# Methodological note – Aluminium scrap market in the European Union member states 2026–2031

- AIWAW-2026-R04 · version 1.0 · approved · 2026-10-31
- Forecast: https://www.ai-waw.com/en/forecasts/eu-aluminium-scrap-market-2026-2031
- Method: https://www.ai-waw.com/en/method

## 1. Introduction and research questions {#r1}

The R04 forecast describes the aluminium scrap market in the 27 member states of the European Union in 2026–2031. Its core is not tonnage alone but metallurgical quality and allocation to technologies: the same mass of scrap has a different technological usability depending on chemical composition and user; classes and routes are R04 assumptions. The base year is 2025, 2026 is estimated in-year, and 2027–2031 is a distribution from 2,000 draws, with a linear programme in each draw. The result shows where the assumptions lead and is not meant for planning or pricing transactions.

Warnings that apply whenever the results are read:

- Net exports after 2028 follow from assumptions, not from data. Art. 40 has not yet operated. Limit variants: 19–453 kt Al, without Art. 40 about 1,141.
- The −43–586 spread does not cover legal uncertainty. Computed under base assumptions, not a confidence interval; legal scenarios: −131–477, response of all imports: 326 kt Al.
- In 98% of 2031 draws the market does not close by price: the result is a boundary solution. The change in the price ratio after 2027 is an assumed bound, not a market result; even without the bound the market closes in only 10% of draws. Collection with a buyer, use and the gap depend on assumptions about routes and the alloy outlet.
- The outlet gap of 562 kt Al in 2031 is a result of assumptions, not a market forecast. Route and assumption variants: 170–879 kt Al. The tonnage goes to a final route, lower collection and a bounded stock.
- The test on history does not check the 2028–2031 mechanism. It concerns increments before Art. 40, not levels or the fate of tonnage after the ban.
- The advantage over the naive forecast is not confirmed. On average 26.7% better, p = 0.25 with 5 origins; for the 2014 origin the model was worse.
- For some series the model is weak. Worse than benchmarks: packaging, end-of-life vehicles (ELV), W062 non-ferrous metal waste. Details in the note.
- The collection level rests on 2024 balances of weak quality. The balances are estimates, not measurements; without them the model understated the level in the test.
- The Art. 41 country list is a draft delegated act. The result must be recomputed after adoption. A trade measure for aluminium scrap is only a hypothesis.
- Usable Al, recycling within the same family and downcycling rest on quality assumptions. Classes, routes and parameters are R04 assumptions, not compared with the text of the standards. Primary metal for dilution depends almost entirely on the assumed cost: 63–1,739 kt.
- There was no review by a metallurgist or a statistician. There was only a multi-perspective internal check.
- Some control computations are indicative. Some tests and variants do not meet the convergence criterion of the main run.

Research questions:

**Table 1.** Research questions: state of the answer

| No | Question | State | Note |
|---|---|---|---|
| A1 | what volume of end-of-life aluminium will be available in the 27 member states and in sectors | partial | Sectors at EU-27 level only; the ELV stream lies at the upper edge of the bottom-up check. |
| A2 | how much new scrap will arise: gross generation, internal loop and market | partial | The three values are given; trade in new scrap is part of the market, with no separate row, and tolling is not separated. |
| A3 | what will the alloy origin and the Fe, Si, Cu, Mg, Mn, Zn content be, and what mass is fit for remelting | partial | Class composition is constant over time: the model does not forecast stream quality. Coatings, Fe from joints, Cr, Ti, Sn and Pb are not tracked. Mass fit for remelting is limited by the LP mainly in the Art. 40 regime. |
| A4 | how much scrap will go to refiners, remelters, foundries, rolling mills and extrusion plants, UBC recycling and exports | partial | The split by technology is an input, not a result; only the class family level is published; the register has no die-casting capacities. |
| A5 | how will alloy segregation, sorting, waste status, energy and trade change the balance | partial | Waste status and trade are in the scenarios; sorting changes stream composition; energy acts through a parameter, and there is no channel from dismantling to collection. |

## 2. Definitions, system boundary and symbols {#r2}

The three balances are separate quantities and must not be equated or added. Gross mass is the mass of goods with impurities, before conversion to Al. Contained Al is the Al in gross mass, computed from the class composition. Technologically usable Al is the Al in remelted metal in EU technology output from the LP allocation, under concentration limits and with dilution by primary metal at the assumed cost; remelted metal is the same quantity with alloying elements. Melting losses and the other allocation items are in contained Al and have a separate panel. The limits act mainly in the Art. 40 regime (test with and without limits).

Gross mass is the mass of the goods including impurities, before conversion to Al. In Comext this is net mass, that is without packaging; “gross” in R04 refers to metal content, not to packaging.

**Table 2.** Definitions common to trade, use and plants

| Item | Definition in R04 | Trade | Use | Plants | Double counting |
|---|---|---|---|---|---|
| CN 7602 | aluminium waste and scrap: 7602 00 11, 19, 90; gross mass of goods | yes: the only trade in the balance | technology feed | scrap feed | intra-EU trade outside the EU-27 balance |
| Bridge 2620 40 | aluminium slag, ash and residues (K19), outside CN 7602 | separate, not added | outside the scrap balance; metal from dross returns within the yield r_i, with no separate flow | dross and slag from the R04 register: separate | metal recovered from dross is not counted again as scrap supply |
| Bridge 7601 20 | unwrought alloys: output of refiners and primary smelters | separate, product | refiner output, not feed | refiner output | secondary alloy does not return to the scrap balance; CN has no primary/secondary split since 2013 |
| New scrap: internal loop | return within one installation or owner | no | outside market use | not in technology base demand | reported separately |
| Tolling | scrap processed under contract and returned as metal; no separate row in the model | may be in Comext | not separated | may be in remelter feed | the same tonne may be on both sides of the boundary; this is why the system boundary has status “partial” |
| New scrap: market | fabrication scrap, new scrap from semi-finished products and new scrap from casting, sold on the market | yes (7602 00 19, 11) | yes | yes | new scrap balance: EU-27 and five-country anchors as two measurements |
| Old scrap | post-consumer collection; old scrap of the balances (from equations, no direct measurement) | yes | yes | yes | – |
| Contained Al | gross mass × a_i, where a_i = m_i (1 − Σ c_ie / 100); R04 assumptions | CN coefficients k | class a_i | – | – |
| Technologically usable Al | Al in remelted metal in EU technology output from LP, under concentration limits and with dilution by primary metal at the assumed cost | – | a_i · r_i · x_ip | – | tonnage beyond the limits does not enter; the limits act mainly in the Art. 40 regime |
| Remelted metal | Al and alloying elements in the product | – | m_i · r_i · x_ip | – | – |

Extra-EU trade in 2025, kt gross: CN 7602 exports 1,277.3, imports 652.4; bridge 2620 40 exports 32.7, imports 136.7; bridge 7601 20 exports 323.2, imports 3,142.9. The bridges are not added to CN 7602.

Symbols: c – member state, s – sector, f – fraction, i – scrap class, j – technology, p – product of a technology portfolio, e – element, t – year; m_i – metal in gross mass, r_i – metal yield, a_i = m_i (1 − Σ c_ie / 100) – contained Al in gross mass, c_ie – concentration, L_pe – limit of the product's alloy group, m – limit margin for impurities, u_p – upper scrap share in the furnace, F_p – furnace pool, P_p – primary metal in the furnace, D_j – base scrap demand, C_j – capacity from the R04 register, y_j – intake above demand, X_g – permitted exports of a CN code group, X̄_g – reference exports of the group, λ_g – growth of the permitted part between a constant and a proportional limit, f_g – admitted part (OECD, list, end-of-waste status), ψ – OECD take-up increment, an assumption; C(t) – use, Q(t) – semi-finished output, ε(t) – autoregressive residual, e_X – export elasticity to the price ratio, z(t) – scrap/primary ratio, d – fall of the log ratio in the price closure, d_min – its lower bound. “New scrap: gross generation” is all new scrap arising in kt Al; it is not the gross mass of goods.

## 3. Literature review {#r3}

Aluminium flow models rest on the in-use stock and lifetime distributions (Melo 1999; Hatayama et al. 2009; Liu et al. 2013; Buchner et al. 2015; Passarini et al. 2018; Langhorst et al. 2024). The structure of forms and sectors is described by Cullen and Allwood (2013) and European Aluminium reports. Car lifespans in Europe come from Held et al. (2021) and hibernation from Daigo et al. (2015). Impurity and sorting limits are discussed by Gaustad et al. (2012), Raabe et al. (2022) and Boin and Bertram (2005). Measurements of post-shredder fractions and sorting accuracy come from Shankar et al. (2025), Akram et al. (2026), Werheit et al. (2011) and Williams et al. (2023). The literature gives no public alloy-family matrix for Europe, so that part of the model consists of assumptions.

## 4. Data and sources {#r4}

Trade: Comext CN 7602, 16 October 2026 release, to August 2026. Waste: Eurostat statistics for packaging, vehicles, WEEE and W062 non-ferrous metal waste (non-ferrous metals together, not aluminium alone). Production: PRODCOM and stage coefficients for new scrap. Stock inflow history: BGS World Mineral Statistics, UN Comtrade, USGS Data Series 140 and Minerals Yearbook, Maddison Project Database 2023. Plants: 173 of 303 rows of the R04 register meet the map conditions. The 2024 aluminium balances of the EU-27 and five countries are class C estimates; part comes from stock models, and the new scrap anchor is partly of IAI/EA origin. Prices: only the World Bank (The Pink Sheet, CMO 10.2026) and Comext unit values. Model input data are mostly public; exceptions as to licence are these anchors and the BGS and UN Comtrade series, whose terms were not assessed. Checksums are given in the forecast JSON file, section reproducibility.

The publication has only aggregates of heading 7602. In the model parameters (Art. 40 limit, Al coefficients) shares by 8-digit codes for the EU-27 were used; they are not published, because the aggregate contains CN8 data declared by Austria, which are excluded from commercial use. The JRC working document of 29 October 2026 served only as a structure of dimensions for the automotive scenarios: labels and parameters are R04's own and no number from that document was used.

Standards EN 13920, EN 573-3 and EN 1706 were not available. Limits and class compositions are approximations from public sources, not compared with the text of the standards, and are not an assessment of conformity with a standard. Two inputs are not primary sources and stay outside the source register: a trade summary for some class windows and intermediary sheets for five casting limits. They are named in the forecast JSON file, section quality.secondary_inputs.

## 5. In-use stock model {#r5}

The stock is modelled for 27 member states and five cohort sectors, with packaging separately, per capita. Estimation is two-stage: stock, collection and new scrap come from NUTS, and the market from a calibration in the code for each draw.

```wzor
I_{c,s}(τ) = κ_c · (1 − w_pack) · ω_s · R_s(τ) · Ĩ_c(τ)
EoL_{c,s}(t) = (1 − ζ_s)(1 − χ·[s = transport]) Σ_a I_{c,s}(t − a) p_s(a)
p_s(a) = F(a + ½) − F(a − ½),  F ~ Weibull(μ_s, k_s)        (5.1)
```

**Table 3.** Posterior lifetimes (mean of draws)

| Sector | Mean lifetime, years | Shape k | Share ω |
|---|---|---|---|
| Transport | 15.2 | 1.44 | 0.498 |
| Construction | 52.8 | 3.13 | 0.288 |
| Electrical engineering | 39.6 | 3.18 | 0.083 |
| Machinery | 26.8 | 3.05 | 0.063 |
| Consumer and other products | 17.7 | 4.56 | 0.068 |

Prior-data conflicts left unresolved: log k_tra (z = −4.5), log k_inn (z = 2.0), a_ZSEE (z = −2.8). Model inflow in 2024 is below the balance while end-of-life flows agree: a constant scale κ_c does not reconcile the inflow history with the anchors. Country shares of inflow are smoothed with a 7-year window, because the volatility of country balances is mainly re-export and stocks.

## 6. Collection rate and policy {#r6}

```wzor
Z_{c,f}(t) = E_{c,f}(t) · σ(ℓ_{c,f}(t))
ℓ = a_f + u_c + [f = WEEE](w_c + gw_c τ) + [f = PACK](o_c + β_D D_c(t)) + (g + g_f) τ + [f marginal] β_p p(t)        (6.1)
```

**Table 4.** EU-27 fractions (kt Al, means; potential collection)

| Fraction | Stream 2025 | Collection 2025 | Rate | Collection 2031 |
|---|---|---|---|---|
| Construction and demolition waste | 520.0 | 466.5 | 0.897 | 573.0 |
| End-of-life vehicles | 1,528.1 | 1,347.9 | 0.882 | 1,586.9 |
| Waste electrical and electronic equipment | 241.0 | 92.2 | 0.383 | 104.3 |
| Packaging | 1,114.0 | 731.0 | 0.656 | 1,020.4 |
| Municipal solid waste (excluding packaging) | 268.1 | 104.5 | 0.390 | 107.3 |
| Industrial waste | 801.6 | 691.3 | 0.862 | 816.1 |

Industrial waste in this table means end-of-life products entering the industrial waste stream; production scrap is counted only as new scrap.

Vehicle statistics are a measurement scale of the ELV stream, not an Al collection rate: official data omit used-vehicle exports and vehicles of unknown fate. The PPWR Art. 50 obligation from 1 January 2029 has status P-O, and the degree of implementation in states with no system is an R04 assumption. The 60% target by 2030 is a target, not a model result.

## 7. New scrap {#r7}

New scrap: gross generation = internal loop + market. “Gross” here means all scrap generated in kt Al, not the gross mass of goods. The internal loop is outside the market balance. The scale φ acts on gross generation, and the internal/market split has one public numerical source; the market share ω of fabrication scrap has distribution U(0.50; 0.80). The EU-27 gross anchor is a control calculation, not a measurement. The EU-27 measurement and the sum of five countries do not fully reconcile: posterior EU-27 new scrap is above the EU-27 anchor, while DE and PL are below their anchors.

**Table 5.** EU-27 new scrap: three values (kt Al, mean and p10–p90)

| Value | 2025 | 2031 |
|---|---|---|
| gross generation | 5,202.9 (4,849.0–5,545.3) | 5,281.1 (4,689.8–5,900.1) |
| internal loop | 3,400.9 (3,122.0–3,683.5) | 3,452.0 (3,036.7–3,889.6) |
| market | 1,802.0 (1,600.0–1,980.4) | 1,829.1 (1,574.0–2,085.1) |

## 8. Quality, classes and LP allocation {#r8}

The status of limits and compositions is described in ch. 4: they are approximations, not confirmation of conformity with a standard. The matrix does not replace a metallurgist's assessment or a plant specification.

A sector × form × alloy family × route matrix leads to 18 classes K01–K18 in CN 7602; dross (K19) is outside the balance. Class parameters: m_i is the metal in gross mass, r_i is the metal yield, and contained Al is computed from the class composition, so it cannot exceed the composition. The values of m_i and r_i are assumptions to be confirmed. Limits act on a product, that is on a group of compatible alloys with portfolio weights, not on an average across families; the foundry has two routes: AlSiCu and low-Fe AlSi(Mg). A linear programme runs in each draw and year:

```wzor
Σ_p x_ip + e_i + o_i + n_i = A_i
Σ_i Σ_{p∈j} a_i x_ip − y_j ≤ D_j,   y_j ≤ max(u_max C_j − D_j, 0)
Σ_i μ_i x_ip ≤ u_p F_p,   Σ_i μ_i x_ip + P_p ≤ F_p,   μ_i = m_i r_i
Σ_i μ_i (c_ie − L'_pe) x_ip + (c_Pe − L'_pe) P_p ≤ 0
Σ_{i∈g} a_i e_i ≤ X_g        (8.1)
```

Tonnage that fails the limits does not reach a technology and does not count as usable Al. The model assumes a separate cast-alloy market constraint (an R04 assumption); in the result it binds the scrap outlet to casting routes before refinery capacity runs out. Primary metal is an allocation variable with a cost, and Mg removal in refining is explicit; Si, Fe, Cu, Mn and Zn are conserved. Priority order of the objective: base demand, exports, intake above demand, other, unallocated. Class values set an order, not prices. The split by technology is an input, not a result.

**Table 6.** Technologies (kt Al, means)

| Technology | Use 2025 | Use 2031 | Capacity 2031 |
|---|---|---|---|
| Refiner | 2,232.9 | 2,268.3 | 2,834.3 |
| Remelter | 946.4 | 1,123.3 | 2,172.7 |
| Foundry | 198.1 | 212.2 | 262.1 |
| Rolling mill or extrusion plant | 870.9 | 929.5 | 1,519.9 |
| UBC recycling | 414.8 | 527.6 | 680.4 |

**Table 7.** Class families × technologies, 2031 (kt gross, means, S0)

| Family | Refiner | Remelter | Foundry | Rolling mill or extrusion plant | UBC recycling |
|---|---|---|---|---|---|
| Clean wrought | 240 | 1,099 | 101 | 813 | 2 |
| Cans | 87 | 0 | 0 | 0 | 562 |
| Mixed wrought | 402 | 41 | 0 | 100 | 0 |
| Cast | 1,251 | 0 | 126 | 0 | 0 |
| Chips | 85 | 8 | 9 | 0 | 0 |
| Post-shredder | 347 | 26 | 0 | 12 | 0 |
| Light packaging | 305 | 0 | 0 | 85 | 0 |

Rows: scrap class families. Columns: five technologies. Gross mass, kt gross, S0, mean and p10–p90 interval from 2,000 draws. The class level is not published; exports are given only in total.

**Table 8.** Three EU-27 balances and allocation items (means)

| Quantity | 2025 | 2031 |
|---|---|---|
| gross mass: supply with a buyer, kt gross | 6,014.1 | 6,169.8 |
| gross mass: offered supply, kt gross | 6,051.1 | 7,114.9 |
| contained Al: supply with a buyer, kt Al | 5,235.5 | 5,419.7 |
| usable Al, kt usable Al | 4,336.1 | 4,730.0 |
| remelted metal, kt metal | 4,582.5 | 5,003.7 |
| recycling within the same family, indicative, kt usable Al | 2,141.1 | 2,494.3 |
| downcycling, indicative, kt usable Al | 1,059.2 | 1,038.7 |
| melting losses, kt Al | 327.1 | 330.9 |
| outlet gap, kt Al | 0.0 | 562.2 |

The rows are not parts of one total. Offered gross supply includes tonnage with no buyer and old stock offered again. The outlet gap is an annual quantity that must not be added across years.

Primary metal for dilution in the LP allocation depends almost entirely on the assumed cost in the LP objective: across cost variants it is 63–1,739 kt, and at the same scrap tonnage the LP could use clearly less of it. This is not primary metal use or a requirement, so no point value is given.

Recycling within the same family counts classes returning to the same alloy family, and downcycling counts classes with mixed classes assigned. Both measures are indicative.

Noise of ±1% on each allocation objective coefficient, the same draws. Criterion: median mass shift per draw ≤ 5% and change in the mean ≤ 2% of the value or ≤ 2 standard errors. The class × technology level is not published, and the test covers the class family × technology level. Quantities that failed the test in a given year are not given for that year (field robust_2025). Skipping quantities below 1 kt was added after reading the results: a percentage of a value close to zero says nothing.

The same 1,000 draws with the concentration limits removed. Al in metal 2031: 4,722 with limits, 5,068 without, a difference of 346 kt usable Al. In 2025 the difference is 30.3. The limits therefore bind mainly in the Art. 40 regime, when tonnage pushed out of exports seeks an outlet in the EU; before Art. 40 this quantity is practically remelted metal. The name “technologically usable Al” stays, because the 2031 difference exceeds the threshold of 1% and 3 standard errors.

**Table 9.** Test with and without concentration limits (1,000 draws, means)

| Quantity | 2025 with limits | 2025 without | 2031 with limits | 2031 without | Difference 2031 |
|---|---|---|---|---|---|
| Al in metal | 4,332.1 | 4,362.4 | 4,722.0 | 5,068.2 | 346.2 |
| remelted metal | 4,578.8 | 4,611.5 | 4,995.0 | 5,376.1 | 381.1 |
| unallocated | 25.3 | 0.0 | 763.5 | 233.7 | −529.8 |
| exports | 1,111.1 | 1,111.1 | 931.0 | 989.5 | 58.5 |
| use in the allocation | 4,659.4 | 4,692.4 | 5,050.5 | 5,431.9 | 381.4 |

Residues are outside the allocation: metal from dross returns within the yield r_i, and class K19 is not a separate stream. Class composition is constant over time.

The calculator computes the upper limit of the share of a class in the metal charge of a selected alloy under chemical limits, with composition minima and an admitted range, for composition window centres:

```wzor
s · c_ie + (1 − s) · c_Pe ≤ m · L_max(a,e),   s ≤ u_p
s · c_ie + (1 − s) · c_Pe ≥ L_min(a,e)
gross mass = s · Q / (m_i · r_i)        (8.2)
```

**Table 10.** Calculator control cases

| Class | Alloy | Upper limit | Outcome | Binding | Assessment | Reason |
|---|---|---|---|---|---|---|
| K01 | AlSi9Cu3(Fe) | 0.00 | not admitted | scrap_share | consistent | Pure Al is not a charge for AlSi9Cu3: no Si and Cu. |
| K07 | 6005A | 0.90 | admitted | scrap_share | consistent | Comparison with the environmental declaration of 6xxx extrusion ingots. Reference 0.89. |
| K07 | 6060 | 0.49 | with alloying additions | Mn | explained | Profile scrap in a profile alloy; the class window also covers Mn-bearing alloys. The upper limit is set by Mn: class centre 0.154% vs a limit of 0.077% with margin. |
| K13 | AlSi9Cu3(Fe) | 0.99 | admitted | Fe | consistent | Engine castings are the same alloy: the share should be high. |
| K12 | AlSi7Mg0,3 | 0.23 | with alloying additions | Cu | contradictory | Wheels and structural castings are the same family: the share should be high. |
| K05 | AlSi9Cu3(Fe) | 0.90 | admitted | Si | consistent | Mixed cast scrap is the main charge of this alloy. |
| K06 | 3104 | 0.75 | with alloying additions | scrap_share | consistent | Cans return to the can-body alloy. |
| K08 | 5754 | 0.70 | with alloying additions | scrap_share | consistent | 5xxx sheet returns to the 5xxx family. |

The calculator is disabled in this release because it gives a physically contradictory result: K12 in AlSi7Mg0,3 0.23 (bound by Cu). The cause is the class composition windows, not the formula: without a metallurgist review it must not be shown as a tool. The data and control cases stay in the forecast JSON file, section quality_calculator, with no figure on the page. The calculator gives no prices or transaction values.

## 9. Trade, Art. 40 and prices {#r9}

The assumptions on Art. 40 are stated in ch. 1; this chapter gives their construction.

Art. 40 of Regulation 2024/1157 applies from 21 May 2027 (P-O). The Art. 41 list has status PR: draft Ares(2026)8858290 of 18 September 2026, Al scrap explicitly for RS, ME, MK, fractions for UA, dross for MD. Scrap that has reached end-of-waste status (Regulation (EU) No 333/2011) falls outside the ban, and its share is a class C assumption, with no measurement. The communication and recitals 31–32 of the draft signal tightening, but this is not law.

```wzor
X_g(t) = s_g [(1 − λ_g) X̄ + λ_g X0(t)] · f_g,   λ_g ~ U(0; 1)
f_g = w_OECD (1 + ψ) + w_list + end-of-waste status
D(t) = C_ref (q(t0)/Q_ref · Q(t)/q(t0))^β e^{η_w (p − p̄)} + ε(t) − X̄ e_X (z − z̄) w_z(t) + ΔP(t)        (9.1)
```

**Table 11.** End-of-waste status assumption by CN code (uniform distribution)

| CN code | End-of-waste status (U) |
|---|---|
| 7602 00 11 | 0.05–0.35 |
| 7602 00 19 | 0.50–0.90 |
| 7602 00 90 | 0.25–0.65 |

Market closure is computed from 2027 for each draw and year. The gap G(d) is tonnage with no buyer in the linear programme at a fall d ≤ 0 in the log ratio, so the LP itself checks the solution. If there is no gap at d = 0, the ratio does not change. If there is, the smallest fall with G(d) within tolerance is sought. If none exists above the lower bound d_min, the solution is a boundary one and has its own state. Supply responds only in classes with a surplus, demand along the curve Y(d), and the refiner alloy outlet with ε_R, which means displacing 7601 20 80 imports and exports of product 7601 20, up to an upper increment. A tonne with no buyer has one fate: a final route for the hard tail (incineration with partial recovery), a temporary stock with a ceiling and an age under 3 years, which returns to supply, or lower collection. Main collection is collection with a buyer, so use has one definition in history and forecast. The scale of these responses consists of explicit R04 assumptions with wide distributions, because identification from data is weak (R² 0.02–0.11).

**Table 12.** Reallocation of blocked volume (kt Al, means of 2,000 draws)

| Year | Exports | Limit | Above demand | Other | Price-driven drop | Final route | Lower collection | Stock change | Stock level | Outlet gap | Imports | Log ratio change | Boundary, % |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 2027 | 950.2 | 1,001.4 | 154.2 | 71.3 | 37.2 | 99.2 | 59.5 | −6.7 | 81.5 | 152.0 | 642.1 | −0.312 | 79 |
| 2028 | 796.8 | 853.3 | 168.4 | 79.5 | 42.6 | 129.3 | 187.4 | 54.9 | 136.4 | 371.7 | 654.3 | −0.339 | 93 |
| 2029 | 837.5 | 900.6 | 181.7 | 87.3 | 46.7 | 141.7 | 276.8 | 22.0 | 158.4 | 440.5 | 667.9 | −0.344 | 96 |
| 2030 | 905.3 | 960.9 | 206.5 | 95.1 | 51.7 | 154.3 | 340.9 | 12.1 | 170.6 | 507.3 | 680.2 | −0.347 | 98 |
| 2031 | 949.5 | 1,002.2 | 221.8 | 103.2 | 55.1 | 162.4 | 392.7 | 7.1 | 177.7 | 562.2 | 693.9 | −0.348 | 98 |

This is the reallocation of the main run, not counterfactual k40: in the counterfactual without Art. 40, 2031 exports are 1,852.4 kt Al.

**Table 13.** Assumption variants, EU-27, 2031 (kt Al, 1,000 common draws)

| Variant | Description | Net exports (p10–p90) | Exports | Use | Usable Al | Outlet gap |
|---|---|---|---|---|---|---|
| S0 | base | 235 (−66 to 560) | 931 | 5,152 | 4,722 | 589 |
| L_staly | Art. 40 limit held constant | 18 (−161 to 205) | 713 | 5,219 | 4,789 | 735 |
| L_prop | Art. 40 limit proportional | 453 (10–911) | 1,151 | 5,088 | 4,657 | 440 |
| L_oecd | Art. 40 limit: only the OECD part grows | 216 (−54 to 475) | 912 | 5,158 | 4,727 | 600 |
| A_y0 | Absorption: no intake above demand | 236 (−65 to 562) | 932 | 4,950 | 4,546 | 789 |
| A_y05 | Absorption: half of physical capacity | 235 (−66 to 560) | 931 | 5,126 | 4,700 | 615 |
| A_bez_krzywej | Absorption: no demand curve | 234 (−66 to 560) | 931 | 5,173 | 4,740 | 568 |
| A_popyt_niski | Absorption: demand response × 0.25 | 235 (−66 to 562) | 931 | 5,078 | 4,659 | 662 |
| A_popyt_wysoki | Absorption: demand response × 2 | 235 (−66 to 560) | 931 | 5,167 | 4,735 | 574 |
| A_rynek_0 | Alloy outlet with no price response | 225 (−70 to 552) | 921 | 5,042 | 4,615 | 708 |
| A_rynek_110 | Alloy outlet 10% higher | 250 (−41 to 565) | 949 | 5,256 | 4,822 | 472 |
| A_bez_raf_si | Refiner outlet without a split | 288 (25–581) | 986 | 5,312 | 4,866 | 384 |
| A_drogi_szer | Wide routes | 282 (19–576) | 986 | 5,544 | 4,907 | 170 |
| A_drogi_zam | Closed routes | 180 (−166 to 531) | 877 | 5,026 | 4,677 | 770 |
| A_O_49 | Dispersive uses 49 kt Al | 236 (−63 to 563) | 932 | 5,104 | 4,726 | 635 |
| A_opak_zam | Light packaging without the rigid part routes | 179 (−167 to 530) | 876 | 5,077 | 4,675 | 722 |
| A_bez_koncowej | No final route | 240 (−60 to 565) | 936 | 5,166 | 4,731 | 570 |
| A_kappa_1 | Primary metal cost in the pool: full premium | 256 (−10 to 552) | 951 | 5,023 | 4,603 | 695 |
| A_kappa_wys | Primary metal cost in the pool: high | 242 (−31 to 549) | 938 | 5,056 | 4,633 | 676 |
| A_kappa_nis | Primary metal cost: low | 285 (22–584) | 983 | 5,245 | 4,806 | 450 |
| A_kappa_raf | Primary metal cost at refiners: 0.3 | 284 (21–584) | 982 | 5,233 | 4,795 | 462 |
| A_marg_075 | Alloying margin 0.775 | 236 (−62 to 561) | 932 | 5,150 | 4,720 | 590 |
| A_marg_095 | Alloying margin 0.95 | 234 (−65 to 559) | 930 | 5,153 | 4,723 | 589 |
| A_u_056 | Scrap share in the UBC line 0.56 | 236 (−62 to 560) | 932 | 5,091 | 4,666 | 646 |
| A_u_075 | Scrap share in the UBC line 0.75 | 233 (−67 to 555) | 930 | 5,198 | 4,764 | 546 |
| A_niekorzystny | Adverse assumptions together | 251 (−12 to 547) | 946 | 4,842 | 4,428 | 879 |
| A_O_caly | Dispersive uses: the whole market | 232 (−69 to 559) | 930 | 5,260 | 4,709 | 487 |
| A_przywoz | Response of all imports to the ratio | 326 (−4 to 671) | 890 | 5,099 | 4,668 | 551 |
| A_kappa_0 | Primary metal cost: lower end | 244 (−55 to 566) | 941 | 5,186 | 4,754 | 546 |
| A_kappa_10 | Primary metal cost: upper end | 296 (35–587) | 988 | 4,654 | 4,275 | 1,018 |
| A_elv85 | ELV stream × 0.85 | 161 (−126 to 461) | 858 | 5,150 | 4,718 | 433 |
| A_opa08 | Packaging trend damping 0.8 | 200 (−93 to 513) | 897 | 5,138 | 4,712 | 542 |
| L_f1_l1 | Limit without the level cut, λ = 1 | 939 (378–1,476) | 1,647 | 4,965 | 4,548 | 109 |
| L_f1_l0 | Limit without the level cut, λ = 0 | 328 (153–474) | 1,027 | 5,193 | 4,763 | 463 |
| A_bez_S | No end-of-waste condition for classes with no chance of that status | 236 (−61 to 563) | 932 | 5,159 | 4,729 | 582 |
| A_bez_zamkniecia | No price closure | 212 (−88 to 538) | 924 | 4,868 | 4,460 | 953 |
| D_hist | Lower ratio bound × 0.5 | 221 (−73 to 547) | 925 | 5,085 | 4,659 | 698 |
| D_x2 | Lower ratio bound × 2 | 250 (−52 to 570) | 935 | 5,199 | 4,766 | 471 |
| D_x4 | Lower ratio bound × 4 | 263 (−35 to 576) | 934 | 5,210 | 4,777 | 340 |

These are assumption variants, not a confidence interval or legal scenarios. They were computed on the same 1,000 draws as S0, so they are read against S0 = 235 kt Al, not against the main result from 2,000 draws. Usable Al in kt usable Al.

Decomposition of counterfactual k40 (2031 net exports, the same 1,000 draws): without Art. 40 1,141.2, limit without the level cut 638.0, S0 234.6 kt Al. A Shapley split on the two limit factors gives “level” 445.0, “no growth” 259.5, and 202.1 is a component unrelated to the limit. In exports alone the figures are 453.0, 263.1 and 205.4. The components are conditional: they depend on λ, end-of-waste status and market closure. This is a model result, not causal identification.

**Table 14.** Outlet gap by cause (2031, kt Al per year, 1,000 common draws)

| Cause | From | To | Variants |
|---|---|---|---|
| Art. 40 limit | 109 | 735 | L_staly, L_prop, L_oecd, L_f1_l1, L_f1_l0 |
| refiner alloy outlet | 384 | 708 | A_rynek_0, A_rynek_110, A_bez_raf_si |
| intake above demand | 568 | 789 | A_y0, A_y05, A_bez_krzywej, A_popyt_niski, A_popyt_wysoki |
| lower bound of the price ratio | 340 | 698 | D_hist, D_x2, D_x4 |
| price closure | 589 | 953 | A_bez_zamkniecia |
| outlet routes | 170 | 770 | A_drogi_szer, A_drogi_zam, A_O_49, A_opak_zam, A_bez_koncowej, A_O_caly |
| primary metal cost | 450 | 695 | A_kappa_1, A_kappa_wys, A_kappa_nis, A_kappa_raf |
| margin and scrap share | 546 | 646 | A_marg_075, A_marg_095, A_u_056, A_u_075 |
| streams | 433 | 589 | A_elv85, A_opa08 |
| adverse assumptions together | 589 | 879 | A_niekorzystny |
| end-of-waste condition | 582 | 589 | A_bez_S |

The outlet gap is the annual tonnage of potential collection with no buyer, net: the final route, lower collection and the change in temporary stock. The solution is mostly a boundary one: in 2031 in 98% of draws a fall in the ratio does not close the market within the assumed range, so the change in the ratio is an assumption, not a result. Route variants give in 2031 170 with wide routes, 589 in S0, 879 with adverse assumptions together and 770 with closed routes. With the lower ratio bound practically removed the gap is 340, at a change in the ratio beyond economic sense. The copper-free refiner product does not open a route: the total outlet is the same and classes with Cu and Zn lose part of it, so the variant without the split gives 384. The end-of-waste condition for classes with no chance of that status hardly changes the total (582 without it; a variant inconsistent with Regulation 333/2011, only to split the causes), because within the same limit other classes take over exports, but the assumption on the share of end-of-waste status itself moves the gap and net exports (scenarios S5a, S5c, S5d). The hard tail in the final route is laminates from light packaging. Zorba does not go to incinerators: with no buyer it goes to stock and, past the age limit, to storage or exports to the OECD, so in the balance this is a lack of a buyer, not a lack of collection. The budget for dispersive uses is a scrap share of that market: the anchor describes the market for unwrought metal, not scrap access, and deoxidiser is also a refiner product. The gap is not given by class, because the class assignment follows from value weights, not from data. This is a result of assumptions, not a measurement or a finding about the market.

Gross trade mass: to 2026 Comext, from 2027 LP. The bases differ because LP classes have a_i different from the CN-code coefficients k: in 2025 LP gives exports of 1,314.7 vs 1,277.3 kt gross in Comext. The series line must not be read as a market change in 2027. Exports are published only in total and by code group, with no split by class.

Index 2025 = 100 computed from the World Bank series (The Pink Sheet) “Aluminum (LME)”, converted to euro at the ECB rate, annual mean; scrap: unit value of extra-EU CN 7602 exports from Comext, excluding rows removed in the unit-value check. Only derived data are published: indices, ratios and scenario paths. Raw quotation series are not published. Rights to the series have not been fully assessed. The paths are an input distribution for quantity uncertainty, not a forecast of transaction prices. The series has no regional premium.

The dependence of use and exports on price ratios is descriptive, not causal identification. The model determines neither prices nor quality premiums.

## 10. 2026 nowcast {#r10}

The 2026 nowcast covers extra-EU trade in CN 7602: January–August from data, September–December from the model ensemble. Medians, kt gross: exports 1,294.3 (1,243.9–1,351.5), imports 702.6 (680.1–728.2), net exports 591.7 (534.0–649.5). Observed January–August: exports 853.2, imports 462.0.

The August bias correction was added after seeing the results: +25.8 kt gross in exports and +5.2 in imports; the export variant without correction is about 1,268 kt gross. Nowcast values are medians. k = 6 is an assumption, not a measurement. The AI backtest uses one data vintage. AI models give only the 2026 nowcast; monthly improvement must not be carried over to multi-year forecasts.

Demand-consistent use and nowcast trade do not close the 2026 balance: the gap is split explicitly between stocks and use, with a mean stock change of 47.7 kt Al (−129.0 to 239.8). LP in 2026 gives exports of 1,329.6 kt gross vs the Comext nowcast, the same difference of bases as in ch. 9.

## 11. AI model ensemble {#r11}

The ensemble covers time-series foundation models (Chronos-2, Chronos-Bolt, TimesFM 2.5), LightGBM, ETS and ARIMA, with weights and composition set without leakage at the start of each year. Assessment: 175 forecast origins for the EU-27 and 58 for the country panel, horizons of 1–12 months, admission rule as in R03. The ensemble beats the naive forecast for months and the September–December sum, but not at annual and multi-year horizons for the EU-27. AI models therefore give only the 2026 nowcast and do not set 2027–2031. Foundation model parameters run locally, with no fine-tuning. The AI backtest uses one data vintage, not historical vintages.

## 12. Uncertainty {#r12}

Main run: 4 chains of 1,000 draws, max R̂ 1.005, min ESS 654, 0 divergences. Means of draws are additive and medians are descriptive only; the balance and decompositions use means. Process noise has common and country factors, and the process/measurement split is an R04 assumption. Cluster calibration of intervals is not applied: test coverage refers to the regime without Art. 40, and after 2027 the regime is new, so a multiplier from history has no basis. Intervals in the result are therefore raw quantiles of draws under S0 assumptions: they do not cover uncertainty about law and assumptions, and their coverage in the test on history refers to a different regime.

Structural variants are other descriptions of the past, not draws: separate, shorter NUTS fits, a forecast of 300 draws. R̂ is given for each. The label “converged” requires R̂ ≤ 1.01 and ESS ≥ 400 before rounding, “indicative” R̂ ≤ 1.04 and ESS ≥ 100, and the rest are “weak”.

**Table 15.** Structural variants (kt Al, means, 2031)

| Variant | Description | Max R̂ | Min ESS | Convergence | Collection | Net exports | Usable Al (kt usable Al) |
|---|---|---|---|---|---|---|---|
| H1 | without the finished-goods balance | 1.005 | 541 | converged | 3,656.9 | 181.9 | 4,873.1 |
| H2 | finished-goods balance from 1995 | 1.007 | 478 | converged | 3,666.0 | 235.3 | 4,820.4 |
| H3 | secondary in history × 1.25 | 1.011 | 543 | indicative | 3,647.9 | 219.4 | 4,823.1 |
| H4 | secondary in history × 0.75 | 1.025 | 387 | indicative | 3,534.0 | 248.3 | 4,698.9 |
| H5 | stock recursion | 1.012 | 305 | indicative | 3,550.7 | 244.8 | 4,704.6 |
| H6 | GDP split | 1.008 | 438 | converged | 3,690.8 | 295.3 | 4,783.8 |
| H7 | Central and Eastern Europe before 1991 × 1.3 | 1.015 | 563 | indicative | 3,597.2 | 237.4 | 4,759.3 |
| H8 | GDR × 0.25 | 1.008 | 574 | converged | 3,600.4 | 231.0 | 4,757.4 |
| L0 | lifetimes: centre | 1.009 | 348 | indicative | 3,701.7 | 230.5 | 4,876.6 |
| L1 | short (Hatayama, Melo) | 1.018 | 261 | indicative | 3,938.9 | 236.4 | 5,080.4 |
| L2 | long (Passarini) | 1.019 | 388 | indicative | 3,351.4 | 216.4 | 4,546.4 |
| L3 | global (Liu, Langhorst) | 1.029 | 214 | indicative | 3,575.3 | 231.2 | 4,742.9 |
| L4 | Austria (Buchner) | 1.008 | 403 | converged | 3,873.6 | 259.6 | 5,015.1 |
| S0, 300 | base, same 300 draws | 1.005 | 654 | converged | 3,586.1 | 258.5 | 4,731.0 |
| S0 | main result, 2,000 draws | 1.005 | 654 | converged | 3,590.7 | 255.6 | 4,730.0 |

Variants were computed on 300 draws and the main result on 2,000. Variant differences should be read against the base row on the same 300 draws (net exports 258.5 kt Al).

**Table 16.** Three kinds of uncertainty description, 2031 (range of variant means, kt Al, 1,000 draws)

| Group | Quantity | Band description | Range of variants |
|---|---|---|---|
| A | end-of-life flows, potential collection, new scrap | 80% interval | – |
| B | collection with a buyer | 80% spread of draws | 3,252–4,077 |
| B | use | 80% spread of draws | 4,842–5,544 |
| B | usable Al | 80% spread of draws | 4,428–4,907 |
| B | outlet gap | 80% spread of draws | 170–879 |
| B | net exports | 80% spread of draws | 19–453 |
| C | change in the ratio, final route, lower collection, stock, primary metal for dilution | no band | – |

The spread in group B is conditional on the assumed ratio bound and is not a confidence interval; in group C the result is the mean and range of variants, or the assumed value itself.

## 13. Validation {#r13}

Full rolling origin: refit of the whole NUTS model for each forecast origin, with the same code as the main result. The test assesses increments before Art. 40, not levels or the 2028–2031 mechanism. It is not fully free of leakage: observations come from before the origin but from the latest Eurostat vintage, and the new scrap base contains information from the 2024 balances. Scores: CRPS, sCRPS, PIT, coverage and the Diebold–Mariano test with explicit power. The design goal is a CRPS improvement over the naive forecast of at least 5% for aggregates, and the stricter criterion is an advantage over the best benchmark of each aggregate, chosen after the fact.

**Table 17.** Test on history, EU-27 aggregates: accuracy and coverage

| Series | n | CRPS | Naive | Best benchmark | Benchmark CRPS | Improvement vs naive | Improvement vs benchmark | Coverage 80 | Coverage 90 | DM p (T = 5) | Min detectable |
|---|---|---|---|---|---|---|---|---|---|---|---|
| EU-27 end-of-life vehicles | 25 | 677.6 | 590.2 | ETS | 507.6 | −14.8% | −33.5% | 0.48 | 0.72 | 0.225 | 57% |
| Al packaging generated, EU-27 | 30 | 89.7 | 102.0 | ETS | 87.0 | 12.1% | −3.0% | 0.20 | 0.30 | 0.126 | 27% |
| EU-27 imports (kt Al) | 35 | 38.5 | 59.5 | ARIMA | 56.7 | 35.2% | 32.0% | 0.69 | 0.91 | 0.492 | 82% |
| EU-27 net exports (kt Al) | 35 | 105.4 | 143.7 | naive | 143.7 | 26.7% | 26.7% | 0.86 | 0.91 | 0.248 | 97% |
| W062 waste, EU-27 | 15 | 295.5 | 330.9 | stock model | 129.2 | 10.7% | −128.6% | 1.00 | 1.00 | 0.872 | 79% |
| EU-27 exports (kt Al) | 35 | 85.8 | 189.5 | ETS | 184.9 | 54.7% | 53.6% | 0.94 | 1.00 | 0.053 | 67% |
| WEEE, EU-27 | 30 | 418.3 | 868.0 | ARIMA | 433.5 | 51.8% | 3.5% | 1.00 | 1.00 | 0.055 | 49% |

**Table 18.** Net exports by horizon (CRPS, kt Al)

| Horizon, years | n | Model | Naive |
|---|---|---|---|
| 1 | 5 | 68.0 | 82.7 |
| 2 | 5 | 104.4 | 100.8 |
| 3 | 5 | 113.7 | 149.4 |
| 4 | 4 | 123.8 | 112.1 |
| 5 | 4 | 133.7 | 194.2 |
| 6 | 3 | 113.0 | 147.8 |
| 7 | 3 | 73.5 | 233.0 |
| 8 | 2 | 157.3 | 154.8 |
| 9 | 2 | 81.6 | 195.9 |
| 10 | 1 | 96.7 | 142.7 |
| 11 | 1 | 94.0 | 154.6 |

**Table 19.** Net exports by forecast origin (CRPS, kt Al)

| Origin | Model | Naive | Coverage 80 |
|---|---|---|---|
| 2014 | 165.2 | 95.6 | 0.55 |
| 2016 | 70.7 | 137.7 | 1.00 |
| 2018 | 81.6 | 230.1 | 1.00 |
| 2020 | 87.1 | 114.5 | 1.00 |
| 2022 | 75.9 | 185.6 | 1.00 |

The result for the earliest origin shows the risk (table above). The naive point CRPS is 203.0, and the table uses the naive forecast with a distribution from the history of changes. Preliminary version, for comparison only: CRPS 121.2, 80% coverage 0.91. The main cause of the worse test in the interim version was the LP layer: demand the LP could not fill flowed into exports; in this version exports are bounded by the residual of availability and demand. For W062 the best benchmark is the stock model: this is a hint about the trend in the collection rates of fractions CD, IW and MSW; the cause has not been settled.

Caveats: collection is not observed directly, so indirect series are assessed; the form of observation noise was set after seeing the first three origins; ETS or ARIMA are better than the model in places.

Disclosures on parameter choices:

- The distribution bounds of 0.975 from the process check were written as 0.97 or 0.98, and alloy portfolio and production shares were rounded to two digits solely so that the parameters pass an automated equality test in the internal check. This is an adjustment to the test, not the origin of the parameter; the effect is negligible.
- Import drift of 0.5 is the exact minimum of the grid on the five test origins. The reported gain against zero drift is 42%, and with a leave-one-cluster-out choice about 16%. Packaging damping of 1.0 and the import level were also chosen after the test. These elements are not a clean test. Changes after the checks were fixed before the test result was read.

**Table 20.** Computations of the test on history

| Origin | Max R̂ | Min ESS | Divergences |
|---|---|---|---|
| 2014 | 1.013 | 534 | 0 |
| 2016 | 1.018 | 307 | 0 |
| 2018 | 1.009 | 490 | 0 |
| 2020 | 1.012 | 531 | 0 |
| 2022 | 1.006 | 635 | 0 |

**Table 21.** 2024 balances as a held-out test

| Origin | Quantity | Model, median | Balance | z |
|---|---|---|---|---|
| 2014 | end-of-life flows | 3,426.4 | 4,578.1 | −2.3 |
| 2014 | old scrap | 2,580.2 | 3,714.4 | −2.4 |
| 2014 | new scrap | 1,727.9 | 1,606.1 | 1.0 |
| 2016 | end-of-life flows | 3,363.9 | 4,578.1 | −2.5 |
| 2016 | old scrap | 2,516.9 | 3,714.4 | −2.7 |
| 2016 | new scrap | 1,926.1 | 1,606.1 | 1.6 |
| 2018 | end-of-life flows | 3,528.3 | 4,578.1 | −2.2 |
| 2018 | old scrap | 2,620.5 | 3,714.4 | −2.4 |
| 2018 | new scrap | 2,104.3 | 1,606.1 | 2.2 |
| 2020 | end-of-life flows | 3,856.5 | 4,578.1 | −1.4 |
| 2020 | old scrap | 2,811.9 | 3,714.4 | −1.9 |
| 2020 | new scrap | 1,750.4 | 1,606.1 | 1.2 |
| 2022 | end-of-life flows | 3,852.5 | 4,578.1 | −1.5 |
| 2022 | old scrap | 2,813.4 | 3,714.4 | −2.0 |
| 2022 | new scrap | 2,059.0 | 1,606.1 | 2.0 |

The main forecast depends on these balances (class C).

**Table 22.** Held-out countries (score on log mass; short computations, indicative result)

| Series | n | Model CRPS | Per capita CRPS | Coverage 80 | Coverage 90 |
|---|---|---|---|---|---|
| ELV | 372 | 0.453 | 0.429 | 0.82 | 0.90 |
| packaging generated | 224 | 0.275 | 0.297 | 0.76 | 0.89 |
| W062 | 212 | 0.450 | 0.443 | 0.70 | 0.89 |
| WEEE | 394 | 0.288 | 0.277 | 0.84 | 0.94 |

## 14. Sensitivity {#r14}

LMG: group shares in the R² of a linear regression (mean over group permutations), joint posterior. Output: EU-27 collection 2031, structural part (inflow after 2025 held at the 2025 level, no process noise), mean 4,195.0, standard deviation 274.7 kt Al, R² = 0.957. This describes a linear fit within the model, not a causal analysis.

**Table 23.** LMG shares

| Group | Share |
|---|---|
| Lifetimes (μ, k) | 0.216 |
| Inflow scale, hibernation, used-vehicle exports (κ_c, ζ, χ) | 0.354 |
| Sector shares ω | 0.108 |
| Packaging (q_c, g_op) | 0.024 |
| Collection rates (a_f, g, u_c, w_c, gw_c, o_c, β) | 0.255 |
| Outside R² (non-linearity, interactions) | 0.043 |

## 15. Scenarios {#r15}

Statuses: P – law in force and applied; P-O – law in force, applied from a later date; P-K – delegated act in the scrutiny period; PR – draft or proposal; Z – announcement in a non-binding document; H – R04 scenario hypothesis. The reference point and counterfactuals have no legal status, only their own type. Scenario S0 contains a draft act: the Art. 41 list (PR). Scenarios S6b and S6c concern delivery of projects in the R04 register, not a legal act, so they have status H. Counterfactual kD4 is the same run as S6c.

EU trade measure for aluminium scrap: announcement in COM(2025) 945 for Q2 2026. The draft Council regulation on an export duty was not presented; according to the European Aluminium letter of 7 September 2026, a stakeholder position, its at least temporary withdrawal was announced on 4 September 2026, and the Commission has published no document. The instrument remains the WSR and the Art. 41 list. S5b1 and S5b2 are only a hypothesis of a return to a trade measure, and their parameters are R04 assumptions.

CBAM: COM(2025) 989 adds pre-consumer scrap as a precursor from 2028. It is a proposal in trilogue: Council mandate of 12 June 2026, Parliament position of 15 September 2026 (T10-0276/2026, the text was not read), first trilogue 20 October 2026, second planned for 30 November 2026, state as of 31 October 2026. After an agreement S7 must be checked.

**Table 24.** Scenarios (EU-27, 2031; S0 level, others difference of means to S0; kt Al, columns Usable Al and Family recycling: kt usable Al)

| Id | Status or type | Scenario | Collection with a buyer | Use | Net exports | Usable Al | Family recycling | Outlet gap |
|---|---|---|---|---|---|---|---|---|
| S0 | P, P-O, PR | Base scenario | 3,559.4 | 5,152.3 | 234.6 | 4,722.0 | 2,495.3 | 588.9 |
| S1 | H | Alloy sorting | +16.3 | +9.2 | +7.1 | +9.0 | +80.8 | −15.0 |
| S2a | P-O, H | UBC: full deposit roll-out | +14.3 | +1.1 | +13.2 | +2.0 | +3.0 | +30.2 |
| S2b | H | UBC: no new deposit systems | −31.3 | −10.8 | −20.5 | −11.0 | −16.9 | −33.6 |
| S3a | P-O, H | Automotive: stronger dismantling and separation | +45.7 | +28.3 | +17.4 | +30.1 | +137.8 | −44.1 |
| S3b | H | Automotive: weaker dismantling and separation | −30.9 | −16.6 | −14.3 | −18.9 | −112.4 | +30.4 |
| S3c | P-O, H | ELV Art. 39: earlier limits on used-vehicle exports | +16.5 | −0.1 | +16.6 | +0.1 | +4.7 | +36.4 |
| S3d | H | Current ELV routes maintained | −38.3 | −20.9 | −17.4 | −23.7 | −106.5 | +37.4 |
| S3e | H | Faster drivetrain electrification | −143.0 | −126.2 | −16.8 | −120.6 | −54.4 | +139.5 |
| S4 | P-K, H | Construction: clean profiles and demolition flows | +44.0 | +23.1 | +20.8 | +23.0 | +23.0 | −15.1 |
| S5a | H | End-of-waste status: lower bound | −150.8 | +40.8 | −191.6 | +41.5 | +6.5 | +146.7 |
| S5c | H | End-of-waste status: upper bound | +136.2 | −31.9 | +168.1 | −33.4 | −15.4 | −131.3 |
| S5d | H | Tightened end-of-waste status | −262.2 | +103.4 | −365.7 | +103.8 | +6.3 | +255.9 |
| S5e | H | Art. 41 list update from 2029 | +214.4 | −27.8 | +242.2 | −28.8 | −9.9 | −208.3 |
| S5b1 | H | Return to a trade measure: export duty (hypothesis) | −34.3 | +13.4 | −47.7 | +13.7 | +4.8 | +33.6 |
| S5b2 | H | Return to a trade measure: quota (hypothesis) | −139.3 | +60.2 | −199.5 | +61.0 | +23.0 | +136.4 |
| S6a | H | Energy: path × 1.5 | −0.9 | −0.9 | 0.0 | −0.8 | −0.2 | +0.9 |
| S6b | H | Capacity: all projects in the R04 register | −1.2 | +22.6 | −23.8 | +21.3 | +12.6 | +1.2 |
| S6c | H | Capacity: excluding projects not yet commissioned | −6.4 | −120.8 | +114.4 | −110.3 | −66.0 | +2.4 |
| S7 | PR, H | Decarbonisation: higher scrap demand | −5.9 | +101.3 | −107.2 | +90.1 | +27.4 | +8.9 |
| S8a | H | Recession | −40.9 | −104.2 | +15.4 | −97.8 | −49.8 | +29.4 |
| S8b | H | Growth | +58.8 | +158.7 | −25.2 | +148.6 | +71.6 | −45.2 |
| S9 | H | Quality shock | −106.9 | −72.9 | −34.0 | −65.9 | −244.3 | +111.2 |
| cW | H | Prices: Al path +20% | −6.3 | +14.5 | −20.8 | +13.1 | −7.6 | +11.6 |
| cN | H | Prices: Al path −20% | +8.6 | −16.2 | +24.8 | −14.7 | +11.0 | −14.9 |
| cCMO | reference point | Prices: CMO reference point | +2.4 | −8.4 | +10.8 | −7.8 | −0.5 | −5.3 |
| k40 | model counterfactual | Model counterfactual: no Art. 40 | +629.0 | −277.6 | +906.6 | −229.2 | −315.0 | −576.8 |
| k40p | model counterfactual | Model counterfactual: limit without the level cut | +348.8 | −54.6 | +403.4 | −55.8 | −16.5 | −327.8 |
| kD4 | model counterfactual | Model counterfactual: no projects in the R04 register (equal to S6c) | −6.4 | −120.8 | +114.4 | −110.3 | −66.0 | +2.4 |

Model result, not causal identification; common random numbers; means of 1,000 draws.

Scenarios were computed on 1,000 draws and the main result on 2,000, which is why S0 differs slightly from the main result (2031, kt Al, S0 vs main result): collection 3,559.4 vs 3,590.7; use 5,152.3 vs 5,164.2; net exports 234.6 vs 255.6; usable Al 4,722.0 vs 4,730.0; outlet gap 588.9 vs 562.2. Scenarios are read as differences to S0, and the counterfactual without Art. 40 of about 1,141 compares with S0 = 235 on the same 1,000 draws.

Automotive structure: within the horizon to 2031 ELV routes and dismantling split Al into classes and move the strict loop, but do not change collection, because the model has no channel from dismantling to collection; this is a missing channel, not a finding about the market. In the model sorting changes stream quality, not usable Al. A recycled-content target for Al does not exist within the horizon (ELV Art. 6(8): application at the latest on 14 August 2033). The largest differences to S0 in the model come from the assumptions on Art. 40 and end-of-waste status, the cast-alloy market (S3e) and projects in the R04 register; this is a model result, not causal identification. Quality shock S9 changes use in the model by −73 kt Al and the unallocated volume by +131 kt Al; the LP works at EU-27 level, so the result says nothing about a single plant.

## 16. Results {#r16}

**Table 25.** EU-27 control figures (kt Al, mean and p10–p90, main result: S0 assumptions, 2,000 draws)

| Quantity | 2025 | 2026 | 2031 |
|---|---|---|---|
| End-of-life flows | 4,472.7 | 4,605.7 (4,261.2–4,952.5) | 5,279.3 (4,899.6–5,673.7) |
| Potential collection | 3,433.5 | 3,567.7 (3,259.7–3,876.2) | 4,208.0 (3,835.7–4,575.6) |
| Collection with a buyer | 3,433.5 | 3,454.1 (3,104.3–3,798.1) | 3,590.7 (3,181.6–4,017.7) |
| New scrap: market | 1,802.0 | 1,809.7 (1,573.1–2,042.4) | 1,829.1 (1,574.0–2,085.1) |
| Supply with a buyer | 5,235.5 | 5,263.8 (4,812.3–5,701.2) | 5,419.7 (4,876.4–5,959.2) |
| Base demand | 4,714.0 | 4,837.4 (4,386.4–5,291.9) | 4,942.3 (4,267.7–5,592.6) |
| Use | 4,714.0 | 4,771.5 (4,317.6–5,215.6) | 5,164.2 (4,576.2–5,730.5) |
| Extra-EU exports | 1,111.5 | 1,127.0 (1,068.3–1,185.4) | 949.5 (652.7–1,277.0) |
| Extra-EU imports | 590.0 | 634.7 (608.4–662.4) | 693.9 (574.7–820.0) |
| Net extra-EU exports | 521.5 | 492.3 (432.4–550.0) | 255.6 (−42.9 to 585.9) |
| Usable Al, kt usable Al | 4,336.1 | 4,400.3 (3,975.7–4,813.1) | 4,730.0 (4,197.3–5,263.8) |

End-of-life flows, potential collection and new scrap: 80% interval. Other rows: 80% spread of draws under base assumptions, conditional on the assumed ratio bound, not a confidence interval. The jump in use in 2027 is mostly a result of absorption assumptions: intake above demand and dispersive uses enter only with tonnage with no buyer. Collection with a buyer in 2026–2028 has kinks because a tonne in stock counts in the year of sale.

**Table 26.** Main member states: collection with a buyer, split in proportion to potential collection, indicative (kt Al)

| Member state | 2025 | 2031 | Annual rate 2025–2031, median of draws, % |
|---|---|---|---|
| Germany | 741.7 | 720.7 (611.7–840.1) | −0.4 |
| France | 517.8 | 543.7 (470.6–621.2) | 0.8 |
| Italy | 435.1 | 449.9 (378.4–528.0) | 0.6 |
| Poland | 257.0 | 303.2 (262.3–346.6) | 2.8 |
| Spain | 239.3 | 249.4 (209.4–291.1) | 0.7 |
| Netherlands | 174.8 | 175.5 (139.4–216.0) | 0.1 |
| Belgium | 159.1 | 168.3 (131.0–210.5) | 1.0 |
| Sweden | 117.5 | 117.3 (96.9–138.9) | 0.0 |
| Austria | 84.2 | 107.0 (90.9–124.6) | 4.1 |
| Finland | 80.5 | 87.7 (71.8–104.9) | 1.5 |

The gap is not computed for member states or fractions: the class assignment follows from value weights, so collection with a buyer is split in proportion to potential collection.

Change in collection 2025→2031: 3,433.5 → 3,590.7 kt Al; end-of-life volume 584.1, fraction mix 22.9, collection rates −449.9. Net exports fall from 521.5 kt Al in 2025 to 255.6 kt Al in 2031 (2,000 draws). Without Art. 40 the model gives 1,141.2 kt Al, which is 906.6 above S0 on the same 1,000 draws (234.6); without projects in the R04 register 349.0. For the assumptions and the unverified mechanism see the warnings. Model result, not causal identification.

## 17. Discussion {#r17}

R04 differs from R03 by its three balances, quality classes and a linear programme in each draw. The main quantitative conclusion on net exports depends on legal assumptions and end-of-waste status, not on a signal in the data. Four items are the most important assumptions: the end-of-waste share, the limit and its growth, the intake of blocked volume and the AlSiCu alloy outlet. The value of the model lies in an explicit comparison of legal and quality variants, not in a point figure. With no metallurgist review the split into closed loop and downcycling is indicative, and the class level is not published.

## 18. Limitations {#r18}

- Alloy families, routes, classes, m_i, r_i and class values are R04 assumptions; with no metallurgist review the result for recycling within the same family, the strict loop and downcycling is indicative.
- Class composition: ranges from the R04 matrix, peer-reviewed measurement only for Twitch fractions (one batch, USA). Limits are approximations from public sources and margin m is an assumption. Class composition is constant over time, so the model does not forecast stream quality; coatings, Fe from joints, Cr, Ti, Sn and Pb are not tracked.
- Model inflow in 2024 (8,127) is below the balance (9,570 kt Al): a constant scale does not reconcile the inflow history with the end-of-life anchors.
- Capacities from the R04 register: coverage below 70% for extruders, sorting plants and die-casting foundries; the scrap share of nominal capacity is an R04 assumption with a wide range.
- Historical use is an identity (collection + new scrap − net exports, no stock change). In small states the mean is sometimes negative (CY, EE, HR, IE, LT, LV, SK in some years), and the country forecast uses shares, so country values jump at 2025/2026 (e.g. LU, SI, MT). Country values to 2025 are not a measurement of use.
- Intra-EU trade: mirror asymmetries are split by turnover; country balances are approximate.
- The LP works at EU-27 level, with no logistics; objective weights and class values set an order, not prices. The model determines neither prices nor quality premiums. The split of scrap by technology is an input, not a result.
- LP gross trade mass differs from Comext: in 2025 exports 1,314.7 vs 1,277.3 kt gross, because LP classes have a_i different from the CN-code coefficients k.
- Comext CN 7602 does not separate new and post-consumer scrap, nor waste and end-of-waste status; the class split of trade is an assumption, and exports are published only in total.
- Market closure (response of demand, collection, imports and the alloy outlet to the price ratio, the lower ratio bound, the temporary stock ceiling, the final route) consists of R04 assumptions with wide distributions; they have no verification in data, because Art. 40 has not yet operated.
- Hibernation has no national measurement; construction lifetimes are 31.5–60 years; the inflow history of Central and Eastern Europe before 1991 is weak. The ELV stream lies at the upper edge of the bottom-up check, because some vehicles of unknown fate are exported.
- For origins before 2024 the new scrap base and coefficients contain information from the 2024 balances; data are the latest Eurostat vintage, not archived vintages.
- W062 covers non-ferrous metals together, not aluminium alone; a simple stock model has a smaller error here than the model, which casts doubt on the growth of collection in the CD, IW and MSW fractions.
- The light packaging class has only a parameter splitting rigid packaging and laminates, with no separate composition; there are no classes for cables, 2xxx and 7xxx or profiles with thermal breaks; tolling is not a separate row.
- The only control was a multi-perspective internal check (AI agent runs). There was no review by an aluminium metallurgist or a statistician.
- Counterfactuals and scenarios are model results, not causal identification.

Caveats in full; the forecast page shows their short versions:

- Net exports after 2028 follow from assumptions, not from data. Art. 40 of Regulation (EU) 2024/1157 is adopted law applying from 21 May 2027, so its operation is not in the historical record. R04 assumptions are: the size of permitted exports and its growth, the share of scrap that has reached end-of-waste status (Regulation (EU) No 333/2011), and the intake of blocked tonnage by EU industry. Limit variants give 2031 net exports of 19–453 kt Al, and without Art. 40 about 1,141 vs 235 in the base scenario (same 1,000 draws; with 2,000 draws the base scenario gives 256). Of the difference of 907 kt Al, a Shapley split gives the “level” of permitted exports 445, “no growth” 259 and 202 outside the limit. The same assumptions move use and usable Al: without Art. 40 use is 4,875 vs 5,152 kt Al, and usable Al 4,493 vs 4,722 kt usable Al. These are conditional model results, not causal identification.
- The −43–586 spread does not cover legal uncertainty. It was computed under the base scenario assumptions and the assumed ratio bound; it is not a confidence interval. In hypothetical scenarios mean 2031 net exports range from −131 kt Al (tightened end-of-waste status) to 477 kt Al (Art. 41 list update), and the intervals of the limit variants extend from −161 to 911 kt Al. A response of all imports to the ratio gives 326 vs 235 kt Al.
- In 98% of 2031 draws the market does not close by price: the result is a boundary solution. The model looks for a fall in the scrap to primary metal price ratio at which all tonnage has a buyer. In 98% of 2031 draws no such fall exists, so the change in the ratio after 2027 is an assumed bound outside the 2005–2025 history, not a market result. This does not follow from the bound alone: with the bound practically removed the market closes in only 10% of draws, because it is limited by chemistry, the alloy outlet and a supply that hardly responds to price. Collection with a buyer, use and the gap depend on assumptions about routes and the alloy outlet, most of all on the split of the refiner outlet into copper-bearing and copper-free alloys: without this split the gap is 384 kt Al. Use in 2031 is base demand 4,942 + intake above demand 222 + dispersive uses 103 − unfilled 103 kt Al; the jump from 4,772 in 2026 to 5,011 in 2027 is mostly a result of absorption assumptions.
- The outlet gap of 562 kt Al in 2031 is a result of assumptions, not a market forecast. It is the annual tonnage of potential collection with no buyer: the model settles it into a final route (incineration of laminates with partial recovery), lower collection and a temporary stock, which is bounded and returns to supply. The gap is 562 (80% spread of draws under base assumptions: 191–931), and collection falls from potential collection of 4,208 to collection with a buyer of 3,591 kt Al, also through a price-driven drop of 55. Route and assumption variants: 170–879 kt Al. On the EU outlet side the largest influence comes from outlet routes: on the same 1,000 draws 170–770 vs 589 kt Al. The end-of-waste condition for classes with no chance of that status changes the composition, hardly the tonnage (582 without it), but the assumption on the share of end-of-waste status itself moves the gap: S5a 736, S5c 458, S5d 845. The gap is not given by class, because the class assignment follows from value weights.
- The test on history does not check the 2028–2031 mechanism. It concerns increments before Art. 40. It verifies neither levels nor what happens to tonnage after the ban, because the provision has not yet operated.
- The advantage over the naive forecast is not confirmed. The naive forecast is the last known value with a distribution of historical changes. In the test on history (5 forecast origins, 2014–2022) the model was on average 26.7% more accurate for net exports, but the difference is not statistically significant (p = 0.25). For the 2014 origin the model was clearly worse than the naive forecast (165.2 vs 95.6 kt Al of error). It also did worse at 3 of 11 horizons: 2, 4 and 8 years ahead.
- For some series the model is weak. Worse than the naive forecast: end-of-life vehicles (ELV) by 14.8%. Losing to the best simple benchmark: packaging by 3.0%, end-of-life vehicles (ELV) by 33.5%, W062 non-ferrous metal waste by 128.6%. Intervals too narrow, coverage of the 80% interval: packaging 0.20, end-of-life vehicles (ELV) 0.48.
- The collection level rests on 2024 balances of weak quality. Collection is not measured directly. The model is tied to the 2024 aluminium balances of the EU-27 and five countries, which are estimates, not measurements. Without them the model understated the level in the test on history: 2024 post-consumer scrap by 24–32% (2,517–2,813 vs 3,714 kt Al), and the deviation is −1.4 to −2.5 standard deviations for end-of-life flows and −1.9 to −2.7 for post-consumer scrap. The main forecast depends on these balances.
- The Art. 41 country list is a draft delegated act. It was not adopted as of 31 October 2026; the adoption deadline is 21 November 2026. After adoption the result must be recalculated. The EU trade measure for aluminium scrap (S5b) is a hypothesis only: according to the European Aluminium letter of 7 September 2026, a stakeholder position, the draft duty was at least temporarily withdrawn on 4 September 2026, and the Commission has published no document.
- Usable Al, recycling within the same family and downcycling rest on quality assumptions. Scrap classes, technology routes and physical class parameters are R04 assumptions to be confirmed, and alloy compositions and limits are approximations from public sources, not compared with the text of the standards. The split into recycling within the same family and downcycling is indicative. Primary metal for dilution in the LP allocation depends almost entirely on the assumed cost in the LP objective: 63–1,739 kt across cost variants. This is not primary metal use or a requirement.
- There was no review by a metallurgist or a statistician. Control is a multi-perspective internal check (AI agent runs). Before use in trading or investment decisions a review by a metallurgist and a statistician is needed.
- Some control computations are indicative. They did not reach full convergence in 3 of the 5 tests on history and in 8 of the 13 variants of use history and lifetimes. The main run meets the requirement.

## 19. Conclusions {#r19}

In the main result EU-27 collection rises to 3,590.7 kt Al in 2031 and technologically usable Al, that is Al in remelted metal, to 4,730.0 kt usable Al. Net exports of 255.6 kt Al in 2031 should be read as an assumption-dependent result. After adoption of the Art. 41 list the result must be recalculated. Before use in trading or investment decisions a review by a metallurgist and a statistician is needed.

## 20. Reproducibility, control and notes on R03 collected after publication {#r20}

All numbers come from deterministic code with fixed seed 20261030. Output files were checked by recomputing from the same fit: 10,003 rows agree to rounding, max difference 0.0005 kt. Control is a multi-perspective internal check (AI agent runs). There was no review by a metallurgist or a statistician, and the model has not been replicated by an outside team.

**Table 27.** Notes on R03 collected after publication: state in R04

| Priority | Recommendation | State | Place | Note |
|---|---|---|---|---|
| P0 | Single system boundary (for Al: CN 7602 with bridges) | partial | note ch. 2; file section system_boundary | One definition table for trade, use and plants; bridges 2620 40 and 7601 20 outside the scrap balance. Remaining: tolling may be on both sides of the boundary, semi-finished scrap may include part of the internal loop, metal from dross has no separate flow, LP gross mass differs from Comext. |
| P0 | Better 2026 nowcast and reconciliation of the balance gap | partial | note ch. 10; file section nowcast | The gap is split explicitly between stocks and use. The August bias correction was added after seeing the results, k = 6 is an assumption, the AI backtest uses one data vintage. |
| P0 | Art. 41 update after the act is published | not done | note ch. 9 and 18; warnings on the page | The act was not adopted as of 31 October 2026, deadline 21 November 2026. The list in the model has status PR; after adoption it must be recalculated. |
| P0 | Repeatable structural validation without leakage | partial | note ch. 13; file section diagnostics | Full rolling origin: NUTS refit for 5 forecast origins. Remaining: 3 of 5 fits with R̂ above 1.01, latest-vintage data, noise form set after three origins, packaging damping and import drift chosen on the test, 2024 balance information in the new scrap base. |
| P1 | New scrap revision: explicit bridge between measurements | partial | note ch. 7; file section new_scrap | Three values, a scale on gross generation and two measurements in the likelihood. The measurements do not fully reconcile, and the gross anchor is a control calculation. |
| P1 | Revision of low-quality fractions | partial | note ch. 8; file section quality | Low-quality classes in the LP, end-of-waste condition for classes with no chance of that status. Classes, parameters m_i and r_i and routes are assumptions with no metallurgist review; the light packaging class has only a parameter splitting rigid packaging and laminates. |
| P1 | Verification of interval coverage | partial | note ch. 12 and 13 | 80% coverage of net exports is 0.86 on 5 origins; intervals too narrow: packaging, end-of-life vehicles (ELV). Cluster calibration is not applied: after 2027 the regime is new. |
| P1 | Uncertainty of common shocks | partial | note ch. 12 | Process noise has common and country factors from residuals; the process/measurement split is an assumption. |
| P1 | Regime changes | not done | note ch. 13 and 18; warnings on the page | History assesses the model before Art. 40, so the blocking mechanism has not been tested against data. |
| Al | Oxide losses and slag fractions (ch. 14.2) | partial | note ch. 8 | Losses are computed as Σ a(1 − r)x, m_i and r_i are assumptions to be confirmed. Dross and slag are outside the LP, and the comparison with the register is an indicative check. |
| Al | Country profiles of foundries and secondary alloys (ch. 14.2) | not done | note ch. 8 and 18 | The LP works at EU-27 level, and the register has no capacities of die-casting foundries. |
| Al | Costs of impurity removal and quality premium (ch. 15) | not done | note ch. 8 and 9 | Class values set an order in the allocation. Mg removal is explicit, but the model determines neither prices nor quality premiums. |
| Al | Cost of bringing scrap to commercial quality (ch. 16.2) | not done | note ch. 18 | Sorting costs are an assumption, with no plant data. |
| Al | Metallurgical control (ch. 18.1, 18.4) | not done | note ch. 20 | There was no review by a metallurgist or a statistician. The only control was a multi-perspective internal check. |

## References {#bibliografia}

- Passarini F., Ciacci L., Nuss P., Manfredi S. (2018). Material Flow Analysis of Aluminium, Copper, and Iron in the EU-28. EUR 29220 EN, Publications Office of the European Union, Luxembourg (JRC111643). https://doi.org/10.2760/1079
- Cullen J.M., Allwood J.M. (2013). Mapping the Global Flow of Aluminum: From Liquid Aluminum to End-Use Goods. Environmental Science & Technology 47(7), 3057–3064. https://doi.org/10.1021/es304256s
- Held M., Rosat N., Georges G., Pengg H., Boulouchos K. (2021). Lifespans of passenger cars in Europe: empirical modelling of fleet turnover dynamics. European Transport Research Review 13, art. 9. https://doi.org/10.1186/s12544-020-00464-0
- Hatayama H., Daigo I., Matsuno Y., Adachi Y. (2009). Assessment of the Recycling Potential of Aluminum in Japan, the United States, Europe and China. Materials Transactions 50(3), 650–656. https://doi.org/10.2320/matertrans.MRA2008337
- Melo M.T. (1999). Statistical analysis of metal scrap generation: the case of aluminium in Germany. Resources, Conservation and Recycling 26(2), 91–113. https://doi.org/10.1016/S0921-3449(98)00077-9
- Buchner H., Laner D., Rechberger H., Fellner J. (2015). Dynamic Material Flow Modeling: An Effort to Calibrate and Validate Aluminum Stocks and Flows in Austria. Environmental Science & Technology 49(9), 5546–5554. https://doi.org/10.1021/acs.est.5b00408
- Liu G., Bangs C.E., Müller D.B. (2013). Stock dynamics and emission pathways of the global aluminium cycle. Nature Climate Change 3, 338–342. https://doi.org/10.1038/nclimate1698
- Langhorst M., Billy R.G., Schwotzer C., Kaiser F., Müller D.B. (2024). Inertia of Technology Stocks: A Technology-Explicit Model for the Transition toward a Low-Carbon Global Aluminum Cycle. Environmental Science & Technology 58(22), 9624–9635. https://doi.org/10.1021/acs.est.4c00976
- European Aluminium (2018). Environmental Profile Report. Life-Cycle inventory data for aluminium production and transformation processes in Europe. Brussels. https://european-aluminium.eu/wp-content/uploads/2023/01/European-Aluminium_Environmental-Profile-Report-2018_full-version.pdf
- European Aluminium, Ramboll (2023). Net-Zero by 2050: Science-based Decarbonisation Pathways for the European Aluminium Industry. Full report. https://european-aluminium.eu/wp-content/uploads/2023/11/23-11-14-Net-Zero-by-2050-Science-based-Decarbonisation-Pathways-for-the-European-Aluminium-Industry_FULL-REPORT.pdf
- Ducker Carlisle (2023). Aluminum Content in Passenger Vehicles (Europe). Public Summary, April 2023. https://european-aluminium.eu/wp-content/uploads/2023/05/23-05-02Aluminum-Content-in-Cars_Public-Summary.pdf
- Daigo I., Iwata K., Ohkata I., Goto Y. (2015). Macroscopic Evidence for the Hibernating Behavior of Materials Stock. Environmental Science & Technology 49(14), 8691–8696. https://doi.org/10.1021/acs.est.5b01164
- Boin U.M.J., Bertram M. (2005). Melting standardized aluminum scrap: A mass balance model for Europe. JOM 57(8), 26–33. https://doi.org/10.1007/s11837-005-0164-4
- Gaustad G., Olivetti E., Kirchain R. (2012). Improving aluminum recycling: A survey of sorting and impurity removal technologies. Resources, Conservation and Recycling 58, 79–87. https://doi.org/10.1016/j.resconrec.2011.10.010
- Raabe D. et al. (2022). Making sustainable aluminum by recycling scrap: The science of "dirty" alloys. Progress in Materials Science 128, 100947. https://doi.org/10.1016/j.pmatsci.2022.100947
- Shankar, Söderhjelm, Apelian (2025). J. Sustain. Metall. 11, 472–484. https://doi.org/10.1007/s40831-024-00989-x
- Akram, Holthe, Ringen (2026). JOM 78, 6113–6127. https://doi.org/10.1007/s11837-026-08309-2
- Werheit et al. (2011). J. Anal. At. Spectrom. 26, 2166. https://doi.org/10.1039/c1ja10096c
- Williams et al. (2023). Sensors 23, 7837. https://doi.org/10.3390/s23187837
- The Aluminum Association, International Alloy Designations and Chemical Composition Limits for Wrought Aluminum and Wrought Aluminum Alloys, August 2018 https://www.aluminum.org/sites/default/files/2021-11/TealSheet.pdf
- The Aluminum Association, Designations and Chemical Composition Limits for Aluminum Alloys in the Form of Castings and Ingot, August 2025 https://www.aluminum.org/sites/default/files/2025-10/Pink%20Sheets%202025.pdf
- Hoffman M.D., Gelman A. (2014) The No-U-Turn Sampler. Journal of Machine Learning Research 15, 1593–1623. https://doi.org/10.48550/arXiv.1111.4246
- Vehtari A., Gelman A., Simpson D., Carpenter B., Bürkner P.-C. (2021) Rank-Normalization, Folding, and Localization: An Improved R̂. Bayesian Analysis 16(2), 667–718. https://doi.org/10.1214/20-BA1221
- Gneiting T., Raftery A.E. (2007) Strictly Proper Scoring Rules, Prediction, and Estimation. JASA 102(477), 359–378. https://doi.org/10.1198/016214506000001437
- Diebold F.X., Mariano R.S. (1995) Comparing Predictive Accuracy. Journal of Business & Economic Statistics 13(3), 253–263. https://doi.org/10.1080/07350015.1995.10524599
- Grömping U. (2006) Relative Importance for Linear Regression in R: The Package relaimpo. Journal of Statistical Software 17(1). https://doi.org/10.18637/jss.v017.i01
- Ansari A.F., Stella L., Turkmen C. et al. (2024) Chronos: Learning the Language of Time Series. arXiv. https://doi.org/10.48550/arXiv.2403.07815
- Das A., Kong W., Sen R., Zhou Y. (2024) A decoder-only foundation model for time-series forecasting. PMLR 235. https://doi.org/10.48550/arXiv.2310.10688
- Ke G., Meng Q., Finley T. et al. (2017) LightGBM: A Highly Efficient Gradient Boosting Decision Tree. NeurIPS 30. https://proceedings.neurips.cc/paper/2017/hash/6449f44a102fde848669bdd9eb6b76fa-Abstract.html