Methodological noteCopper scrap market in the European Union member states 2026–2031

1. Introduction and research questions

The R03 forecast describes the copper scrap market in the 27 member states of the European Union in 2026–2031: post-consumer scrap collection, new (post-industrial) scrap, the grade structure by CN 7404 subheading, scrap use by smelters, refineries and processors, and net exports outside the EU. The measure is copper content (kt Cu); gross mass of grades is reported separately. The base year is 2025, 2026 is a nowcast from monthly data, and 2027–2031 is a forecast expressed as a distribution of 10,000 draws. Release 1.1 takes into account independent reviews of data, mathematics, industry, the portal and the text (ch. 20).

Research questions:

• Q1 – how much post-consumer scrap EU member states will collect in 2026–2031 and with what uncertainty;
• Q2 – how much new (post-industrial) scrap copper semis production will generate (link to the R02 forecast);
• Q3 – what the grade structure (CN 7404 00 10, 7404 00 91, 7404 00 99) and the average Cu content of supply will be;
• Q4 – how supply will be split between smelters and refineries, processors melting scrap directly, and net exports outside the EU;
• Q5 – how legal acts (Regulation 2024/1157, CRMA, Regulation 2026/1738, the WEEE Directive) and technologies (sensor-based sorting, digital product passport) change the result.

Principles of the method:

• physics before learning – the stock model and the mass balance are hard constraints: collection does not exceed end-of-life products, grades add up to supply, and supply equals use, net exports and stock change in every draw; machine learning works only on observable series (CN 7404 trade, WEEE) – in this version there is no residual series of the physical model to learn (ch. 11), and AI models affect only the 2026 nowcast;
• full posterior distribution (NUTS sampler) with convergence diagnostics (R̂, ESS, divergences) and posterior predictive checks;
• probabilistic evaluation – CRPS, quantile loss, PIT histogram, interval coverage, Diebold–Mariano test; an AI model enters the ensemble only if it meets the inclusion rule: more accurate than the benchmark, or at most 2% worse with better interval coverage;
• explainability and sensitivity – SHAP for learning models, LMG (parameter-group shares in the R² of a linear regression) for the stock model, re-estimation under alternative priors and a structural variant of the inflow scale;
• quality, not only quantity – grades, Cu content and impurity risk by fraction;
• primary sources with reliability class A/B/C, no exchange quotations, price agencies or press as sources of figures; reproducibility (random seeds, library versions, SHA-256 checksums of data).

Main results. EU-27 post-consumer scrap collection declines slightly: 1,454.9 kt Cu in 2025 and 1,426.3 (1,319.7–1,533.6) kt Cu in 2031 (median and 80% interval; average change −0.3% per year). The growth of collection in release 1.0 came from an unestimated ramp of the stock inflow scale in 2000–2020; this variant is shown only as a structural variant (+7.0% collection by 2031, ch. 14). New scrap is 595.8 (473.8–745.2) kt Cu, total supply 2,028.7 (1,856.6–2,207.0) kt Cu. EU scrap use rises to 2,074.4 (1,770.8–2,412.7) kt Cu (smelters and refineries 997.5, processors 1,066.8; medians). Net exports outside the EU fall from 98.3 kt Cu in 2025 to −52.8 (−305.1 – 192.7) kt Cu in 2031. The average fall of 152.6 kt Cu has two main causes: Art. 40 of Regulation 2024/1157 (ban on waste exports to non-OECD countries from 21 May 2027, in the base scenario) – −96.5 kt, and higher smelter scrap throughput from the D4 projects – −84.9 kt; other factors (processor demand, use level, collection, new scrap) add +28.8 kt. Without Art. 40, net exports in 2031 would be 45.1 kt Cu (median). Net imports occur in 35–61% of the 2027–2031 draws – as the balance residual, without a price mechanism (ch. 16). The collection rate (excluding the dormant stock) in 2031 is 65.7%, and the share of grade 7404 00 10 in supply Cu is 53.9%.

The note is structured as a research article: definitions and symbols (ch. 2), literature review (3), data (4), method – stock model, collection rate, new scrap, grades, market balance, nowcast, AI ensemble and uncertainty (5–12), validation and sensitivity (13–14), scenarios and results (15–16), discussion (17), limitations (18), conclusions (19), reproducibility (20) and bibliography. The model architecture diagram on the forecast page links each block to the relevant chapter.

2. Definitions and symbols

• collected post-consumer scrap – copper in scrap from end-of-life products that entered trade (collection);
• new (post-industrial) scrap – scrap from product makers sold on the market, excluding run-around scrap returned within the plant and losses;
• end-of-life products – copper in products that leave the in-use stock in a given year and enter the waste stream;
• collection rate – collection / end-of-life products (excluding the dormant stock – cables in the ground and installations taken out of use); it differs from the end-of-life recycling rate in balances, which has the dormant stock in the denominator;
• apparent use – production + imports − exports of refined copper and semis (stock inflow);
• scrap use – scrap melted by smelters, refineries and processors in the EU (excluding run-around scrap);
• net exports outside the EU – exports − imports of CN 7404 scrap in trade with non-EU countries, in Cu content;
• national market balance – supply (collection + new scrap) − domestic use = scrap trade balance with all partners.

Waste fraction abbreviations (as in the JRC report for the EU-28): C&D – construction and demolition waste, IW – industrial waste (cables, machinery), ELV – end-of-life vehicles, MSW – municipal solid waste, WEEE – waste electrical and electronic equipment.

3. Literature review

The review covers 40 peer-reviewed items with a DOI, checked in the publishers' registry; the descriptions below paraphrase their contribution and show how it was carried into the model. Works whose content was verified only from publisher metadata or from their citation in the JRC 2018 report are marked.

Dynamic stock and flow models of copper

The basis is the formulation in which the flow of end-of-life products is a convolution of historical inflow to use with a lifetime distribution (Elshkaki 2005; Müller 2006 – from publisher metadata). Reviews of dynamic material flow analysis show common principles and large differences in extrapolation (Müller et al. 2014), and open software frameworks organise the product–material–element data structure (Pauliuk, Heeren 2020). For copper, global dynamic models identified lifetimes as the main source of uncertainty in recycling indicators (Glöser et al. 2013), and European analyses the size of the urban mine (Ciacci et al. 2017) and the share of new scrap in fabrication (Soulier et al. 2018, via JRC 2018). Longer-horizon EU forecasts (Ciacci et al. 2020) and cohort models with Monte Carlo (Born, Ciftci 2024) are a point of reference; R03 adds a resolution of 27 member states, Bayesian calibration to balances and waste statistics, and a quality layer.

Product lifetimes

Databases and reviews of lifetime distributions (Murakami et al. 2010; Oguchi et al. 2010) justify the Weibull distribution (Weibull 1951) and the cohort age definition. Sectoral distributions for metals (Yokoi et al. 2022a; Elshkaki et al. 2018) and the normal distributions of the JRC report for the EU-28 (Passarini et al. 2018) differ for infrastructure and consumer goods – hence a prior covering both sources (ch. 5) and variants centred on each of them (ch. 14). The analysis of copper stock by final product in Japan (Yokoi et al. 2022b) is a template for the member state × sector × product breakdown.

Collection efficiency and scrap quality

Collection and sorting rates by waste stream for Western Europe (Ruhrberg 2006, via JRC 2018) form the prior of fraction rates. Reviews of metal recycling challenges (Reck, Graedel 2012; Loibl, Tercero Espinoza 2021) point to the growing role of complex post-consumer scrap, the difficulty of removing lead and the low technological readiness of sensor-based copper sorting. EU WEEE analyses show large copper losses outside formal collection (Wagner et al. 2023; Pauliks et al. 2026), and multivariate reconciliation of sales, stock and lifetimes improves WEEE estimates (Wang et al. 2013). Copper lost in steel scrap (Daehn et al. 2017; Panasiuk et al. 2022) is a quality loss outside CN 7404 – hence the importance of the Cu limit in the steel fraction in the end-of-life vehicles regulation.

Bayesian inference and uncertainty in flow analysis

Probabilistic flow models with uncertainty propagation (Gottschalk et al. 2010 – from publisher metadata) and systematic frameworks for uncertainty assessment depending on data quality (Laner et al. 2014) precede Bayesian flow analysis: incremental updating (Lupton, Allwood 2018), learning data noise together with prior elicitation (Dong et al. 2023), process hierarchies with balance relaxation and posterior predictive checks (Wang et al. 2024), and system quantification with incomplete data (Mason et al. 2025). Differences in transformation rates between similar economies (Graedel et al. 2004) justify a member-state hierarchy instead of a single EU parameter.

Machine learning, sensors and the market

Machine learning predicts waste volumes from socio-economic data (Kannangara et al. 2018); in R03 it forecasts observable series (2026 nowcast) and serves checks, and does not replace the mass balance. A systematic review of optical sensors with machine learning (Kroell et al. 2022) shows a predominance of laboratory-scale studies – hence a sorting scenario acting mainly through quality. Economic models with a price mechanism (Diersen et al. 2025) justify treating relative prices as a latent factor. Chinese demand and scrap import policy (Zhang et al. 2015; Wang et al. 2017; Dong et al. 2019; Dong et al. 2020) explain the destination structure of EU exports, and reviews of demand forecasts (Schipper et al. 2018; Kuipers et al. 2018; Yoshimura, Matsuno 2018; Watari et al. 2021) give the long-term background and the distinction between copper and alloys.

4. Data and sources

Source classes: A – official data and primary documents (Eurostat, EU legal acts, company annual and sustainability reports, peer-reviewed literature); B – calculations on class A data (conversion to Cu, ai-waw.com flow balances, company websites and releases); C – model assumptions and results. The forecast file contains 102 described sources (without repeating the same work). Trade mass is converted to copper content with grade coefficients:

where c for 7404 00 10 is 0.90, and in exports to China and Hong Kong 0.965 (the Chinese standard GB/T 38471-2023 admits for import raw material with at least 97–98% Cu; the same coefficient is used in the portal's EU analysis), for 7404 00 91 – 0.61, for 7404 00 99 – 0.65 (standard deviations 0.04, 0.02, 0.03, 0.10). The grade coefficients come from 1989–2010 industry data cited in the 2011 JRC end-of-waste report – a secondary source (class B). Equation (4.1) also applies to imports and intra-EU trade.

Intra-EU trade is mirror-asymmetric (export and import declarations of the same country pair differ); the EU-27 balance is computed from trade with extra-EU partners, where the asymmetry does not arise. Inflow to use of the nine member states with long series comes from their apparent use; for the others – from the 2024 balance level, the core trend and relative GDP (a class C proxy). Aurubis data (the largest component of the smelter feed index) refer to October–September fiscal years, treated as calendar years; until August 2024 the primary smelter segment also included a plant in the USA.

5. In-use stock model

Copper enters the in-use stock of five sectors (construction, infrastructure, transport, industrial machinery, consumer and electrical products). Inflow of sector s in member state c:

where Ĩ is smoothed apparent use, ψ the inflow scale from the 2024 balance, κ a consistency correction of inflow against the end-of-life, collection and 2024 inflow anchors, constant over time for all cohorts and forecast years. The κ correction is not a measure of copper exported in finished goods (net importers of goods also have values below 1). Release 1.0 used a κ ramp in 2000–2020 (inflow of 2000–2020 cohorts raised relative to older ones) without an estimated mechanism – this variant is in ch. 14. Lifetime is Weibull-distributed, and end-of-life products are the convolution of inflow with the probability of retirement at age x:

In equation (5.3) ζ denotes hibernation (cables in the ground, installations taken out of use – infrastructure and construction), d – dissipation, and in equation (5.4) T – the sector → waste fraction transition matrix (JRC report for the EU-28). The member-state hierarchy applies to the inflow correction:

with σ_κ = 0.30 and s_τ = 0.25 (non-centred parameterisation); 2024 inflow from the balances is an observation of κ_c. Sector shares ω ~ Dirichlet(100 · ω⁰), ζ_inf ~ Beta(2; 6), ζ_bud ~ Beta(1; 9). The prior of log μ_s is centred on the geometric mean of the JRC and Yokoi et al. (2022a) lifetimes; the shape k lies between the Weibull equivalent of a normal distribution (k ≈ 3.7) and the shape from the literature; for consumer products – WEEE category distributions.

The posterior EU inflow correction is κ_EU = 0.99 (member states from 0.50 – Portugal to 3.14 – Bulgaria), hibernation ζ_inf = 0.48, ζ_bud = 0.19. With literature (JRC) parameters EU-27 end-of-life products in 2024 would be 3,513.9 kt Cu, in the stock (DMFA) model of the ai-waw.com ‘Copper stock’ analysis 3,141.0 kt, in the EU-27 balance 2,364.1 kt; the posterior fit (parameter means, before process noise) gives 2,294.7 kt. The Shapley decomposition of this difference (Table 4) attributes it mainly to hibernation and the inflow correction, and less to lifetimes. With κ constant over time, neither κ nor ζ changes the growth rate of end-of-life flows.

6. Collection rate and policy

Collection of fraction f in member state c is the product of end-of-life products in the fraction and a collection rate modelled on the logit scale:

After 2025 the trend in equation (6.2) is damped (φ = 0.8) so as not to extend the 2005–2025 improvement mechanically:

Annual observations (WEEE collected, end-of-life vehicles, W062 waste) are linked to collection through unknown scale factors with a Student distribution (ν = 4), robust to single outlying years; since release 1.1 the collection rates (end-of-life recycling rates) of the 12 national balances for 2024 with a standard deviation from their p05–p95 interval are also observed:

Priors: a_f ~ N(logit(JRC rate), 0.5²), country effects u, w, v ~ N(0, τ²), τ ~ HN(0.5), g ~ N(0; 0.3²), g_W ~ N(0; 0.5²); m_W has a median of 0.018 (copper in WEEE mass × share of recoverable fractions), m_V – 0.0085 (copper in a vehicle × sorting), log σ 0.35.

FI, GR, SE – balance variant without package D (circularity from JRC), standard deviation × 1.5.

National collection rates in 2025 include: Netherlands 0.79, Austria 0.78, Poland 0.75, Germany 0.64, France 0.62. The order of member states is close to the balances (Spearman rank correlation 0.81) but not the same: Poland is in position 5 in the model (balances: 9), Sweden 4 (7), Germany 9 (5); above the balances lie CZ, FR, GR, PL, SE, below – AT, BE, DE, ES, FI, IT, NL (ch. 18). The common trend g has a posterior mean of 0.298, the WEEE trend g_W – 0.654 (per decade, logit scale). Policy enters the model in two places. Adopted law is in the base scenario: Regulation (EU) 2026/1738 on end-of-life vehicles (removal of wiring harnesses, electric motors and large circuit boards before shredding from 1 September 2029 – a logit shift of the Cu recovery rate from vehicles; controls on exports of used vehicles from 1 September 2031) and Art. 40 of Regulation (EU) 2024/1157 (ch. 9). Targets missed in the data (WEEE 65%) and future measures (CRMA programmes, product passport, sorting) are named scenarios (ch. 15).

7. New scrap

New (post-industrial) scrap is the share of post-industrial losses in the production base: copper semis + wire rod + 0.67 × alloy semis (Cu content, R02 forecast – the same draws). The base is split between member states half by their share of semis production and half by their share of inflow to use (new scrap also arises at finished-goods makers).

Prior θ₀ ~ N(logit 0.08; 0.7²), τ_θ ~ HN(0.5), s_η = 0.05; the posterior EU share is 0.092. The EU anchor is the EU-27 balance (847.8 ± 87.3 kt Cu, likelihood without double counting: 12 national balances and the remaining 15 member states = EU-27 balance − sum of 12). The definition in all balances and in R03 is the same: post-industrial scrap sold on the market, without run-around scrap and losses; scrap used at wire-rod mills, in direct melting and in foundries is on the use side. For the EU-27 in 2024 the model gives 600.0 kt Cu (posterior median; z = −2.8 against the balance; in the published series, after nowcast weighting – 597.1 kt). The difference reflects the dispute between bottom-up and top-down balances: the 12 national balances (detailed models of mills and processors) give 483.0 kt in total (model 435.7), so the remaining 15 member states would have to supply 43–49% of EU new scrap with about 18% of the production base – the share hierarchy does not accept that. The EU-27 balance assumes a new-scrap share of about 25% of fabrication, consistent with empirical estimates (Soulier et al. 2018), not only an assumption. EU-27 scrap use in R03 is therefore lower than in the balance (2024: 1,918.9 vs 2,078 kt Cu).

FI, GR, SE – balance variant without package D, standard deviation × 1.5.

8. Grades and quality

The Cu content of each source fraction and of new scrap is split into the grades CN 7404 00 10 (copper), 7404 00 91 (brass) and 7404 00 99 (other alloys) by shares from a Dirichlet distribution; gross mass follows from the grade Cu content:

The Dirichlet concentration is α_π = 40. Prior shares are calibrated by weighting to the Cu structure of EU member-state exports in 2021–2025: data 0.431 / 0.225 / 0.344, prior 0.561 / 0.254 / 0.185, after weighting 0.542 / 0.253 / 0.204 (ESS 5,982). The weight of the observation is weak because exports are of lower grade than total supply – EU smelters and processors retain more grade 7404 00 10. Result: the 7404 00 10 share in EU-27 supply Cu is 54.1% in 2025 and 53.9% in 2031, the average Cu content per tonne of scrap 0.750 and 0.750. Grade Cu content is constant over time, so the quality index changes only through the grade and fraction structure (comparison with the literature use structure – ch. 17).

9. Market balance and trade

In 2005–2025 scrap use follows from the balance identity – supply minus net exports (Comext data in Cu). In the forecast, EU use consists of smelter demand (feed index from D4 company documents), processors (direct-melt scrap – R02), a level shift in 2021 (break in the D_H index: inclusion of the former Metallo plants in Aurubis from June 2020), a zero-mean AR(1) process and throughput projects; net exports are the balance residual:

Posterior means: h = 1.440, m = 1.339, δ = −310.4 kt, ρ_M = 0.482, σ_nx = 112.4 kt; priors h, m ~ logN(log(2024 balance / 2024 index); 0.25), σ_nx ~ HN(100), ρ_M ~ Beta(3; 2). The smelter feed index D_H is 733.1 kt Cu in 2024 and 732.5 kt in 2025 (Kovohuty anodes, refined in Brixlegg, counted once). Member states: use C_c = s_c (C_EU − ΔP_EU) + ΔP_c (s_c – the 2021–2025 share), national balance = Z_c + N_c − C_c; intra-EU trade sums to zero. The stock change ΔS is allocated to member states in proportion to collection: national trade balances add up to EU net exports, and market balances (supply − use) to net exports plus ΔS. Use is split between smelters and refineries and processing in every draw: the ε component is allocated in proportion to h·D_H and m·SD, and projects by type of user.

Art. 40 of Regulation (EU) 2024/1157 (from 21 May 2027) is adopted law and part of the base scenario: exports of scrap as waste to non-OECD countries are allowed only to countries and wastes on the Art. 41 list. The draft list (Ref. Ares(2026)8858290 of 18 September 2026, adoption by 21 November 2026) allows copper scrap explicitly (B1010, waste code 17 04 01) only to Serbia; China and Hong Kong are not on the list, and metals were rejected for India, Indonesia, Malaysia, Pakistan, Thailand and Vietnam. Non-ferrous fractions for Ukraine and cables for Moldova were left out as immaterial for CN 7404. The exception is end-of-waste scrap (Regulation 715/2013). Its volume share is a class C assumption by CN code: for China and Hong Kong (45.9% of Cu exports outside the EU in 2023–2025; 7404 00 10 – 181.4, 7404 00 91 – 250.9, 7404 00 99 – 146.7 kt Cu) it is U(80%; 95%), U(50%; 80%) and U(30%; 60%), for other non-OECD countries – U(5%; 25%). The blocked share b_t:

where s_k,g – the share of destination k and code g in Cu exports outside the EU in 2023–2025, e_k,g – the product-status share, W41 – countries on the list (Serbia), f_t – the part of 2027 after 21 May, X_ext – gross exports outside the EU, r_OECD – redirection to OECD countries U(30%; 60%) (with the receiving-facility audit requirement, Art. 44 of the regulation), a – additional EU use U(30%; 70%), l – spare capacity U(5%; 15%) of use. The rest of the blocked exports lowers the collection of low-quality fractions; in 2027 a part U(0; 50%) of the rest goes temporarily to stocks (half of that in 2028, then zero) – ΔS in equation (9.3) averages 9.0 kt Cu in 2027 and 6.6 kt Cu in 2028. Without a price mechanism the split between these channels is an assumption. Sensitivity of the effect to the product-status share for China and Hong Kong:

EU scrap exports in Cu content were 582.4 kt in 2015 and 461.6 kt in 2025; net exports (after deducting imports from outside the EU) were 98.3 kt Cu in 2025. The main destinations in 2023–2025 were: China (44.5%), India (15.2%), Thailand (6.4%), Pakistan (4.9%), Malaysia (4.4%); non-OECD countries took 83.6% of the Cu exported. Intra-EU flows are shown on the arc map (pairs ≥ 3 kt Cu) with the declaration asymmetry marked.

Scrap use by smelters and refineries (R03) corresponds in 2031 to 99.6% of secondary refined copper output in the R01 forecast, and processor use to 117.9% of R02 direct-melt scrap. The differences come from definitions: R01 counts all cathode output from secondary raw materials (including materials outside CN 7404, e.g. residues and slimes), R02 – scrap as the residual of the copper balance in semis and wire rod; processor use in R03 also includes foundries and alloy makers, which R02 does not cover. Both are order-of-magnitude consistency checks, not components of one sum.

10. 2026 nowcast

Comext data for January–August 2026 (extra-EU trade) give EU net exports in Cu of 115.9 kt vs 57.0 kt in the same period of 2025. Three estimates of 2026 were considered: a Kalman filter (local level) on the monthly balance – 155.5 ± 37.4 kt; a year-to-date ratio separately for exports and imports – 194.9 kt; a September–December monthly forecast with time-series foundation models (ch. 11) – exports 543.4 kt (90% interval: 506.8–571.1), imports 360.6 kt (329.8–395.9). The monthly variant is the observation: 182.7 ± 42.5 kt Cu; the standard deviation combines model uncertainty (28.0), data revision and the spread of the three methods. The range of estimates (155.5–194.9 kt) shows that the result depends on the assumption about imports in September–December; the value will be updated after the next Comext release.

Model draws are weighted by the observation according to equation (10.1) (f_T = 1.00) and resampled: the ESS is 3,279 out of 10,000, with 3,977 distinct draws. The median of 2026 net exports changes from 99.3 to 177.2 kt Cu, EU collection from 1,462.2 to 1,471.2 kt Cu. Higher net exports with unchanged supply lower EU use in 2026 by about 59 kt Cu vs 2025 (ch. 18). Waste statistics are annual with a publication lag, so they do not enter the 2026 nowcast.

11. AI model ensemble

The ensemble includes time-series foundation models used without fine-tuning (Chronos – Ansari et al. 2024; Chronos-2 – Ansari et al. 2025; TimesFM – Das et al. 2024) and a locally trained model (LightGBM – Ke et al. 2017); weights were downloaded only from the developers' official repositories, and computations are deterministic (fixed seed, CPU). The control models N-HiTS (Challu et al. 2023) and TFT (Lim et al. 2021) did not enter the ensemble.

The backtest uses a rolling origin (base years 2012–2020, horizons 1–6 years; monthly series – 1–6 months); a model sees only data up to the base year, and covariates with their publication lag. The measure is sCRPS – CRPS from 19 quantiles (Gneiting, Raftery 2007) divided by the MASE scale of the unit (Hyndman, Koehler 2006). Rule for including a model in the ensemble:

Conclusions from Table 13: on the member-state panel the ensemble improves CN 7404 trade accuracy by a few per cent relative to the random walk; on the EU-27 aggregate there is no improvement for exports and net exports (net exports outside the EU: 1.508 vs 1.438); for extra-EU imports and WEEE the ensemble is more accurate, but these series do not set the result. The advantage of foundation models is large on monthly data (exports outside the EU: sCRPS change −28.0%), which justifies their use in the 2026 nowcast. Rule (11.1) admits a model at most 2% worse with better coverage. Stacking with weights estimated on short series overfitted (large drift weight at early origins), so production weights are equal, except for sets with enough cases. LightGBM has an accurate median but too narrow intervals and did not enter the ensemble. Conformal calibration of the ensemble (quantile regression – Romano et al. 2019; widening only) changes little because panel coverage is close to nominal.

The trade balance history in the structural model is input data, so the residuals of the physical model on observable variables are identically zero and there is no residual series to model; the other model variables have only 2024 cross-sectional anchors. The AI models therefore affect only the 2026 nowcast. Explainability (SHAP – Lundberg, Lee 2017; LightGBM, net exports outside the EU) shows as the most important features: Series level at the origin – 31.2%; WEEE collected × Cu (level) – 12.4%; Construction output (level) – 7.8%; Forecast horizon – 6.5%; Series volatility (5 years) – 6.0% of the sum of mean |SHAP|. Covariates (construction, registrations, WEEE, Comext unit value) do not improve accuracy significantly, so the note attributes no predictive role to them.

12. Uncertainty: draws, copula, summation, calibration

The posterior comes from the NUTS sampler (Hoffman, Gelman 2014) in the PyMC library (Abril-Pla et al. 2023): 4 chains × 1,000 draws after tuning, time 27.4 min. The 10,000 forecast draws are sampled with replacement from 4,000 posterior draws (seed 20261028). Process noise of the collection logit is an AR(1) process (ρ_Z = 0.6, standard deviation = σ_W + 0.1) starting from zero in 2025, with innovations from a t copula across member states (Demarta, McNeil 2005): the correlation matrix from Kendall's τ of changes in the log of scrap turnover in 2006–2025, shrunk with weight 0.3, ν = 5, mean correlation 0.081, mean tail dependence 0.065.

Summation is bottom-up: the EU-27 is the sum of 27 member states in every draw, and the mass constraints are met by construction:

The aggregated EU model (without member-state heterogeneity) is computed only for information – its ratio to the sum of member states in 2031 (median) is 1.031. Net-export intervals are widened by conformal calibration on the backtest from 4 origins (n = 24; errors within an origin are strongly correlated, and leave-one-origin-out folds share target years): multipliers c₈₀ = 1.37 and c₉₀ = 1.36; coverage before calibration 0.62 and 0.75, after calibration 0.67 and 0.71 for 80% and 90% intervals – still below nominal. The growth of the multiplier with the horizon and the widening of 2026 are assumptions, not backtest results; domestic use preserves mass, and the physical constraints on net exports are applied after widening:

Mass checks in 10,000 draws: collection above end-of-life products – 0 cases, negative collection – 0; maximum deviation of the grade sum from supply 0.000122 kt, of the EU balance 0.000904 kt, of the sum of national trade balances from EU net exports 0.000458 kt, of national balances with stocks 0.000092 kt. The use constraint binds after widening in at most 0.82% of draws (2031). Member-state series add up to the EU-27 in every draw, so means add up, not medians.

13. Validation

The structural backtest has origins 2016, 2018, 2020, 2022. It is conditional: structural parameters (lifetimes, sector shares, inflow correction, hibernation) and observation scale factors are fixed at the posterior means of the main run, and NUTS calibrates on data up to the origin only the collection rates and their trends, new scrap and market parameters; no full refit on data before the origin was done. Waste panels are evaluated on increments log y_t − log y_t0 (measurement constants and anchor levels cancel; the naive benchmark is a zero increment), net exports – on the level in kt Cu. The results are no proof of an advantage over the naive forecast.

Waste panels – CRPS of log-mass increments; net exports – CRPS in kt Cu before conformal calibration.

The relative CRPS difference vs the naive forecast (conditional; Table 14) is 0.42 for WEEE, 0.28 for W062 waste and 0.26 for vehicles. For EU-27 net exports the model CRPS is 94.677 vs 92.596 kt Cu for the naive forecast (relative difference −0.02) – accuracy practically equal to the naive forecast; use in the backtest is computed as in the forecast (the 2021 level shift from the 2022 origin, a zero-mean AR(1) component). The Diebold–Mariano test (with the Harvey, Leybourne and Newbold correction) has low power with 4 origins. The PIT test rejects uniformity for: EU-27 net exports outside the EU (kt Cu), Non-ferrous metal waste W062, member states (log mass), WEEE collected, member states (log mass). Across all backtest cases the coverage of 50%, 80% and 90% intervals is 52.8%, 78.8% and 87.7% (the ‘Calibration: PIT histogram and interval coverage’ chart on the forecast page); for net exports coverage before calibration (0.62 and 0.75) was too low, hence the conformal calibration (ch. 12).

In the main run no parameter has R̂ > 1.01 or ESS < 400 (R̂ – Vehtari et al. 2021); the highest R̂ (1.007) belongs to the EU inflow correction (log κ_EU), the lowest ESS (896) to the common collection-rate trend. Tree depth hits the limit in 0.4% of transitions; the few divergences in the 2016 backtest fit do not affect the forecast.

FI, GR, SE – balance variant without package D, standard deviation × 1.5.

EU-27 collection in 2024 in the model (1,443.5 kt Cu) lies within ±1σ of the EU-27 anchor (1,355.1 ± 139.8 kt). The largest member-state deviation: France (model 227.1 vs 188.0 kt).

14. Sensitivity

The shares of parameter groups in the variance of EU-27 collection in 2031 (structural part, 2026–2031 inflow at the 2025 level) are computed with LMG – group shares in the R² of a linear regression of collection on the parameters, averaged over group orderings (Grömping 2006; relation to the Shapley value – Owen 2014), on the joint posterior (3,000 draws, full vectors of member-state parameters κ_c, u_c, w_c, v_c, g_c, gw_c): mean 1,457.1 kt Cu, standard deviation 62.1 kt, R² = 0.70. Posterior parameters are correlated, so the shares are computed on the joint distribution, not under independent marginals.

Collection rates explain the largest part of the variance (Table 18), followed by the member-state inflow correction and hibernation; the part not explained by the linear regression is 0.30. LMG shares are an additive measure without interactions, not Shapley effects of conditional variances. A structural variant of the inflow scale (a separate NUTS run) shows the source of the collection growth of release 1.0:

The growth of end-of-life flows in release 1.0 came from the inflow-scale ramp (effective inflow of 2000–2020 cohorts raised relative to older ones), not from stock dynamics supported by data; with a scale constant over time, end-of-life flows decline slightly. The ramp variant also differs from the base by lacking the 2024 inflow observation from the balances. Parameters whose posterior mean lies far from the prior (at least 1.5 prior standard deviations):

The data shift the spread of country trends (τ_g), the transport lifetime shape and infrastructure hibernation; results are sensitive to this within the range of the prior variants (table below).

2024 and 2031 columns – posterior fit, structural part (before the nowcast and process noise).

A variant is first assessed by importance weighting of the main-run draws with the ratio of prior densities (Pareto smoothing – Vehtari et al. 2024); when k̂ ≥ 0.7 (data dominate the prior), the variant is estimated with a separate NUTS run – such results are indicative. 2031 collection changes relative to the base variant by at most 39.6 kt Cu (54.5 kt between variants), less than the width of the 80% interval. Sensitivity to the product passport scenario assumption (increase of the WEEE collection rate by 2031): Δr = 0.00 → 0.0 kt; Δr = 0.02 → 5.4 kt; Δr = 0.05 → 13.3 kt; Δr = 0.08 → 20.8 kt.

15. Scenarios

Named scenarios use the same 10,000 draws as the base scenario (common random numbers), so the scenario − base difference has a small Monte Carlo error; the effect intensity is drawn, and the uncertainty of the effect size is part of the scenario. The base scenario contains adopted law (Regulation 2026/1738 and Art. 40 of Regulation 2024/1157); the counterfactuals each remove one of these acts. Channels: logit shift of collection rates (CRMA, WEEE), absolute rate increase (passport), quality and marginal recovery of shredder fraction losses (sorting), blocking part of exports (Art. 40, extended list), multipliers of end-of-life flows, new scrap and use (recession), project completion (smelter demand):

CRMA programmes and the WEEE target have the strongest effect on collection (Table 22). In the collection scenarios (WEEE target, sorting, passport, 2026/1738 counterfactual) EU use depends on supply only slightly – smelter and processor demand is exogenous in the model – so most of the change in supply goes to the trade balance; the change in net exports follows from this assumption and is not a separate result. Art. 40 lowers net exports: the counterfactual without Art. 40 gives a 2027–2031 difference of 71.8 / 105.7 / 100.4 / 98.7 / 96.5 kt Cu, and the extended list 58.5 kt Cu in 2031. The counterfactual without the D4 projects is computed only for the decomposition of the net-export change (ch. 16) and is not a named scenario. CRMA acts on collection more strongly than the WEEE target because it shifts the rates of three fractions (construction and demolition, industrial waste, WEEE) and the WEEE target of one; the upper bound of the CRMA effect is an assumption (an instrument without binding targets). The construction recession acts over time (collection difference 2027–2031: −30.2 / −39.6 / −26.0 / −16.4 / −9.6 kt Cu; net exports: −4.8 / 9.2 / 40.8 / 52.6 / 57.1 kt Cu): the effect on end-of-life flows fades out by 2031, while the drop in processor demand (ratio to base from R02) persists, so in 2031 the scenario changes mainly use and the trade balance. Sensor-based sorting acts mainly through quality (share of grade 7404 00 10).

16. Results

Net exports in 2024–2025 are Comext data in Cu (122.8 and 98.3 kt), and use in those years follows from the balance identity. In 2015–2025 net exports ranged from −76.6 to 208.5 kt Cu. The median path 2026–2031: 177.2 / 50.6 / −26.8 / −36.2 / −46.1 / −52.8 kt Cu. The mean change 2025 → 2031 (98.3 → −54.3 kt Cu) decomposes into: Art. 40 – −96.5, smelter throughput projects (D4) – −84.9, other factors – 28.8 kt Cu; without Art. 40 the 2031 median is 45.1 kt Cu. Share of draws with net imports in 2026–2031: 0 / 35 / 56 / 58 / 60 / 61% (without Art. 40: 0 / 22 / 36 / 39 / 42 / 43%). Net imports are the balance residual (σ_nx and use uncertainty around a median close to zero after Art. 40); the model has no price mechanism of competition between buyers, so this is not a market forecast.

The change in EU-27 collection from 1,454.6 to 1,426.3 kt Cu (means 2025 → 2031) decomposes (Shapley) into: end-of-life products (volume) −29.7 kt; waste fraction mix −2.7 kt; collection rates 4.2 kt. The decline of the end-of-life flow is offset only slightly by collection rates (trend from data, damped after 2025; Regulation 2026/1738 for vehicles). By fraction: construction and demolition waste −23.1 kt; industrial waste (cables, machinery) −27.9 kt; end-of-life vehicles 17.4 kt; municipal solid waste 0.1 kt; waste electrical and electronic equipment 5.3 kt. New scrap adds 12.9 kt, and supply changes from 2,045.4 to 2,030.0 kt. Collection per capita in the EU-27 is 3.23 kg Cu in 2025 and 3.16 kg Cu in 2031. In some member states collection grows despite the EU-27 decline (table below):

The growth of collection in Czechia (36.3 → 45.5 kt Cu) and Poland (111.4 → 122.6 kt Cu) comes mainly from growing end-of-life flows (Czechia 51.7 → 64.6 kt; Poland – end-of-life effect 6.0 and rate effect 5.3 kt): inflow to use in these member states grew after 1995, so the 2000–2015 cohorts reach retirement age. The inflow correction κ_c is constant over time and scales the level, not the growth rate. In the EU-27 the decline of end-of-life flows of older Western European cohorts offsets this growth.

17. Discussion – comparison with the literature

The R03 collection level (Table 26) is consistent with European values and higher than global ones, in line with higher EU collection rates. R03 new scrap (600.0 kt Cu in 2024) is lower than the value from the 25%-of-fabrication assumption (JRC 2018 report, via Soulier et al. 2018), because national balances count new scrap bottom-up (ch. 7). The share of grade 7404 00 10 in supply mass is lower than in the 2008–2010 use structure (ECI 2010, via JRC 2011), consistent with EU smelters and processors melting more high-grade scrap than supply alone provides. Longer-horizon EU forecasts (Ciacci et al. 2020) and cohort models (Born, Ciftci 2024) point to a growing post-consumer scrap flow in the coming decades; R03 does not confirm growth up to 2031 – with an inflow scale constant over time, the EU-27 end-of-life flow declines slightly and grows only in member states with a later build-up of stock (Czechia, Poland).

18. Limitations

The structural-model backtest is conditional: stock parameters (lifetimes, sector shares, inflow scale, hibernation) are fixed at the posterior means of the main run (data to 2025), without a full refit on data before the origin – the results are no proof of an advantage over the naive forecast. The relative CRPS difference vs the naive forecast (conditional, log-mass increments) is: WEEE 0.42, W062 non-ferrous metal waste 0.28, end-of-life vehicles 0.26. For EU-27 net exports outside the EU the model CRPS is 94.7 kt Cu vs 92.6 kt Cu for the naive forecast (relative difference −0.02) – accuracy practically equal to the naive forecast. The Diebold–Mariano test has low power with 4 origins. Net-export intervals were widened by conformal calibration from 4 origins (n = 24; errors within an origin are strongly correlated): multipliers 1.37 (80%) and 1.36 (90%); coverage after calibration (leave-one-origin-out) 0.67 and 0.71 – below nominal, so net-export intervals may still be too narrow; the growth of the multiplier with the horizon and the widening of 2026 are assumptions, not backtest results. The AI model ensemble (TimesFM, Chronos-Bolt, Chronos-2) is not more accurate than the naive forecast for the EU-27 aggregate (net exports: sCRPS 1.508 vs 1.438) and there is no residual series of the physical model to learn – it affects only the 2026 nowcast from monthly data (ensemble sCRPS 0.801, TimesFM 0.719 vs 1.042 for the naive forecast). In 2027–2031 net imports occur in 35–61% of draws; this is the residual of the balance (supply minus smelter and processor demand), and the model has no price mechanism – it is not a forecast of net imports.

• Process noise of collection rates and the t copula across member states are order-of-magnitude assumptions (correlations from scrap turnover), not estimates from collection residuals – national intervals may be too narrow or too wide.
• Scrap use in 2026 is not tied to R01/R02: the high net-export nowcast lowers EU-27 use by about 59 kt Cu vs 2025, which the secondary refined copper (R01) and semis (R02) forecasts do not confirm; the nowcast has no stock component.
• Market scope: the smelter feed index includes complex materials, blister copper and electronic scrap (CN 8549), not separated from CN 7404; the copper balance of finished goods does not enter the stock inflow (the 2022–2024 level contains it through the balances, earlier dynamics do not).
• Trade in 2024 in Cu: R03 net exports outside the EU 122.8 kt vs 124.9 kt in the portal's EU-27 balance – the difference comes from the Cu coefficients of alloys (7404 00 91 and 99) and the least-squares balance reconciliation.
• National collection rates 2024: the order is close to the balances (rank correlation 0.81) but not the same – above the balances CZ, FR, GR, PL, SE (Poland 0.73 vs 0.63, position 5 vs 9), below AT, BE, DE, ES, FI, IT, NL (Germany 0.64 vs 0.68, position 9 vs 5).
• EU-27 new scrap in the model (2024: 600.0 kt Cu) lies below the EU-27 balance (847.8 kt; z = −2.8): the 12 national balances count new scrap bottom-up, the EU-27 balance top-down, with a share of about 25% of fabrication consistent with empirical estimates (Soulier et al. 2018). EU-27 scrap use in R03 is therefore lower than in the balance (2024: 1,918.9 vs 2,078 kt Cu).
• Inflow to use of 18 member states (BE, NL, GR and 15 member states without a balance) is a proxy construction; the history of Central European member states before 1990 is uncertain.
• Inflow scale and hibernation lower end-of-life flows similarly and are told apart only by the time pattern; the variant with a 2000–2020 inflow-scale ramp (release 1.0, also without the 2024 inflow observation) gives growing collection to 2031, the base variant does not.
• Art. 40 of Regulation 2024/1157 is adopted law (from 21 May 2027). The assumptions are the content of the Art. 41 list – draft Ref. Ares(2026)8858290 of 18 September 2026 (copper scrap explicitly only to Serbia; China and Hong Kong not on the list; adoption by 21 November 2026) – and the volume share of end-of-waste scrap (class C, by CN code). Net exports in 2031 with a product-status share for China and Hong Kong of 50% / 80% / 95% (all codes): −69.7 / −39.3 / −24.4 kt Cu (medians).
• Blocked exports are settled by redirection to OECD countries (with the receiving-facility audit requirement), additional EU use up to spare capacity and lower collection of low-quality fractions; the stock change is only temporary (2027–2028). Without a price mechanism the split between these channels is an assumption.
• Prior variants with k̂ ≥ 0.7 are computed with separate, converged NUTS runs; their results are indicative.
• The size of the Regulation 2026/1738 effect on vehicles is at the upper end of the possible range (full effect in 2031, without the exemption for post-shredder technologies).
• National use history from the identity (collection + new scrap − trade balance) lacks market meaning in some years due to the mirror asymmetry of intra-EU trade; in 2024–2025 the sum of national use exceeds EU use, and from 2026 they are equal by construction, so some national series jump between 2025 and 2026.
• There is no scenario of releasing the dormant stock (replacement of cable networks) – hibernation is constant over time.
• KGHM's 2018 report gives output from scrap higher than from all purchased materials – the value entered the smelter feed index without correction.
• Waste statistics (W062) are published every 2 years; 2024 data were not available on the release date.

19. Conclusions

• Q1: EU-27 post-consumer scrap collection declines slightly in 2026–2031 – 1,426.3 (1,319.7–1,533.6) kt Cu in 2031 vs 1,454.9 kt Cu in 2025; the growth of release 1.0 does not follow from the data; collection grows in member states with a later build-up of stock (Czechia, Poland).
• Q2: new (post-industrial) scrap is 595.8 (473.8–745.2) kt Cu in 2031 and follows R02 semis output; it is lower than the EU-27 balance assumption.
• Q3: the share of grade 7404 00 10 in supply Cu stays at 53.9%, and the average Cu content at 0.750; sensor-based sorting acts mainly through quality.
• Q4: EU use rises to 2,074.4 kt Cu, and net exports outside the EU fall to −52.8 kt Cu in 2031 (80% interval −305.1 – 192.7) – the sign of the trade balance is uncertain, and net imports in some draws are a balance result, not a market forecast.
• Q5: the fall in net exports (on average 152.6 kt Cu) is explained by Art. 40 of Regulation 2024/1157 (−96.5) and smelter throughput projects (−84.9); the effect of Art. 40 depends on the share of end-of-waste scrap (net exports 2031: −69.7 / −39.3 / −24.4 kt Cu at 50 / 80 / 95% for China and Hong Kong); CRMA and the WEEE target raise collection by 39.7 and 28.4 kt Cu; Regulation 2026/1738 – by 15.5 kt Cu.

20. Reproducibility and reviews

The result is produced deterministically from data D1–D6 and the R01/R02 files (read-only). Stages: stock model fit (NUTS) for 2025, prior variants, the structural variant and backtest origins; forecast draws, nowcast, scenarios, variance decomposition and change decomposition; structural backtest with conformal calibration; AI ensemble (series, backtest, stacking, calibration, SHAP); assembly of the portal file and the note. Random seeds: model 20261028, nowcast 20261029, scenarios 20261035; R02 – version 2.1, seed 20261013. Release 1.1 followed independent reviews A1–A5 (data and sources, mathematics, industry, portal and reproducibility, text) and a verification of the corrections (mathematics, industry); the findings and their status are recorded in the correction plan, and the publisher approves the version in the portal's editorial panel before publication.

Bibliography

Domain literature (DOIs checked in the publishers' registry):

• Born K., Ciftci M.M. (2024) The limitations of end-of-life copper recycling and its implications for the circular economy of metals. Resources, Conservation and Recycling. doi:10.1016/j.resconrec.2023.107318
• Ciacci L., Fishman T., Elshkaki A. et al. (2020) Exploring future copper demand, recycling and associated greenhouse gas emissions in the EU-28. Global Environmental Change. doi:10.1016/j.gloenvcha.2020.102093
• Ciacci L., Vassura I., Passarini F. (2017) Urban Mines of Copper: Size and Potential for Recycling in the EU. Resources. doi:10.3390/resources6010006
• Daehn K.E., Cabrera Serrenho A., Allwood J.M. (2017) How Will Copper Contamination Constrain Future Global Steel Recycling? Environmental Science & Technology. doi:10.1021/acs.est.7b00997
• Diersen I., Bhuwalka K., Olivetti E. (2025) The Opportunity for Utilizing End‐of‐Life Scrap to Meet Growing Copper Demand. Journal of Advanced Manufacturing and Processing. doi:10.1002/amp2.70031
• Dong D., Tercero Espinoza L.A., Loibl A. et al. (2020) Scenarios for anthropogenic copper demand and supply in China: implications of a scrap import ban and a circular economy transition. Resources, Conservation and Recycling. doi:10.1016/j.resconrec.2020.104943
• Dong D., Tukker A., Van der Voet E. (2019) Modeling copper demand in China up to 2050: A business‐as‐usual scenario based on dynamic stock and flow analysis. Journal of Industrial Ecology. doi:10.1111/jiec.12926
• Dong J., Liao J., Huan X. et al. (2023) Expert elicitation and data noise learning for material flow analysis using Bayesian inference. Journal of Industrial Ecology. doi:10.1111/jiec.13399
• Elshkaki A. (2005) Dynamic stock modelling: A method for the identification and estimation of future waste streams and emissions based on past production and product stock characteristics. Energy. doi:10.1016/j.energy.2004.02.019
• Elshkaki A., Graedel T.E., Ciacci L. et al. (2018) Resource Demand Scenarios for the Major Metals. Environmental Science & Technology. doi:10.1021/acs.est.7b05154
• Glöser S., Soulier M., Tercero Espinoza L.A. (2013) Dynamic Analysis of Global Copper Flows. Global Stocks, Postconsumer Material Flows, Recycling Indicators, and Uncertainty Evaluation. Environmental Science & Technology. doi:10.1021/es400069b
• Gottschalk F., Scholz R.W., Nowack B. (2010) Probabilistic material flow modeling for assessing the environmental exposure to compounds: Methodology and an application to engineered nano-TiO2 particles. Environmental Modelling & Software. doi:10.1016/j.envsoft.2009.08.011
• Graedel T.E., van Beers D., Bertram M. et al. (2004) Multilevel Cycle of Anthropogenic Copper. Environmental Science & Technology. doi:10.1021/es030433c
• Kannangara M., Dua R., Ahmadi L. et al. (2018) Modeling and prediction of regional municipal solid waste generation and diversion in Canada using machine learning approaches. Waste Management. doi:10.1016/j.wasman.2017.11.057
• Kroell N., Chen X., Greiff K. et al. (2022) Optical sensors and machine learning algorithms in sensor-based material flow characterization for mechanical recycling processes: A systematic literature review. Waste Management. doi:10.1016/j.wasman.2022.05.015
• Kuipers K.J., van Oers L.F., Verboon M. et al. (2018) Assessing environmental implications associated with global copper demand and supply scenarios from 2010 to 2050. Global Environmental Change. doi:10.1016/j.gloenvcha.2018.02.008
• Laner D., Rechberger H., Astrup T. (2014) Systematic Evaluation of Uncertainty in Material Flow Analysis. Journal of Industrial Ecology. doi:10.1111/jiec.12143
• Loibl A., Tercero Espinoza L.A. (2021) Current challenges in copper recycling: aligning insights from material flow analysis with technological research developments and industry issues in Europe and North America. Resources, Conservation and Recycling. doi:10.1016/j.resconrec.2021.105462
• Lupton R.C., Allwood J.M. (2018) Incremental Material Flow Analysis with Bayesian Inference. Journal of Industrial Ecology. doi:10.1111/jiec.12698
• Mason A.R., Bide T., Wang J. et al. (2025) Bayesian material flow analysis of the construction aggregate cycle in England. Resources, Conservation and Recycling. doi:10.1016/j.resconrec.2025.108135
• Murakami S., Oguchi M., Tasaki T. et al. (2010) Lifespan of Commodities, Part I. Journal of Industrial Ecology. doi:10.1111/j.1530-9290.2010.00250.x
• Müller D.B. (2006) Stock dynamics for forecasting material flows–Case study for housing in The Netherlands. Ecological Economics. doi:10.1016/j.ecolecon.2005.09.025
• Müller E., Hilty L.M., Widmer R. et al. (2014) Modeling Metal Stocks and Flows: A Review of Dynamic Material Flow Analysis Methods. Environmental Science & Technology. doi:10.1021/es403506a
• Oguchi M., Murakami S., Tasaki T. et al. (2010) Lifespan of Commodities, Part II. Journal of Industrial Ecology. doi:10.1111/j.1530-9290.2010.00251.x
• Panasiuk D., Daigo I., Hoshino T. et al. (2022) International comparison of impurities mixing and accumulation in steel scrap. Journal of Industrial Ecology. doi:10.1111/jiec.13246
• Pauliks N., Fishman T., Istrate R. et al. (2026) Assessing historic material losses in European waste from electrical and electronic equipment. Resources, Conservation and Recycling. doi:10.1016/j.resconrec.2026.109034
• Pauliuk S., Heeren N. (2020) ODYM–An open software framework for studying dynamic material systems: Principles, implementation, and data structures. Journal of Industrial Ecology. doi:10.1111/jiec.12952
• Reck B.K., Graedel T.E. (2012) Challenges in Metal Recycling. Science. doi:10.1126/science.1217501
• Ruhrberg M. (2006) Assessing the recycling efficiency of copper from end-of-life products in Western Europe. Resources, Conservation and Recycling. doi:10.1016/j.resconrec.2006.01.003
• Schipper B.W., Lin H.C., Meloni M.A. et al. (2018) Estimating global copper demand until 2100 with regression and stock dynamics. Resources, Conservation and Recycling. doi:10.1016/j.resconrec.2018.01.004
• Soulier M., Glöser-Chahoud S., Goldmann D. et al. (2018) Dynamic analysis of European copper flows. Resources, Conservation and Recycling. doi:10.1016/j.resconrec.2017.10.013
• Wagner M.A., Habib H., Herreras L. et al. (2023) Unlocking the potential of e-waste: A material quantification analysis of Cu, Cr, In, Mg, Nb, and Nd in the EU. Resources, Conservation and Recycling. doi:10.1016/j.resconrec.2023.107243
• Wang F., Huisman J., Stevels A. et al. (2013) Enhancing e-waste estimates: Improving data quality by multivariate Input–Output Analysis. Waste Management. doi:10.1016/j.wasman.2013.07.005
• Wang J., Ray K., Brito‐Parada P. et al. (2024) Bayesian material flow analysis for systems with multiple levels of disaggregation and high dimensional data. Journal of Industrial Ecology. doi:10.1111/jiec.13550
• Wang M., Chen W., Zhou Y. et al. (2017) Assessment of potential copper scrap in China and policy recommendation. Resources Policy. doi:10.1016/j.resourpol.2016.12.009
• Watari T., Nansai K., Nakajima K. (2021) Major metals demand, supply, and environmental impacts to 2100: A critical review. Resources, Conservation and Recycling. doi:10.1016/j.resconrec.2020.105107
• Yokoi R., Nakatani J., Hatayama H. et al. (2022b) Dynamic analysis of in-use copper stocks by the final product and end-use sector in Japan with implication for future demand forecasts. Resources, Conservation and Recycling. doi:10.1016/j.resconrec.2022.106153
• Yokoi R., Watari T., Motoshita M. (2022a) Future greenhouse gas emissions from metal production: gaps and opportunities towards climate goals. Energy & Environmental Science. doi:10.1039/d1ee02165f
• Yoshimura A., Matsuno Y. (2018) Dynamic Material Flow Analysis and Forecast of Copper in Global-Scale: Considering the Difference of Recovery Potential between Copper and Copper Alloy. Materials Transactions. doi:10.2320/matertrans.m2017399
• Zhang L., Cai Z., Yang J. et al. (2015) The future of copper in China–A perspective based on analysis of copper flows and stocks. Science of The Total Environment. doi:10.1016/j.scitotenv.2015.07.021

EU institution reports:

• Muchová L., Eder P., Villanueva A. (2011) End-of-waste criteria for copper and copper alloy scrap: Technical proposals (JRC64207, EUR 24786 EN). European Commission – Joint Research Centre (JRC-IPTS). doi:10.2791/57777
• Passarini F., Ciacci L., Nuss P., Manfredi S. (2018) Material Flow Analysis of Aluminium, Copper, and Iron in the EU-28 (JRC111643, EUR 29220 EN). European Commission – JRC. doi:10.2760/1079

Methodological literature – models and computation (records checked in Crossref and DataCite):

• Abril-Pla O., Andreani V., Carroll C. et al. (2023) PyMC: a modern, and comprehensive probabilistic programming framework in Python. PeerJ Computer Science 9, e1516. doi:10.7717/peerj-cs.1516
• Ansari A.F., Shchur O., Küken J. et al. (2025) Chronos-2: From Univariate to Universal Forecasting. arXiv. doi:10.48550/arXiv.2510.15821
• Ansari A.F., Stella L., Turkmen C. et al. (2024) Chronos: Learning the Language of Time Series. Transactions on Machine Learning Research. doi:10.48550/arXiv.2403.07815
• Challu C., Olivares K.G., Oreshkin B.N. et al. (2023) NHITS: Neural Hierarchical Interpolation for Time Series Forecasting. Proceedings of the AAAI Conference on Artificial Intelligence 37(6), 6989–6997. doi:10.1609/aaai.v37i6.25854
• Das A., Kong W., Sen R., Zhou Y. (2024) A decoder-only foundation model for time-series forecasting. Proceedings of the 41st International Conference on Machine Learning (PMLR 235). doi:10.48550/arXiv.2310.10688
• Demarta S., McNeil A.J. (2005) The t Copula and Related Copulas. International Statistical Review 73(1), 111–129. doi:10.1111/j.1751-5823.2005.tb00254.x
• Diebold F.X., Mariano R.S. (1995) Comparing Predictive Accuracy. Journal of Business & Economic Statistics 13(3), 253–263. doi:10.1080/07350015.1995.10524599
• Gneiting T., Raftery A.E. (2007) Strictly Proper Scoring Rules, Prediction, and Estimation. Journal of the American Statistical Association 102(477), 359–378. doi:10.1198/016214506000001437
• Grömping U. (2006) Relative Importance for Linear Regression in R: The Package relaimpo. Journal of Statistical Software 17(1). doi:10.18637/jss.v017.i01
• Harvey D., Leybourne S., Newbold P. (1997) Testing the equality of prediction mean squared errors. International Journal of Forecasting 13(2), 281–291. doi:10.1016/S0169-2070(96)00719-4
• Hoffman M.D., Gelman A. (2014) The No-U-Turn Sampler: Adaptively Setting Path Lengths in Hamiltonian Monte Carlo. Journal of Machine Learning Research 15, 1593–1623. doi:10.48550/arXiv.1111.4246
• Hyndman R.J., Koehler A.B. (2006) Another look at measures of forecast accuracy. International Journal of Forecasting 22(4), 679–688. doi:10.1016/j.ijforecast.2006.03.001
• Ke G., Meng Q., Finley T. et al. (2017) LightGBM: A Highly Efficient Gradient Boosting Decision Tree. Advances in Neural Information Processing Systems 30 (bez DOI – strona wydawcy)
• Lim B., Arık S.Ö., Loeff N., Pfister T. (2021) Temporal Fusion Transformers for interpretable multi-horizon time series forecasting. International Journal of Forecasting 37(4), 1748–1764. doi:10.1016/j.ijforecast.2021.03.012
• Lundberg S.M., Lee S.-I. (2017) A Unified Approach to Interpreting Model Predictions. Advances in Neural Information Processing Systems 30. doi:10.48550/arXiv.1705.07874
• Owen A.B. (2014) Sobol' Indices and Shapley Value. SIAM/ASA Journal on Uncertainty Quantification 2(1), 245–251. doi:10.1137/130936233
• Romano Y., Patterson E., Candès E. (2019) Conformalized Quantile Regression. Advances in Neural Information Processing Systems 32. doi:10.48550/arXiv.1905.03222
• Vehtari A., Gelman A., Simpson D., Carpenter B., Bürkner P.-C. (2021) Rank-Normalization, Folding, and Localization: An Improved R̂ for Assessing Convergence of MCMC. Bayesian Analysis 16(2), 667–718. doi:10.1214/20-BA1221
• Vehtari A., Simpson D., Gelman A., Yao Y., Gabry J. (2024) Pareto Smoothed Importance Sampling. Journal of Machine Learning Research 25(72). doi:10.48550/arXiv.1507.02646
• Weibull W. (1951) A Statistical Distribution Function of Wide Applicability. Journal of Applied Mechanics 18(3), 293–297. doi:10.1115/1.4010337

Data and legal acts:

• Eurostat Comext DS-045409 – CN 7404 00 10/91/99 trade, EU member states, intra- and extra-EU partners, 2005–2025 and Jan–Aug 2026. Eurostat. 2026-10-28 (accessed). https://ec.europa.eu/eurostat/api/comext/dissemination/sdmx/2.1/data/DS-045409
• Eurostat – generation of waste (env_wasgen, W062 non-ferrous metal waste), biennial 2010–2022. Eurostat. 2026-10-28 (accessed). https://ec.europa.eu/eurostat/databrowser/view/env_wasgen/default/table
• Eurostat – waste electrical and electronic equipment (env_waselee, env_waseleeos), 2005–2024. Eurostat. 2026-10-28 (accessed). https://ec.europa.eu/eurostat/databrowser/view/env_waselee/default/table
• Eurostat – end-of-life vehicles (env_waselvt), 2005–2023. Eurostat. 2026-10-28 (accessed). https://ec.europa.eu/eurostat/databrowser/view/env_waselvt/default/table
• Eurostat – population on 1 January (demo_pjan), GDP (nama_10_gdp), construction output (sts_copr_a), car registrations. Eurostat. 2026-10-28 (accessed). https://ec.europa.eu/eurostat/databrowser/
• Regulation (EU) 2024/1252 on critical raw materials (CRMA). Publications Office of the EU. 2024-04-11. http://data.europa.eu/eli/reg/2024/1252/oj
• Commission Implementing Regulation (EU) 2026/1116 – waste streams with critical raw material recovery potential. Publications Office of the EU. 2026-05-26. http://data.europa.eu/eli/reg_impl/2026/1116/oj
• Regulation (EU) 2024/1157 on shipments of waste. Publications Office of the EU. 2024-04-11. http://data.europa.eu/eli/reg/2024/1157/oj
• Directive 2012/19/EU on waste electrical and electronic equipment (WEEE). Publications Office of the EU. 2012-07-04. http://data.europa.eu/eli/dir/2012/19/oj
• Directive (EU) 2024/884 amending Directive 2012/19/EU. Publications Office of the EU. 2024-03-13. http://data.europa.eu/eli/dir/2024/884/oj
• Directive 2000/53/EC on end-of-life vehicles. Publications Office of the EU. 2000-09-18. http://data.europa.eu/eli/dir/2000/53/oj
• Regulation (EU) 2026/1738 on end-of-life vehicles. Publications Office of the EU. 2026-07-08. http://data.europa.eu/eli/reg/2026/1738/oj
• Commission Regulation (EU) No 715/2013 – criteria determining when copper scrap ceases to be waste. Publications Office of the EU. 2013-07-25. http://data.europa.eu/eli/reg/2013/715/oj
• Regulation (EU) 2023/1542 concerning batteries and waste batteries. Publications Office of the EU. 2023-07-12. http://data.europa.eu/eli/reg/2023/1542/oj
• Regulation (EU) 2024/1781 – framework for ecodesign requirements for sustainable products (ESPR). Publications Office of the EU. 2024-06-13. http://data.europa.eu/eli/reg/2024/1781/oj
• Regulation (EU) 2023/956 establishing a carbon border adjustment mechanism (CBAM). Publications Office of the EU. 2023-05-10. http://data.europa.eu/eli/reg/2023/956/oj
• Commission Communication COM(2025) 125 – European Steel and Metals Action Plan. Publications Office of the EU. 2025-03-19. http://publications.europa.eu/resource/celex/52025DC0125
• Commission Communication COM(2025) 945 – RESourceEU action plan. Publications Office of the EU. 2025-12-03. http://publications.europa.eu/resource/celex/52025DC0945
• Draft Commission Delegated Regulation supplementing Regulation (EU) 2024/1157 – list of countries to which exports of certain non-hazardous wastes are authorised (Ref. Ares(2026)8858290, with annex; copper scrap – Serbia, annex pp. 28–29). European Commission. 2026-09-18. https://ec.europa.eu/info/law/better-regulation/have-your-say/initiatives/14861
• Commission Delegated Regulation C(2026) 7390 amending Annexes III, IIIA, IIIB and IV to Regulation (EU) 2024/1157 (Council document ST 14953/26). Council of the European Union. 2026-10-26. https://data.consilium.europa.eu/doc/document/ST-14953-2026-INIT/en/pdf