did well in every test), and finally 16 models that use the country's trade with others and its GDP.

## 3. The 16 models of generation 4

They range from "2021 held flat" through share and GDP models to machine learning. Models that use Russia's own trade after 2022 were not allowed: that trade is the war's outcome. The full list is in the developer reference.

## 4. The three back-tests, and why 2013 matters

Each model predicted years it had not seen: 2008–2012 from 2007, 2013–2017 from 2012, and 2017–2021 from 2016. The 2013 test is the hard one: in 2014–2016 the shock came from Russia itself (the oil price halved, the rouble lost half its value), which partner-side data cannot see. It warns how wrong a prediction can be when Russia's own economy moves.

## 5. Results and the decision rule

The rule, fixed before the results: a model replaces "2021 held flat" only if it is at least 15% more accurate in every test (or on average, worse in none). Only the model for sales to Russia passed (about 18% better). A linear model and a random forest looked better for trade with Ukraine on two tests, but not with more tests.

## 6. Which model is used for which flow

| flow | expected level | why |
| --- | --- | --- |
| Sales to Russia | a model on the country's exports to everyone else | clearly better in all three tests |
| Purchases from Russia | 2021 held flat | depends on Russian supply, which partner data cannot see |
| Trade with Ukraine | 2021 held flat | no model passed with more tests |
| Services | 2021 flat plus the travel recovery from COVID | 2021 was a COVID year for travel |
| Re-export hubs and Belarus | 2021 held flat | their trade was reshaped by goods passing to Russia |
| Company taxes, shipping, laundered oil | pre-war trends | not covered by the back-tests |

Commodity prices in the models follow the IMF forecast of October 2021, the last before the war, not realised prices.

## 7. How the prediction's error enters the ranges

The expected level is never treated as exact. Its range comes from the models' own errors in the back-tests: a 90% band, wider for small flows and longer horizons (for a large flow over 2022–2026, roughly ×0.6 to ×1.6). In the 2013 test, when the shock came from Russia, these bands covered only 77–79% of outcomes instead of 90%; this is listed under [limitations](https://warbalance.org/how-it-works/).

### Belarus

Belarus's imports from Russia (about EUR 18bn a year) are steadier than the average partner's, so they get a narrower band of their own (×0.72–1.44). Even so the range is wide enough to cross zero: its net is −3.83% of GDP, likely −13.53% to +7.11%. Belarus is a party to the war and not ranked.

### Trade with Ukraine

Tested on Ukraine's own 2004–2021 trade and eight similar partners, the models' gains roughly halved and they lost in several years, so both purchases from and sales to Ukraine hold 2021 flat.

### Russia's neighbours and Moldova

Each of Russia's neighbours was tested on its own history. Mongolia, Yemen, Tajikistan and Moldova get their own, narrower bands where those still covered the test years. Moldova's gas for Transnistria is not counted: Transnistria never paid for it.

### Imports from Russia

Five ways to narrow the range for "paid to Russia" were tested; only a band re-measured year by year passed (large flows ×0.57–1.62).

### Small flows start from the 2019 to 2021 median

A small flow from Russia swings from year to year (one cargo more or less). Where purchases were under EUR 0.1bn a year, the no-war level starts from the median of 2019, 2020 and 2021, so one unusual year cannot set it. Since round 47, flows of EUR 0.1–1bn start from 2021 again: in the back-tests that was 13% more accurate for them. It was applied at the owner's request, though it is below the 15% bar for a new model.

The back-tests also showed that small buyers' purchases grew faster than a flat path. Their middle value is therefore raised by the measured median error, ×1.57 over 2022–2026. Their range stays where it was, by design: only the middle moves. Exceptions keep 2021 where 2019–2020 do not describe the country (Brunei's refinery opened in late 2019, New Zealand's closed in 2022; Syria's pre-war data are only Russia's own reports).

### Pipeline fuel buyers

Buyers whose Russian fuel came mainly by pipeline had steady volumes, so seven countries get narrower bands (Germany ×0.74–1.31, Hungary, Slovakia, Czechia, Poland, Finland, North Macedonia). Poland had decided in 2019 to end its Yamal gas contract in 2022, so its no-war path drops that gas from 2023.

### Power of Siberia

China's no-war path includes the gas the Power of Siberia contract (2014) already foresaw: 15.4 bcm in 2022 rising to 38 bcm. Only deliveries beyond that count as war-driven. China raised Russia's earnings from trade by 0.74% of GDP.

### Japan's Sakhalin-2 LNG

LNG is about a quarter of what Japan bought from Russia. Its no-war band comes from Japanese buyers' long-term Sakhalin-2 contracts (about 5 Mt a year): the floor is the contracted volume, the top the pre-war maximum. The rest of Japan's purchases keeps the generic band. The middle value does not change.

### Forced losses: strikes, export bans and a pipeline halt

Ukraine's drones hit Russian refineries, oil ports and the Druzhba pipeline. That limited what Russia could export, whatever its buyers wanted. A buyer that wanted Russian fuel and could not get it did not choose to buy less, so the loss is taken out of its baseline: its change is counted as if the loss had not happened. The same applies to two other losses the buyers did not choose.

- Strikes. Oil product exports (diesel, fuel oil, gasoline, naphtha): about 5 Mt in 2024, 9 Mt in 2025 and 19 Mt in January–August 2026, or 4%, 6% and 20% of what Russia exported before the war (likely 2–11 Mt, 4–16 Mt and 12–29 Mt). In 2022–2023 no loss shows in the export figures. Crude: ports stopped for days at a time, but the crude that idle refineries could not use was exported instead, so crude exports did not fall; only the high end of the range counts a small loss. Druzhba: flows to Hungary and Slovakia stopped for about a week in August 2025. Russia's export bans from March 2024 answered strike-driven shortages and are counted inside these figures.
- Russia's 2023 export ban. From 21 September 2023 Russia banned gasoline and diesel exports to cut its own pump prices; diesel by pipeline was allowed again in October, gasoline on 17 November. About 0.9 Mt was lost (likely 0.5–1.4 Mt: gasoline about 0.65 Mt, diesel about 0.25 Mt). Armenia, Belarus, Kazakhstan and Kyrgyzstan were exempt and get nothing.
- The Druzhba halt of 2026. A Russian strike near Brody in Ukraine stopped the pipeline to Hungary and Slovakia from 27 January to 22 April 2026: 86 days (counted 80–92), about three months of their Russian crude. Czechia had stopped by its own choice in 2025 and is not affected.
- Who it applies to. Seaborne losses are shared over the countries that were actually buying that product at the time, in proportion to what they bought; pipeline losses go to the pipeline's buyers. A buyer's baseline is never lowered by more than it would itself have bought of that product. Buyers whose own embargo had already stopped the product (the EU, the UK, the US and others, from 2023) get nothing: the loss changed nothing for them.
- What does not change. The price windfall (what was actually bought, times the war-time price rise) and the laundered oil are not touched.

In total, EUR 13bn (likely 8–21bn) of expected purchases over 2022–2026 is taken out for 60 buyers: strikes EUR 11.8bn, the 2023 ban EUR 0.3bn, the Druzhba halt EUR 1.0bn (Slovakia 0.65bn, Hungary 0.35bn). Türkiye, China, Singapore, India and Brazil are most affected by the strikes. What they did not buy because of these losses no longer counts as money kept from Russia, so the world's net change in payments to Russia is about EUR 3.7bn smaller (at Russia's share of the money).

### Deliberate choices

- Actual figures as inputs. The models use each country's actual 2022–2026 GDP and trade with others. If a country's economy shrank because of the war, it would have traded less with Russia anyway; that is not a choice to credit or blame.
- Ranges as tested. They are not widened for Russia-side shocks: stretching every range for a rare shock would blur the result more than it adds.

## 8. Models we tried

"2021 held flat" was tested against several alternatives. A new model is adopted only if it is at least 15% more accurate at every starting year from 2005 to 2016, its ranges are not wider, and the placebo test (the method rerun for years without a war) still finds the true value inside the range at least 90% of the time. The typical miss over five years, in the back-tests:

| what was tried | typical miss | result |
| --- | --- | --- |
| The current rule (2021 held flat, median start for small buyers) | 32% (small buyers 60%, mid-size 38%, large 16%) | kept |
| 2021 held flat for everyone | 33% | the base for the rules below |
| Product detail (round 44) | no better than the total | rejected |
| Mid-size buyers start from 2021, small buyers raised by their measured bias (round 47) | 13% better for mid-size buyers | applied at the owner's request, below the bar |
| Buyer demand: the buyer's economy and imports (round 48) | 26% | better on average, but not 15% better at every starting year |
| Russia's supply (round 49) | 30% | better on average, but not 15% better at every starting year |
| Predicted share of Russia in the buyer's imports (round 50) | large buyers 15.6% against 16% | barely different |

None passed. Some were better on average but not at every starting year, or only for some buyers. The simple rule is about as accurate as any of them, and it is easy to check.

[How it works](https://warbalance.org/how-it-works/) · [Limitations](https://warbalance.org/limitations/) · [Changelog](https://warbalance.org/changelog/) · [Open in the map](https://warbalance.org/#view=how)

Cite as: War Balance, figures to August 2026, https://warbalance.org/method/counterfactual/