Methodological note – Copper semi-finished products output in the EU 2026–2031

ai-waw.com · version 2.0 · 2026-10-17 · AIWAW-2026-R02

Contents: 1 · 2 · 3 · 4 · 5 · 6 · 7 · 8 · 9 · 10 · 11 · 12 · 13 · 14 · 15 · 16

1. Method summary and hybrid model diagram

Version 2.0 replaces the version 1.x bottom-up model (output = EU use / (1 − net export share), country shares as random walks, capacity outside the model) with a three-layer hybrid model: 1) the semis demand forecast in four customer sectors – construction, energy and grids, industry and transport; 2) two parallel modules – the foreign trade model (imports, exports, competition) and the production plant model (capacity, investment, closures); 3) the output forecast by country and product with physical constraints, a copper balance check and probabilistic scenarios. Levels are anchored in International Wrought Copper Council (IWCC) data for the EU-27 (EU27+UK net of the United Kingdom).

The 2027–2031 EU-27 level is set by an equal-weight ensemble of five models (the hybrid model and four statistical models); the hybrid model gives the split by country and product, the distribution (intervals) and the scenarios.

Result: EU-27 copper and alloy semis (excluding wire rod) in 2031 – median 3,946 kt (80% interval 3,293–4,706; 2025: 3,886 kt), +0.3% a year on average; wire rod – 2,342 kt (+0.4% a year).

Model diagram (arrows – direction of information flow; modules in one step run in parallel on the same draws):

1. Semis demand forecast – construction, energy, industry, transport (IWCC end uses × Eurostat indicators, GDP and investment per IMF and EC → sector × product matrix) ↓
2. Parallel modules: [a] Foreign trade model – imports, exports, competition (Comext CN 7407–7411 → IWCC products, partner GDP per IMF, US tariff, non-EU supplier shares) ∥ [b] Production plant model – capacity, investment, closures (plant register, company reports, ICSG; foil separately; project probability and delay) ↓
3. Output forecast by country and product – physical constraints (capacity, company lower bounds), copper balance (check: R01, cathodes, scrap) and probabilistic scenarios (Monte Carlo, 2026 nowcast, model ensemble, calibration of the final distribution)

Each Monte Carlo draw passes through all layers; named scenarios use the same random numbers (common random numbers), so the difference against the base in a given draw results only from the parameter change.

2. Scope, definitions, data and symbols

Total semis (semis) = copper semis + alloy semis per IWCC, excluding wire rod; wire rod (semis.wire_rod) is the intermediate for wire and is not added to semis. Unit – kt of product mass; Cu content – @kt_cu series. Areas: EU-27, countries with significant sold output (≥ 40 kt of semis or wire rod on average in 2020–2024) and “other EU countries and unallocated output” (REST_EU27). History 2012–2025, nowcast 2026, forecast 2027–2031.

Foil: the IWCC product “Copper Plate, Sheet & Strip (Inc. Foil)” includes foil by name, and IWCC lists heading 7410 – copper foil regardless of technology – among copper and alloy tariff codes (IWCC statistics page). The code list is, however, general (it also includes matte, scrap and powders) and does not define the product scope, and IWCC does not mention electrodeposited foil explicitly. Including electrodeposited foil in the anchor is therefore an ASSUMPTION (class C): it is treated as part of the anchor (also in CN 7410 imports and in projects), and in the plant module as a separate sub-form, so that new foil plants do not displace rolled strip. Sensitivity: without EU foil projects (SK Nexilis, Volta) EU-27 copper plate, sheet, strip and foil in 2031 would be lower by about 26 kt on average (about 8% of the product, about 0.65% of total semis; Poland −9, other countries −17 kt).

3. IWCC anchor and removal of the United Kingdom

EU-27 output and use of product p are derived from IWCC data for the EU27+UK region net of the United Kingdom: output – the UK share of form output per PRODCOM (EU28 − EU-27, 2015–2019; EU28 values are rounded Eurostat estimates); use – UK output plus EU-27→GB net shipments (Comext) divided by the EU supply share φ_GB (assumption 0.70, range 0.55–0.85). The UK share uncertainty is drawn in every draw.

ω_p – share of product p in form f per IWCC 2025.

4. Module 1 – semis demand in four sectors

IWCC copper end uses (End Use Summary, Cu content, world 2012–2025) are allocated to four sectors by IWCC group and category description. “Electrical” (industrial transformers and motors) is split: share θ – energy, the rest – industry; θ ~ U(0.25; 0.55). Cooling and air conditioning (equipment) belong to industry, air-conditioning tube in buildings – to construction. Construction is measured by buildings (F41) and specialised works (F43) – building installations – while civil engineering (F42, incl. networks) enters the energy index; F41 and F43 are published from 2015, before that the construction index is total F.

The sector s index is the average of log changes of end-use indicators weighted by their Cu content (world 2025):

Sector forecast: 2026 – ratio of the averages of the available months of 2026 and the same months of 2025 (seasonally adjusted), with standard deviation σ_s · √((12 − m)/12) for m observed months; 2027–2031 – regression of the index change on a macroeconomic driver x (EU GDP for energy and transport, EU gross fixed capital formation for construction and industry), OLS 2011–2025. The driver is drawn around the forecast (uncertainty from historical volatility, class C):

Ω – covariance matrix of sector residuals (correlations from 2011–2025). Driver uncertainty: GDP σ = 1.16 pp, GFCF σ = 2.76 pp, correlation 0.67 (2011–2025 excluding 2020–2021) – IMF and EC forecast error statistics are not in the forecast data. GDP comes from the IMF October 2026 release and investment 2026–2027 from the EC Spring 2026 forecast (the latest releases at the forecast date); then a bridge through a regression of investment on GDP.

The sector × product matrix is built from a prior end use × product matrix (product shares in the end use's Cu – an assumption from IWCC category descriptions, e.g. “Aircon tube” → copper tube, “Water distrib., heating” → copper tube and brass fittings), multiplied by the IWCC end-use structure and column-scaled to the 2025 Cu use of IWCC EU-27 products. Columns are sector shares in product use w_p,s; wire rod takes the copper wire weights. In each draw the weights are perturbed by a Dirichlet distribution (κ_D = 50) and shifted with the drawn θ.

Implied sector structure of EU-27 Cu use: construction 23.6%, energy and grids 21.0%, industry 39.0%, transport 16.4% – practically the same as the global one (IWCC does not publish the EU structure). The construction share in tubes and brass in the EU is probably higher (water and heating installations, fittings) – a limitation described in chapter 15.

Product use in the EU-27 (IWCC, net of UK) responds to the product's sector index. The elasticity β is common (DerSimonian–Laird estimator for 8 products; wire rod duplicates copper wire – same sector weights, use correlation 0.99 – and takes the common β), and the own trend α_p is shrunk to a common mean with the same estimator and damped by half (λ = 0.5). The (β, ᾱ) distribution in the draws comes from a block bootstrap of years (3-year blocks, 2,000 replications, the same years for all products), because the product estimates are not independent:

The local level is chosen when the autocorrelation of use changes is negative (a sign of measurement noise in IWCC data); the start then comes from a Kalman filter, otherwise from 2025. The DerSimonian–Laird estimator gives τ² = 0 (the Q test shows no heterogeneity), so β and α are common to all products; product differences come from the sector matrix, trade and the shocks u_p,t.

β is identified mainly by the 2020–2021 fall and rebound; without those years the elasticity is clearly lower. Robust estimation confirms the full-sample value, and the wide bootstrap distribution carries this uncertainty into the draws and scenarios (chapter 12).

5. Module 2a – foreign trade: imports, exports, competition

Extra-EU imports and exports (Comext DS-045409, mass) are converted to IWCC products with an explicit CN6 → product key; code 7408 11 (wire with a cross-section > 6 mm) is wire rod, foil (7410) belongs to plate, sheet and strip. The key is reconciled with history: the difference between IWCC (P − D) and Comext (X − M) net exports is the reconciliation residual R (stocks, captive output, classification mismatch, United Kingdom trade; for copper wire – cable trade outside chapter 74).

Imports are import penetration m = M/D on a logit scale (m ∈ (0; 1)) with a drift (10-year average, damped by half) and a shock; exports grow with external demand – partner GDP (IMF) weighted by the product's 2023–2025 extra-EU export structure – with elasticity η (DerSimonian–Laird; distribution from a block bootstrap of years) and a damped trend c_p^X; the reconciliation residual reverts to its 5-year mean as an AR(1) process with product-specific parameters:

The 50% US tariff on semi-finished copper products (Proclamation 10962, from 1 Aug 2025) enters as a permanent reduction of the export level to the USA: r – exports to the USA in the available months of 2026 vs the same months of 2025 (Comext); since 2025 already contains 5 months after the tariff, the 2026 level changes by 1 − w_US + w_US · r / (7/12 + 5/12 · r), where w_US is the US share of the product's 2023–2025 extra-EU exports.

CN8 subheadings show that imports grow unevenly within families (e.g. coiled vs straight tubes, foil, fine wire); the model works at the IWCC product level, so this structure is context. EU trade defence – as a legal fact: the anti-dumping proceeding on copper tubes from China, Mexico, Vietnam and Uzbekistan (OJ C/2026/1506, 12 Mar 2026) with import registration (Implementing Regulation 2026/1161). No duty rates had been published by the forecast date, so a duty variant cannot be parameterised from EUR-Lex documents – the base makes no assumption on the outcome (version 2.1). CBAM does not cover chapter 74 (Regulation 2023/956).

6. Module 2b – production plants: capacity, investment, closures

Plant-module forms: wire rod, wire, bars and profiles, rolled sheet and strip, foil, tubes. Foil (national PRODCOM 24.44.25; the rest of the EU in the “other” item) is separated from “copper.sheet_strip” with its 2025 share (13.0%), changed only by foil projects. Capacity of area a in form f: the documented value when the plant register gives the capacity of all the area's plants in the form and it is not below output; otherwise a lower bound – maximum 2021–2025 output divided by the peak-year utilisation (class C). For wire rod in Poland (the KGHM Polska Miedź S.A. plant in Orsk) the company report gives “about 270 kt a year” as an output level, not capacity – so the general rule applies (2021–2025 maximum: 271.0 kt in 2024); company documents announce higher output at Orsk (a new Contirod line, Upcast II) without quantities, so they have no quantitative term. The “other” item has capacity not below the remainder of ICSG regional capacity (Dec 2020) net of the United Kingdom, Norway and Switzerland shares.

Maximum utilisation u^max and peak-year utilisation u^peak are class C assumptions: the available company data (table above, including sales and deliveries instead of output) give ratios of 0.34–0.91 and do not measure full-load utilisation; the ratio of IWCC 2020 output to ICSG Dec 2020 capacity is 0.5–0.8 depending on the form. Condition (6.3) ensures that capacity is not below the output achieved in 2021–2025.

Projects, closures and relocations (primary company documents; parameters – class C assumptions): start with probability p, delay ~ Poisson (mean 1 year under construction, 2 years planned), ramp-up 50% in the start year and 100% from the next year; when the documented date has passed without confirmation – start in 2027 at the earliest. New output Q = ΔC · r · u^new fully raises EU output: a share ϑ ~ U(0.3; 0.7) replaces imports (at most half of the product's imports), the rest raises exports – existing EU plants are not displaced. After a closure a share ψ of output is taken over by other EU plants (by default ψ ~ U(0.3; 0.8)), the rest goes to imports. A relocation only changes the split between countries.

Rationale for event parameters: SK Nexilis (Stalowa Wola) – plant completed without a confirmed start-up, and the anchor customer (Northvolt, about 80% of offtake per the SKC release) filed for bankruptcy on 12 Mar 2025 – p = 0.5 and utilisation 0.2–0.5. Rai – Hailiang board resolution of 3 Aug 2026 and the job-protection plan agreement: relocation of solid brass rod capacity to Germany and Italy, p = 0.9; scope – part of the 27.5 kt output (solid rod share not stated). KME Stolberg – portfolio transferred to other KME plants, ψ = 1. Nexans Lens – wire rod increase U(0; 85 kt): upper bound = “over +50%” (release of 22 Oct 2024) × 170 kt capacity (plant page); the line mainly remelts scrap instead of cathodes, so the volume increase is uncertain. Projects without quantities in documents (modernisations, electrorefining in Córdoba) have no quantitative term; items on the forecast exclusion list were removed from the register and events.

Wire: ICSG covers only alloy wire in the region including the United Kingdom, Norway and Switzerland, without copper wire drawing at cable makers, while the model capacity covers all EU-27 wire (a lower bound from output, including cable makers); the figures differ in scope. Foil: ICSG gives 10 kt, less than PRODCOM output, so the model's foil capacity is a lower bound from output plus projects.

• The register does not cover all the area's plants in the form or the capacity scope is wider than the form (e.g. all of the plant's products) – the register capacity is then not a lower bound.

7. Layer 3 – output by country and product

EU-27 product output follows from the demand and trade balance identity, increased by new event output and reduced by closure losses not taken over in the EU. The lower bound 0.2 · D prevents negative output in extreme trade draws (the shortfall reduces imports). Split into areas: the base share with a random walk from 2025 applies to EU output without events, and events are added directly to the area concerned; then projection onto capacity and company lower bounds, and a lack of free capacity in the EU means imports:

Z_a,f,t – relocations (zero sum). The surplus above capacity goes to areas with free capacity in proportion to their output. Company lower bounds (the area's share of the EU sum of forms not below that of the reporting year) are applied in the model and again in the published draws, at the expense of other areas: Spain – wire rod, wire ≥ 5.5% (company reports 2024); Spain – wire rod, wire, bars and profiles, rolled sheet and strip, tubes ≥ 6.9% (company reports 2024); Greece – tubes ≥ 20.9% (ElvalHalcor report 2025); Greece – bars and profiles, tubes ≥ 6.0% (ElvalHalcor report 2025). Spanish wire rod drawn into wire within the same company (La Farga, Cunext) is outside PRODCOM sold output and goes to the “other” item. Polish wire rod = KGHM Polska Miedź S.A. data. Coherence in every draw: EU-27 = Σ areas, aggregates = Σ products; wire rod outside the sum.

Check: the largest share of published draws with a share below a company lower bound – 0.00%.

8. 2026 nowcast

2026 is conditioned on the STS production index NACE 24.44 (January–August 2026 vs January–August 2025: +0.7%): countries with an index – ensemble Kalman filter in logs; rest of the EU – importance weighting and systematic resampling (30,000 proposals, ESS 27,603). The 2026 prior was widened by c₀ = 1.34 – the conformal multiplier of hybrid model errors at a 1-year horizon in the backtest. The index covers the whole NACE 24.44 activity (including wire rod and refining) and only some countries, and the semis nowcast median combines it with the model prior – hence the semis nowcast change differs from the index change.

9. Backtest and model ensemble

As-of backtest: origins 2015, 2017, 2019 and 2021, horizon 1–5 years, 12 EU-27 series. The backtest hybrid model (HYBRID) uses only data to the origin (sector, demand and trade regressions; drivers – historical averages, as origin-year IMF and EC forecasts are not archived), without national capacity and events. The HYBRID-R diagnostic plugs in realised sector indices and is not in the ensemble. Ensemble: mean of log medians of members with equal weights (DRIFT excluded as it reduced accuracy).

ETS, ARIMA and BSTS values come from the same backtest records as the others (the backtest block of the forecast file).

In the backtest (4 origins 2015–2021, horizon 1–5 years) the model ensemble is not more accurate than the random walk: mean MAPE 8.9% vs 8.5% (v1.3 ensemble on the same data – 9.0%). The hybrid model alone has a MAPE of 9.1%, and 8.3% with sector indices known in advance: most of the error comes from annual use and trade shocks, not from the business-cycle forecast. The forecast is therefore mainly about the level, structure and range of uncertainty and the effects of scenarios, rather than an edge over simple extrapolation.

The ensemble sets the EU-27 product level in 2027–2031 (top-down correction: a year multiplier common to the product's draws and areas); the hybrid model contributes structure and spread.

10. Conformal calibration of the final distribution

For each series (products and the aggregates: semis total, copper, alloys) the multiplier c of the hybrid model spread sd₀ was set on ensemble backtest errors (80% and 90%) with an out-of-sample coverage condition and an origin-level lower bound (maximum over origins of the root mean square standardised error); c ∈ [0.5; 2.5]. LOO coverage is computed without leakage: the multiplier and the lower bound are set without the given origin. Published spread: the change from 2026 has standard deviation T_p,t = c_p · sd_0;p,t (non-decreasing over time) for products and aggregates simultaneously – components independent across products, common to the group (copper, alloys) and common to the semis total are set iteratively. The FINAL distribution is calibrated: after projection onto capacity and company lower bounds the standard deviation is checked and, if it falls below 0.97 of the target, the target is raised and the whole step repeated (up to 4 times).

The final-to-target spread ratio in 2027–2031 lies in 0.976–1.866. Above target are: total semis in 2027–2028 (1.118 and 1.091), alloy tubes (1.26–1.87) and alloy wire (up to 1.20); the intervals there are wider than the target, because the common group and total components needed for the alloy aggregate exceed the target of products with a small c. The intervals of these series are therefore conservative (not narrowed). LOO 90% coverage is below nominal in 4 series (copper tubes, total alloys, alloy bars and alloy sheet); with 19 points from 4 correlated origins, a result of 16/19 at probability 0.9 has P(≤ 16) ≈ 0.30, so it does not indicate narrowed intervals.

11. Cu content and copper balance check

Cu content: product mass × Cu share (copper semis 0.99–1.00, wire rod 0.999–1.0, alloys 0.622–0.72), drawn in every draw. The balance without double counting covers first-processing products: wire rod, copper semis excluding wire (wire is drawn from rod) and alloy semis. The balance is a CHECK: output is not adjusted, because the ceilings on cathode imports and direct-melt scrap are assumptions rather than a physical limit; the share of draws in which Cu in products exceeds metal availability is reported, together with the headroom:

Components: maximum net cathode imports (CN 7403 11) 2012–2025 – 729 kt (2025: 541); maximum direct-melt scrap – 984 kt Cu; net scrap exports CN 7404 2023–2025 – 151 kt (mass). R01 – a draw from the R01 forecast distribution (version 1.2) by quantile interpolation, one u per draw, independent of R02 draws.

In 2031 Cu in products exceeds the assumed metal availability in 15.3% of draws – in the upper tail of the output distribution combined with low draws of EU refined copper output. The median headroom is positive, but the p10 headroom is negative from 2030 and the p5 headroom from 2028. Refined copper use by semis producers (R02) exceeds EU-27 output (R01) – the gap is covered by net cathode imports; the link block shows this relation as a check, not a sum.

12. Probabilistic scenarios

The base scenario is the full Monte Carlo distribution. Named scenarios shift hybrid model parameters on the same random numbers; the ratio of the scenario result to the base result in a given draw (area × product; a 0.5 kt offset in the numerator and denominator guards against division by very small values) multiplies the calibrated published draw, after which the scenario capacity and company lower bounds are applied; the copper balance – as a check.

• Accelerated grid build-out – Energy and grid demand grows over the whole horizon at the top-tercile rate of 2011–2025 (Eurostat production indicators C271, C2732, F42) – a historical parameter. The direction is consistent with the EU Action Plan for Grids COM(2023) 757 (EUR 584 billion of grid investment this decade); the plan's scale was not converted into semis demand.
• Extra-EU import surge – New capacity just outside the EU: the Hailiang plant in Tangier (tubes, rods, foil; the company document points to the European market) supplies part of its output to the EU, and the new CRW Metal rod mill in Türkiye (per the company – for the Turkish market) displaces part of EU wire rod exports to Morocco, Tunisia and Serbia. The share χ is an upper bound – an assumption.
• Closures under high energy costs – Stress test: in countries with an industrial electricity price above the 2025 EU-27 average (Eurostat nrg_pc_205), 10% of form output closes from 2027 and 15% in total from 2029, weighted by the form's energy intensity; part of the output is taken over by other EU plants, the rest by imports.
• Construction recession – In 2027–2029 construction output repeats the worst three-year period of 2001–2025 (Eurostat, construction F: 2008–2010), with transmission to industry (fittings, heat pumps, air conditioning – C28); from 2030 – the base path.

The effects of scenarios acting through sector demand are proportional to β – with β from the sample excluding 2020–2021 they are several times smaller.

13. Results and comparison with version 1.3

Largest differences of 2031 medians against version 1.3: Other EU countries and unallocated output – wire rod 45 kt; Other EU countries and unallocated output – semis total 34 kt; Spain – wire rod 30 kt; Germany – semis total −23 kt. They follow from the plant module (SK Nexilis foil with a lower probability, the Rai relocation, the Lens increase), exports to the USA after the tariff and the new construction indicators.

Percentiles (including medians) do not add up across areas and products; draws and means do.

14. Sensitivity and variance shares

15. Limitations

• The 2027–2031 EU-27 level is set by an equal-weight ensemble of five models; the hybrid model (weight 1/5) gives the structure, intervals and scenarios. Plant events (projects, closures) act mainly on the split between countries.
• Hybrid model factors explain about 37% of the variance of the 2031 semis total (wire rod – about 85%); the rest is annual product shocks, country shares, the model ensemble and interval calibration – scenarios move the median less than the interval width.
• The demand elasticity to sectors is identified mainly by 2020–2021 (β = 0.9; without those years 0.3); β and η uncertainty comes from a block bootstrap of years and is carried into the draws, but the grid and recession scenario results depend on β (variant in the note).
• The copper end-use structure (IWCC End Use Summary) is global and is applied to the EU unchanged; the allocation of end uses to 4 sectors and the sector × product matrix are assumptions (class C) with sampled uncertainty. The construction share in tubes and brass in the EU is probably higher than in the global structure.
• Capacity is documented for only part of the plants; the rest – a lower bound from 2021–2025 output and an assumed utilisation. The “about 270 kt” figure for the Orsk rod mill is an output level in the company report, not capacity. Wire drawn at cable makers has no independent capacity measure.
• Including electrodeposited foil (CN 7410) in the IWCC product “Copper Plate, Sheet & Strip (Inc. Foil)” is an assumption (class C) – IWCC lists heading 7410 only in a general code list; foil is a separate sub-form in the plant module, and its share of the product is taken from PRODCOM 2025 (constant apart from projects). Without EU foil projects, total semis in 2031 would be lower by about 26 kt (about 0.65%).
• Project probabilities and delays are assumptions by status; many projects in company documents do not state a capacity increase and have no quantitative term (e.g. the new Contirod and Upcast II lines at the KGHM plant in Orsk).
• Trade: the CN 7407–7411 → IWCC product key is explicit, and the IWCC–Comext reconciliation residual is modelled statistically; exports to the USA after the 50% tariff are a permanent level reduction from 2026 data, without diversion of other exporters' supplies to the EU market.
• The copper balance is a check (output not adjusted): the R01 distribution is drawn independently of R02, and the ceilings on cathode imports and direct-melt scrap are assumptions. The p10 metal headroom is negative from 2030 and the p5 headroom from 2028 (upper output tail with low refined copper output).
• The “other EU countries and unallocated output” item is about 47% of 2025 semis: mainly wire drawn at cable makers, wire rod in Belgium, France and Greece and Spanish wire rod drawn into wire within the same company (La Farga, Cunext – outside PRODCOM sold output), countries without data.
• GDP and investment forecast uncertainty is taken from historical volatility (class C), as IMF and EC forecast error statistics are not in the forecast data; the recession scenario parameter comes from 2008–2010 (construction output F).
• The backtest covers the simplified hybrid model (without capacity, events and the sampled matrix); a backtest of the full hybrid and the sensitivity of conformal multipliers to the origin set – in version 2.1. No archived vintages of IWCC, Comext and IMF/EC forecasts. The intervals of total semis in 2027–2028 and of alloy tubes and alloy wire are wider than the calibration target (up to 1.9 times), because the common group and total components exceed the target of products with a small conformal multiplier.
• The copper price, conversion margins, aluminium substitution (cables, windings, heat exchangers), copper intensity of electric vehicles and data-centre demand are not separate model variables (they act only through historical trends).

16. Reproducibility

Deterministic run (seed 20261013; Python 3.12.10, numpy 2.4.2, pandas 3.0.1): data download and preparation, model, publication and check scripts in the forecast repository; the same code and data give an identical result file. R01 forecast in the balance check: version 1.2 (MD5 307d328c43d6…).

Methodological bibliography

• Angelopoulos A.N., Bates S. (2023). Conformal prediction: a gentle introduction. Foundations and Trends in ML 16(4), 494–591.
• DerSimonian R., Laird N. (1986). Meta-analysis in clinical trials. Controlled Clinical Trials 7(3), 177–188.
• Evensen G. (2003). The Ensemble Kalman Filter: theoretical formulation and practical implementation. Ocean Dynamics 53, 343–367.
• Glasserman P. (2004). Monte Carlo Methods in Financial Engineering. Springer (common random numbers).
• Hyndman R.J., Athanasopoulos G. (2021). Forecasting: Principles and Practice, 3rd ed. OTexts.
• Künsch H.R. (1989). The jackknife and the bootstrap for general stationary observations. Annals of Statistics 17(3), 1217–1241.
• Smith J., Wallis K.F. (2009). A simple explanation of the forecast combination puzzle. Oxford Bulletin of Economics and Statistics 71(3), 331–355.
• Vovk V., Gammerman A., Shafer G. (2005). Algorithmic Learning in a Random World. Springer.