Interdependence among crude oil, gold, and stock markets: a systematic bibliometric review of spillovers and co-movements (2010–2025)

Spillovers and co-movements among crude oil, gold, and equity markets sit at the heart of portfolio diversification, hedging, energy-risk assessment, and financial-stability policy. This systematic review, reported in accordance with the PRISMA 2020 guidelines and adapted to a bibliometric evidence-synthesis design, maps how that field is built intellectually, synthesizes what its empirical literature has established, identifies where its newest fronts lie, and situates it against prior bibliometric reviews. The review question was framed with a PICOS structure; the corpus comprises 1,081 unique documents drawn from Web of Science and Scopus under a PRISMA 2020 selection flow. The search window is 2010–2026 and the analytical window 2010–2025 (n = 1,077): the four 2026 Early Access records are excluded from every statistic that depends on publication year, and their removal from the remaining analyses moves no reported share by more than 0.4 percentage points. Within the strictly defined crude oil–gold–equity triad this is, to our knowledge, the most extensive corpus assembled so far—about 58% larger than the closest triad-focused comparator—and the first to treat the three assets jointly rather than as separate dyads. We combine performance analysis with complementary science-mapping techniques: co-occurrence of lemmatized keywords with Louvain community detection (γ = 1.0; Q = 0.231), author co-citation, bibliographic coupling confined to 2022–2025, and country-level co-authorship, with normalized indicators (h-index, g-index, citations per document per year, and the MCP/SCP ratio). Output grew at a 27.4% compound annual rate over 2010–2024 and then leveled off from 2023 (year-over-year rates of 4.1%, 0.0%, and −3.4%). Keyword co-occurrence resolves into four thematic clusters; author co-citation into three intellectual schools (safe haven/GARCH, connectedness/Diebold–Yilmaz, and COVID/crypto/alternative havens); and bibliographic coupling into nine emerging fronts, among them a small green bonds and energy-transition front that earlier reviews do not register. Latin America and the Caribbean account for at most 1.76% of corpus documents while supplying 9.5% of world commodity exports, a representation ratio of 0.19. The findings inform diversification and hedging strategies, regulatory debates on the financialization of energy and commodity markets, the emerging interface between fossil-energy assets and green finance, and regional research agendas.

Crude oil is the world's most heavily traded commodity, at once a production input for nearly every sector and a financial asset shaped by speculation and portfolio allocation (Hamilton, 2009; Kilian and Murphy, 2014). How its price shocks propagate to the rest of the economy—and to financial markets in particular—is a central question of energy economics. Gold sits in a very different place within the financial system: a precious metal supported by industrial and jewelry demand, a long-standing store of value and inflation hedge, and a safe haven when markets turn stressful or geopolitics turns uncertain (Baur and McDermott, 2010; Beckmann et al., 2015).

Equity markets connect to both commodities through several channels: cost pressure on corporate earnings, inflation expectations, flight-to-quality in turbulent periods, and diversification gains in calmer ones (Mensi et al., 2013; Bouri et al., 2017b). The 2008 global financial crisis, the 2014–2016 oil price shocks, the 2020 COVID-19 pandemic, and the Russia–Ukraine war since 2022 each made the same point: the three markets are far more tightly bound than conventional wisdom assumes, with abrupt regime shifts and volatility spillovers that call for serious analytical machinery (Diebold and Yilmaz, 2012; Tang and Xiong, 2012).

Why this matters in practice is easy to state. Portfolio managers lean on stable correlations to diversify, yet those correlations tend to spike precisely when stress hits and diversification is needed most (Baruník and Krehlík, 2018). Regulators and central banks need tools that anticipate how energy and commodity shocks propagate to financial stability (Bonato et al., 2024). Strategic allocators must contend with financialization (Tang and Xiong, 2012), where institutional money treats commodities as financial assets in their own right. And as the energy transition rewires commodity markets, the interface between fossil-energy assets, traditional havens, and new instruments such as green bonds is becoming a policy question in its own right.

Prior bibliometric reviews and positioning of the study

As the field has grown, so has the number of bibliometric reviews trying to organize parts of it. Among the reviews identified and compared here, none treats oil, gold, and stocks together as a single analytical object, and none reaches the post-pandemic consolidation examined in this study. Table 1 brings together eight related bibliometrics published between 2020 and 2025, ordered by how close they sit to our object of study.

Four are direct competitors by subject. Bashir (2022), in Environmental Science and Pollution Research, analyzes 684 documents on the oil price–stock market nexus with the bibliometrix R package, but stays within the oil–stocks dyad and leaves gold out. Singh et al. (2023), in a meta-literature review, cover 594 documents on market connectedness and volatility spillovers from 1991 to 2021 and distinguish four research streams, but stop before the field's most expansive years (2022–2025). Anas (2025), in the Journal of Economic Surveys, reviews 617 documents on hedge and safe-haven assets (1996–2022), including gold and Bitcoin as alternative havens, though the lens is the safe-haven property at large rather than the commodity–equity triad. Patel et al. (2025), in the Review of Financial Economics, study 200 articles on the energy–financial markets linkage (1984–2023) using co-citation, co-authorship, and cartographic mapping, but keep to the energy–finance dyad with a smaller corpus.

Four others overlap more loosely, in method or in scope. Liu et al. (2020), in the International Review of Economics & Finance, offer a scientometric review of the oil price–stock market nexus across 1,342 documents — a corpus larger than the present one in absolute terms, but built on the oil–stock dyad, without gold and without the triadic interdependence framing used here, so the two corpora are not measurements of the same object. Su et al. (2021), in Emerging Markets Finance and Trade, produced an early bibliometric study of volatility spillovers that set much of the methodological template for this line of work. Khan et al. (2025), in Qualitative Research in Financial Markets, examine volatility during crisis periods with a broad bibliometric toolkit. And Almeida and Gonçalves (2023), in the Journal of Risk and Financial Management, pair systematic review with bibliometrics to study cryptocurrencies as hedging assets across 146 documents.

Against that backdrop, this study stakes out three claims, each bounded by the comparison set assembled in Table 1 rather than by the bibliometric literature at large. On breadth, its 1,081 unique documents, obtained by deduplicating Web of Science against Scopus, form the most extensive corpus assembled so far for the strictly defined oil–gold–stocks triad, roughly 58% larger than the closest triad-focused comparator (Bashir, 2022; n = 684); larger corpora exist on adjacent but narrower objects—Liu et al. (2020) analyze 1,342 documents on the oil–stock dyad—so the claim is one about scope within a defined object, not about corpus size in energy-finance bibliometrics generally. On integration, and again within the set of reviews compared in Table 1, it is the first to treat the three assets simultaneously as one interdependent system rather than as separate dyads. On coverage, the 2010–2025 analytical window captures 832 documents published between 2021 and 2025 (77.0% of the corpus), the phase of fastest growth that earlier reviews, ending sooner, could not observe. Methodologically, these claims rest on a Louvain community detection applied to a co-occurrence network of lemmatized keywords, which lets the thematic structure surface on its own rather than through categories imposed in advance.

Theoretical framework

Work on spillovers and co-movements among crude oil, gold, and stocks has grown around three theoretical pillars that coexist uneasily: the safe-haven, directional-connectedness, and financialization hypotheses. They are best understood not as rival explanations of one phenomenon but as answers to successive questions about a single causal chain. Financialization concerns why commodity and equity prices became coupled at all; connectedness concerns how, in which direction, and at which frequencies shocks now travel across that coupling; and the safe-haven hypothesis concerns what the coupling leaves available to an investor seeking protection when it tightens. Spillovers and co-movements are the observable traces of that chain—spillovers denoting the directional transmission of shocks and second moments, co-movements the contemporaneous alignment of returns—and each pillar reads those traces with a different instrument. Each asks different questions and favors different methods, yet all three bear on the same empirical objects, which leaves the field conceptually dense but only loosely integrated.

The safe-haven hypothesis

Gold's safe-haven role has been studied at length. Baur and Lucey (2010) drew the now-standard distinction between a hedge (negative or zero correlation on average) and a safe haven (negative or zero correlation specifically during turmoil), a framework that has shaped almost everything since. Ciner et al. (2013), cited 537 times in our dataset, test hedge and safe-haven properties across several asset pairs, including gold–stocks and oil–stocks. The arrival of cryptocurrencies, Bitcoin above all, has since unsettled gold's claim to be the haven of reference: Bouri et al. (2017b), the most-cited paper in our corpus at 883 citations, asks whether Bitcoin hedges or shelters commodities, gold included.

The directional connectedness hypothesis

Diebold and Yilmaz (2012, 2014) reshaped the study of financial spillovers by deriving directional measures from forecast-error variance decompositions. Later refinements, TVP-VAR extensions (Antonakakis et al., 2020) and frequency-domain decompositions (Baruník and Krehlík, 2018), let researchers separate net transmitters of shocks from net receivers within multivariate systems. Its reach into commodity–equity systems is by now well documented (Mensi et al., 2017; Tiwari et al., 2020).

The financialization hypothesis

Tang and Xiong (2012) advanced the financialization hypothesis: from the mid-2000s, the wave of institutional money into commodity markets reshaped price dynamics and introduced correlations with equities that owe little to macroeconomic fundamentals. The argument has fueled a long debate over whether commodities and equities can still be treated as analytically separable, and it sits in unresolved tension with the other two views—with the safe-haven hypothesis (can gold be a safe haven and a financialized asset at once?) and with the connectedness hypothesis (how stable is the direction of spillovers once the market's composition shifts structurally?). The three hypotheses are not woven into a common theory; they run as semi-independent traditions that happen to share methods. Making that fragmentation visible is one of the quieter contributions of the mapping that follows, and it marks an open theoretical agenda.

What the empirical literature has established

Read as a body of evidence rather than a set of methods, the literature supports a few substantive regularities. Gold's protective role is real but conditional. Baur and Lucey (2010) find that gold acts as a hedge for stocks on average and as a safe haven in extreme negative episodes, but only over limited horizons: the protection fades within roughly fifteen trading days. Ciner et al. (2013) confirm the hedge property across asset pairs while showing that it weakens once exchange-rate risk enters the picture, and Beckmann et al. (2015) find that the hedge and safe-haven properties switch on and off with regime changes rather than holding uniformly across markets. The consensus, then, is not that gold protects, but that it protects under identifiable conditions—and that those conditions shift.

The evidence on oil–equity transmission points the same way. Volatility spillovers between commodity and stock markets intensified visibly after the 2008 crisis (Mensi et al., 2013), and directional measures show that the roles of shock transmitter and receiver rotate with market conditions rather than staying fixed (Diebold and Yilmaz, 2012, 2014). Refinements of that framework add two robust qualifications: connectedness is time-varying, spiking around systemic events (Antonakakis et al., 2020), and it is frequency-dependent, with short-run linkages behaving differently from long-run ones (Baruník and Krehlík, 2018). Dependence-structure work reinforces the practical message: co-movements are strongest exactly when diversification is most needed (Mensi et al., 2017; Tiwari et al., 2020).

Two open disputes complete the picture. Tang and Xiong (2012) document that index investment raised commodity–equity correlations from the mid-2000s in ways fundamentals cannot fully explain, which challenges the treatment of commodities as an independent asset class. And the claim that Bitcoin can replace gold as the haven of reference has not survived stress-testing: digital assets show haven-like behavior under moderate uncertainty (Bouri et al., 2017a,b) but shed it in systemic crises such as the COVID-19 crash (Smales, 2019; Conlon and McGee, 2020). Both disputes are visible in the maps that follow—the first as the persistent centrality of the financialization school, the second as the rise and consolidation of the crypto-haven front.

Integrating the three pillars

Read together, the three pillars describe one sequence rather than three separate literatures. Financialization supplies the structural precondition: once index investment turned commodities into a portfolio asset, the correlation between commodity and equity returns ceased to be governed by fundamentals alone (Tang and Xiong, 2012). Connectedness supplies the measurement apparatus for the resulting system, decomposing forecast-error variance into directional flows and separating net transmitters from net receivers at short and long horizons (Diebold and Yilmaz, 2012, 2014; Baruník and Krehlík, 2018; Antonakakis et al., 2020); spillover and co-movement estimates are the quantities it produces. The safe-haven hypothesis supplies the welfare-relevant outcome: whether the asset investors turn to under stress still decouples once the system tightens, which Baur and Lucey (2010) and Beckmann et al. (2015) show holds only conditionally, briefly, and asymmetrically across regimes. The two open disputes sit at the joints of that sequence rather than inside any single pillar. Whether commodities remain an analytically separable asset class is a question about how far financialization has displaced fundamentals before connectedness measures the result; whether digital assets can replace gold is a question about whether a new candidate satisfies the safe-haven condition inside a transmission structure that connectedness has already mapped. Reading the pillars this way also makes the empirical maps that follow interpretable as a test rather than a description: if the field had integrated the three, its citation and keyword behavior would show schools that overlap conceptually, whereas a field that keeps them methodologically adjacent but conceptually apart will show dense connection with weak separation—which is precisely the pattern the modularity values in Sections 3.4 and 3.7 return.

Objectives and review questions

Five objectives follow from the gaps above. The first is to trace the field's temporal evolution and mark its phases, separating growth from stabilization through compound annual growth rates (CAGR) and year-over-year rates. The second is to identify the journals, authors, countries, and institutions that structure its output. The third is to read its intellectual structure by triangulating keyword co-occurrence, author co-citation, and bibliographic coupling. The fourth is to map where production comes from geographically and to measure the regional deficits that stand out. The fifth is to surface the theoretical, methodological, and geographic gaps that follow, and to set out a research agenda anchored in the bibliometric evidence.

These objectives are operationalized through three review questions. Q1: how is the field structured intellectually, and which schools does the research community itself treat as its main references? Q2: which fronts are forming in the most recent period (2022–2025), and how do they sit relative to the established traditions? Q3: which geographic, thematic, and methodological asymmetries shape the dominant agenda, and which of the resulting gaps actually matter?

Study design and reporting guidelines

This study is a systematic bibliometric review. It pairs performance analysis with science mapping, following Donthu et al. (2021). The combination lets quantitative measures of output and impact sit alongside a structural reading of the conceptual and collaboration networks that give a field its shape (Aria and Cuccurullo, 2017). The design is quantitative throughout: it works from objective bibliometric metadata, not from qualitative content analysis or subjective reading of the texts. The review is reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) statement (Page et al., 2021), adapted to the bibliometric review design following the guidelines of Donthu et al. (2021); the PRISMA 2020 flow of identification, screening, eligibility, and inclusion is reported in Section 3.1 (Figure 1).

Review question and eligibility framework (PICOS)

Because the units of analysis in a bibliometric evidence synthesis are bibliographic records rather than participants, the standard PICOS framework was adapted to define the review question and the eligibility of documents (Table 2). The framework is used here as an eligibility scaffold rather than as a claim that bibliographic records behave like clinical populations, and its methodological value added is specific: it forces every component of the review question to terminate in a decision rule that can be applied to a record before any analysis is run, and it allows PRISMA 2020 items 5–8 to be reported against those rules instead of against a narrative description of the search. Concretely, the population fixes what counts as an admissible record; the exposure fixes the analytical content a record must display to be topically relevant, and is therefore the component the search equation operationalizes term by term (Section 2.5); the comparators fix the contrasts the synthesis is designed to support, and so determine which metadata fields must be complete—year, source, country of affiliation, keywords, references—for a record to be usable at all; and the outcomes fix the bibliometric quantities to be extracted, which is what makes the analysis plan auditable in advance rather than reconstructed afterwards. Where a component would otherwise have to be stretched, it is kept deliberately narrow: the comparator element does not construct control groups, it names the axes of contrast, and the outcome element does not measure effects on markets, it names the bibliometric quantities synthesized. Table 2 states the decision rule attached to each element. The population is the body of peer-reviewed scientific literature on the interdependence among crude oil, gold, and equity markets; the exposure of interest is the analytical treatment of spillovers, co-movements, connectedness, and hedge/safe-haven behavior among these assets; the comparators are the contrasts drawn across assets, periods, intellectual schools, emerging fronts, and world regions, as well as against prior bibliometric reviews; the outcomes are the bibliometric quantities synthesized (production growth, productivity, citation impact, intellectual and thematic structure, emerging fronts, geographic distribution, and research gaps); and the eligible study designs are empirical peer-reviewed journal articles applying quantitative or econometric methods to the triad.

Systematic review protocol

A review protocol was defined a priori, specifying the research question and PICOS framework (Section 2.2), the data sources and search equations (Sections 2.4–2.5), the eligibility criteria (Section 2.6), the study-selection and data-extraction procedure (Section 2.7), and the analysis plan (Section 2.8). The review was not registered in PROSPERO because that registry accepts only health-related systematic reviews and does not admit bibliometric or science-mapping reviews; in the absence of a discipline-specific registry, transparency is instead ensured by full reporting of the search equations, the deduplication procedure, and the network-analysis configuration, together with sensitivity analyses. Reporting follows PRISMA 2020 (Page et al., 2021) adapted to bibliometric reviews per Donthu et al. (2021).

Data sources

The search ran on January 27, 2026, across two databases with complementary coverage: the Web of Science Core Collection (SCI-EXPANDED, SSCI, and ESCI indices) and Scopus. Pairing them is well justified, since their coverage of financial research complements rather than duplicates (Martín-Martín et al., 2018; Pranckute, 2021), especially for regional journals and Early Access work.

Search strategy

The query was built from four conceptual blocks joined by AND: (1) gold and precious metals; (2) oil and energy commodities; (3) stock markets; and (4) terms for interdependence, spillovers, and co-movements. We adapted the syntax to each database's conventions while keeping the meaning equivalent.

Scopus syntax (TITLE-ABS-KEY field): TITLE-ABS-KEY[(gold OR “precious metal * ”) AND (oil OR petroleum OR “crude oil” OR “energy commodit * ”) AND (“stock market * ” OR “equity market * ” OR stock OR equit *) AND (spillover * OR “co-movement * ” OR comovement * OR correlation * OR connectedness OR volatility OR dynamic * OR interdependen * OR linkage *)].

Web of Science syntax (TS = Topic): TS=[(gold OR “precious metal * ”) AND (oil OR petroleum OR “crude oil” OR “energy commodit * ”) AND (“stock market * ” OR “equity market * ” OR stock OR equit *) AND (spillover * OR “co-movement * ” OR comovement * OR correlation * OR connectedness OR volatility OR dynamic * OR interdependen * OR linkage *)].

The four blocks are joined by AND and confined to the title–abstract–keyword field, so breadth within any single block does not translate into topical breadth in the result set: a record is retrieved only when a precious-metal term, an energy-commodity term, an equity term, and an interdependence term all appear in its indexed text. Generic terms such as stock, correlation, volatility, or dynamic therefore cannot pull records in on their own—a study of exchange-rate volatility, for instance, is retrieved only if gold or another precious metal, oil or another energy commodity, and equity markets are simultaneously present in the same record. The co-presence requirement is what performs the topical screening in this design, and it is applied by the database at identification rather than by the reviewers afterwards.

We set the search window at 2010–2026. The choice of 2010 is deliberate: it is the year of the field's two most-cited foundational papers (Baur and Lucey, 2010; Baur and McDermott, 2010), which marks the start of its modern thematic consolidation. The upper bound takes in the 2026 Early Access records available at search time (n = 4). Search window and analytical window are kept distinct throughout the manuscript: the analytical window is 2010–2025 (n = 1,077 documents), and the four 2026 records are excluded from every statistic that depends on publication year—annual counts, year-over-year rates, CAGR, the phase totals in Table 3, and the growth panel of Figure 2—while being retained in the analyses that do not, namely journal and author distributions, citation and normalized indicators, geographic distribution, author co-citation, bibliographic coupling, and keyword co-occurrence. Because those four records are 0.37% of the corpus, recomputing every share reported below on the 1,077 documents published through 2025 moves no value by more than 0.4 percentage points, so the treatment is immaterial to every conclusion drawn.

Eligibility criteria

Three inclusion criteria applied: (a) document type limited to Article and Article-Early Access, with proceedings, book chapters, editorials, notes, and reviews left out; (b) English-language publication, with a single Turkish record admitted because it carried bilingual title, abstract, and keywords and so did not disturb the English-metadata analyses; and (c) a complete bibliographic record (title, abstract, keywords, affiliations, and references). Records not meeting all three criteria were excluded at the identification stage.

Study selection and data extraction

The initial search retrieved 829 records from the Web of Science and 558 records from Scopus (n = 1,387). Deduplication was performed in two steps. First, exact matching by normalized DOI identified 306 duplicate records. In the second step, the 29 documents without DOI were compared through fuzzy matching of normalized titles (Jaccard threshold ≥ 0.90) and cross-verification with the first author and year. This procedure identified no additional duplicates; the 29 DOI-less records corresponded to publications in regional journals (Romanian Journal of Economic Forecasting, Pacific Business Review International, African Review of Economics and Finance, among others) indexed in only one database. The final corpus thus comprised 1,081 unique documents, of which 829 originated from Web of Science and 252 were exclusive contributions from Scopus (23.3% of the final corpus).

Topical relevance was enforced ex ante, by the co-presence requirement built into the search equation together with the document-type, language, and date filters, all of which were applied inside the databases at the identification stage. Records lacking any one of the three asset families, or lacking an interdependence construct, were never retrieved and therefore could not reach the eligibility stage; consistent with that design, no record was excluded at eligibility and the PRISMA flow reports zero exclusions there. No additional manual title–abstract screening was applied on top of the retrieved set. The residual risk this leaves—a small number of records in which one of the three assets is treated only marginally, for instance as a control variable—is appraised in Section 3.8 and bounded in Section 4.6; its expected direction is dilution, since such records contribute weakly connected nodes that attenuate the clusters rather than create them. The 1,081 documents that remained form the final corpus: 1,030 regular articles and 51 Early Access articles (4 from 2026 and 47 from 2025). For each included document, structured metadata were extracted directly from the database exports: title, abstract, author keywords, authors and affiliations, source title, publication year, document type, cited references, and citation counts. Data extraction was performed programmatically from the exported records to avoid manual transcription error; the extracted fields feed the performance and science-mapping analyses described below.

Data analysis

The analysis was carried out in Python 3.12.3. Louvain and Leiden were run in parallel on the same weighted network to address Louvain's known sensitivity to poorly connected communities (Traag et al., 2019), with Leiden retained as a robustness check. Performance analysis covered publication output, source distribution, author productivity, geographic distribution, and citation impact. Science mapping comprised four complementary components: (i) a co-occurrence network of authors' keywords, (ii) author co-citation, (iii) bibliographic coupling restricted to 2022–2025 to detect emerging fronts, and (iv) country-level co-authorship with centrality computations.

Communities were detected with the Louvain algorithm (Blondel et al., 2008) on a weighted co-occurrence network, under explicit settings: resolution γ = 1.0 (the standard value); random seed = 42 (with stability checked at seeds 100 and 2024); 100 iterations (convergence confirmed); edge weight equal to co-occurrence frequency; a minimum node frequency of 8 mentions; and a minimum of 3 co-occurrences per edge. Before building the network, author keywords were lemmatized with a verified dictionary (16 mappings) and JEL codes were dropped. Robustness was checked two ways: (i) sensitivity to resolution and seed (nine configurations across γ ε {0.8, 1.0, 1.2} and three random seeds), which kept the four-cluster solution stable at γ = 1.0; and (ii) replacing Louvain with Leiden (Traag et al., 2019), which returned a comparable four-cluster partition (Q = 0.238; NMI = 0.74 against the Louvain solution).

Three parameters govern the bibliographic coupling solution and are reported here in full: the subcorpus (documents published 2022–2025), the reference floor (at least 10 cited references, which removes records whose reference lists are truncated in the export and cannot support a coupling measure), and the coupling threshold (at least eight shared references between two documents, set so that the network stays sparse enough for Louvain to separate communities while retaining a main component large enough to be informative). Community detection on the coupled network used the same Louvain configuration as the co-occurrence analysis (γ = 1.0, seed = 42, 100 iterations). Unlike the co-occurrence solution, the coupling solution was not re-estimated across alternative values of the reference floor or the coupling threshold; the fronts reported in Section 3.6 are therefore conditional on the two values stated here, and the smallest of them are the most exposed to that choice.

Study selection and PRISMA flow

The identification, screening, eligibility, and inclusion of documents followed the PRISMA 2020 flow (Page et al., 2021), adapted to the bibliometric review design following Donthu et al. (2021). Of the 1,387 records identified across Web of Science (n = 829) and Scopus (n = 558), 306 were removed as exact DOI duplicates and no further duplicates were detected among the 29 DOI-less records; the 1,081 unique documents that remained constitute the final corpus analyzed below. Figure 1 lays out the complete flow.

Study characteristics: temporal trends in scientific production

Figure 2 traces scientific production from 2010 to 2026. Every rate and phase total in this section is computed on the 1,077 documents published between 2010 and 2025; the four 2026 Early Access records enter the corpus total and the truncated 2026 bar of Figure 2A only. The path is one of fast growth giving way to a recent plateau, which is better captured by a handful of complementary metrics than by any single summary number. The compound annual growth rate (CAGR) over the 2010–2024 window—the last year with consolidated output at the time of the search—is 27.4%, computed as ((n 2024 / n 2010)(^1/14) – 1) × 100, with n 2010 = 6 and n 2024 = 178. Extending the window to 2025, the CAGR drops to 25.1%, reflecting the incipient deceleration in the field. The year 2026 was excluded from the rate calculations because it contained only Early Access records (n = 4) that did not represent consolidated output and would, if included, contaminate the metric through truncation. This correction follows the recommendations of Bornmann and Mutz (2015) for reporting growth rates in the presence of incomplete temporal cut-offs.

Reading the series year by year reveals three phases. A foundational phase (2010–2015) holds just 36 documents (3.3% of the corpus), with output low and erratic. A consolidation phase (2016–2020) holds 209 (19.3%), a sixfold rise. An expansion phase (2021–2025) holds 832 (77.0%), set off by a 92.8% jump between 2020 and 2021, the first full pandemic year. The plateau from 2023 onward is something earlier reviews could not have seen: year-over-year rates drop to 4.1% (2022 → 2023), 0.0% (2023 → 2024), and −3.4% (2024 → 2025). The field looks to have reached a mature phase once the pandemic impulse faded, leaving open whether the next shock—geopolitical, energy-related, or technological—reignites expansion or the domain settles into qualitative consolidation.

Journals, authorship, and citation performance

The 1,081 documents appeared in 254 journals. Figure 3A shows the 15 most productive, which together carry 518 documents (47.9% of all output). Resources Policy leads with 77 documents (7.12%), and Energy Economics follows with 72 (6.66%)—the two flagship outlets of resource and energy economics, a first indication of how deeply the field is anchored in energy-market questions. A Bradford's Law reading places the core zone at seven journals (2.76% of all titles), which hold 354 documents (32.75% of the total).

Figure 3B shows the 15 most prolific authors. Kang S leads with 52 documents, ahead of Mensi W with 49 and Bouri E with 31. Bouri's reach is exceptional for his count: his 2017 Bitcoin paper (Bouri et al., 2017b) drew 883 citations. Collaboration runs high, at 3.11 authors per document on average, and 90.8% of papers are multi-authored.

Across the 1,081 documents there are 31,996 citations, a mean of 29.60 per paper. The distribution is sharply right-skewed (median = 9; SD = 61.30; maximum = 883). Most work has landed: 935 documents (86.5%) have been cited at least once, while 146 (13.5%) remain uncited.

Normalized measures fill out the picture. The corpus reaches an h-index of 87 (87 documents with at least 87 citations each), a g-index of 140, an m-index of 5.44 (the h-index over the field's active years, a marker of how recent its consolidation is), and 5.93 citations per document per year. Skew dominates here too: the 90th percentile sits at 72 citations and the 99th at 329, while one paper (Bouri et al., 2017b) alone gathers 883, or 2.76% of all citations in the field.

An MCP/SCP ratio of 6.19, with 929 documents (85.9%) authored across countries against 150 (13.9%) from a single one, marks a field that collaborates internationally as a rule. That tempers the picture of geographic concentration in Section 3.5: output may cluster in a few countries, but most papers still join institutions from several.

Intellectual structure: author co-citation

Keyword co-occurrence (Section 3.7) reads the field's thematic structure from the terms authors attach to their own work. That structure need not match the intellectual structure—the theoretical and methodological traditions researchers themselves treat as reference points—which only author co-citation can recover (Small, 1973; Zupic and Cater, 2015). Cited references pulled from the 1,081 documents yield 9,169 distinct cited authors. Setting a 30-citation floor leaves a network of 210 core authors linked by 14,738 edges, on which Louvain (γ = 1.0) resolves three intellectual schools (Figure 4).

School A (hedge and safe haven, classical GARCH) is anchored by Baur (415 citations), Reboredo (252), Engle (186/181), Ciner (163), Arouri (158), Kroner (152), and Sadorsky (134). It carries the seminal tradition of conditional-volatility modeling applied to gold's hedge and safe-haven roles, tracing back to Baur and Lucey (2010).

School B (connectedness and spillovers) is anchored by Mensi (416 citations), Diebold (321), Antonakakis (227), Balcilar (189), Tiwari (185), Kang (155), Pesaran (149), and Naeem (148). It forms around the directional-connectedness framework of Diebold and Yilmaz (2012, 2014) and its refinements (TVP-VAR, frequency connectedness), and it is the dominant methodological stream in recent work.

School C (cryptocurrencies, COVID, and alternative havens) is anchored by Bouri (370 citations), Shahzad (233), Corbet (192), Ji (192), Salisu (169), Baker (120), and Akhtaruzzaman (117). The most recent of the three, it carries the debate over whether digital assets can stand in for gold as a haven through systemic and geopolitical crises. The strongest co-citation pairs—Baur–Mensi (233), Diebold–Mensi (221), and Bouri–Mensi (215)—show Mensi bridging all three schools, in keeping with his standing as the corpus's most-cited author and his structural role in the field's consolidation.

Geographic distribution of production

Figure 5 maps where the work comes from. Counting is at the country level and follows whole counting: a document with authors affiliated in several countries is credited once to each of them, so country figures are participations, not documents, and they cannot be added across countries to obtain a share of the corpus. To keep that distinction visible, country and regional figures are reported below as participations per 100 corpus documents, and the share of unique documents for any group of countries is bounded above by the corresponding participation figure and below by the largest single-country value in the group. China leads with 205 participations (18.96 per 100 documents), India follows with 152 (14.06), and the United States is third with 113 (10.45). The field's Asian tilt is hard to miss: the five Asian countries among the ten most productive (China, India, Pakistan, South Korea, and Vietnam) record 576 participations, or 53.3 per 100 corpus documents, so that between 19% and 53% of the corpus carries at least one Asian affiliation—a range whose lower bound alone already exceeds any other single country in the field.

A distinctive finding is Tunisia's position with 110 participations (10.18 per 100 documents), which places it fourth in the corpus and constitutes an outlier within the African landscape. This Tunisian concentration is explained by the activity of a research group centered on the École Supérieure de Commerce de Tunis and systematic cooperation with institutions in France and Lebanon. This exception qualifies the diagnosis of African under-representation; the continent's presence comprises Tunisia (110), South Africa (27), Nigeria (26), Ghana (15), Morocco (7), and Egypt (6), for 194 participations, or 17.95 per 100 corpus documents—again an upper bound on the share of unique documents with an African affiliation, and one in which a single national cluster supplies 56.7% of the total.

Latin America and the Caribbean, by contrast, record only 19 participations, at most 1.76% of corpus documents, spread across Brazil (7), Mexico (5), Chile (3), Colombia (2), and Peru (2). The shortfall is theoretically costly: the region supplies large shares of several leading global commodities (copper, soybeans, coffee, silver, and oil) and lives with exactly the price volatility this literature studies. Its near-absence, then, signals not low exposure but limited say in framing the questions and categories that define the field.

At the country level, the co-authorship network (density 0.490 over the 25 most productive countries; 147 edges) splits into four collaboration communities (Figure 6). The United Kingdom (0.110), Pakistan (0.103), and Turkey (0.101) serve as the main bridges across them. Modularity is Q = 0.153, pointing to a densely linked collaboration space rather than one carved into isolated blocks.

Figure 5 and Table 4 are read together but answer different questions, and neither restates the other. Table 4 is the ranked inventory: it gives the exact participation count for each of the 25 most productive countries, assigns each to a region, and is the source for every country figure quoted in the text. Figure 5 is the spatial and comparative reading: panel A distributes those counts across the world surface, which is what makes visible the pattern no ranked list conveys—that entire commodity-producing regions are blank rather than merely low-ranked, and that the field's footprint follows research capacity rather than resource endowment—while panel B sets Latin American and Caribbean participation beside the region's weight in world commodity exports, which is the benchmark the representation ratio in Section 4.3 rests on and which appears nowhere in the table. The table supplies the values; the figure supplies the geography and the comparison.

Concentration of this kind is not only a descriptive fact about who publishes. Because the questions a literature asks follow from the markets its authors live with, a field produced overwhelmingly by research communities in commodity-importing or financially central economies will estimate the parameters that matter to an importer or an allocator—transmission from oil shocks to equity indices, the hedging value of gold inside a diversified portfolio—rather than those that matter to an exporter. Section 4.2 draws out what this implies for the external validity of the field's results, and Section 4.3 quantifies the resulting asymmetry.

Emerging fronts: bibliographic coupling 2022–2025

Co-citation carries a built-in lag, since it depends on accumulated citations and so struggles to detect fronts still forming (Boyack and Klavans, 2010). Bibliographic coupling sidesteps that problem by linking documents that share cited references. We restricted it to the 567 documents from 2022–2025 carrying at least 10 references; with a threshold of eight shared references, the main component held 403 documents and 3,261 edges (Figure 7). Louvain detection returned nine emerging fronts (Q = 0.404), the four largest of which together hold 81.1% of the coupled documents. The modularity of that partition indicates well-separated communities, but the solution remains conditional on the two parameters just stated—the reference floor and the coupling threshold (Section 2.8)—and the smaller fronts are correspondingly the most sensitive to them.

Front F1 (TVP-VAR connectedness and post-COVID portfolios) gathers 144 documents (modal year 2024; mean TC = 19.1), built around Connectedness, COVID-19, TVP-VAR, Gold, Spillovers, and Portfolio Diversification. This represents the natural evolution of School B, as identified in the co-citation analysis. Front F2 (Bitcoin, safe haven, and cryptocurrencies) holds 117 documents (modal year 2022; mean TC = 17.4), built around Bitcoin, COVID-19, Gold, Safe Haven, Cryptocurrencies, and Hedge. It maps most closely onto School C from the co-citation analysis.

Front F3 (green bonds and energy transition) holds 42 documents (modal year 2023; mean TC = 16.7), built around Gold, Oil, Volatility Spillover, Green Bonds, Connectedness, and Financial Markets. It does not surface in the keyword co-occurrence map because its core terms have not yet cleared the frequency threshold, which makes it a distinctive find of bibliographic coupling: a nascent stream on how traditional commodities interact with energy-transition assets. Being among the smaller communities in the partition, its detection is the most exposed to the coupling parameters, so it is reported here as an emerging signal awaiting threshold-sensitivity confirmation (Sections 4.5 and 4.6) rather than as a settled structural result. Front F4 (geopolitical risk and precious metals) holds 24 documents (modal year 2023; mean TC = 13.6), tracking the entry of the Russia–Ukraine war into the field and the wider geopolitical turn in commodity finance. The remaining fronts are smaller: connectedness and alternative investments (F5, 23 documents), COVID-19 spillovers and TVP-VAR extensions (F6, 17), exchange rates, oil prices, and stock markets (F7, 13), and two residual fronts (F8–F9, 23 documents combined).

Thematic structure: keyword co-occurrence analysis

Figure 8 shows the keyword co-occurrence structure built from the preprocessing in Section 2.8. The final network has 72 keywords and 241 edges, and Louvain resolves it into four thematic clusters with modularity Q = 0.231. The reported percentages count documents containing at least one keyword from a given cluster; since a document can belong to several clusters at once, the shares exceed 100%, the kind of overlap one expects in an interdisciplinary field. The strongest co-occurrence pairs are Crude Oil–Gold (45), Gold–Oil (45), Bitcoin–Gold (34), and Gold Price–Oil Price (31), confirming Gold as the most structurally central node in the field (170 mentions). Its position mirrors Mensi's in the co-citation network: both act as bridges among the field's separate traditions (Table 5).

Risk of bias and quality appraisal

Because the units of synthesis are bibliographic records rather than clinical or interventional studies, the standard instruments for appraising risk of bias in systematic reviews (e.g., Cochrane RoB 2 or ROBINS-I) do not apply. Instead, the risk of bias was appraised at the level of the evidence base and the bibliometric procedure, along seven dimensions specific to this design. First, database-coverage bias: reliance on Web of Science and Scopus, without Dimensions, Google Scholar, or regional databases (SciELO, Redalyc, Latindex), can under-represent regionally indexed work; this was made explicit and its likely direction (an over-statement of the Latin American deficit) is discussed in Section 4.6. Second, topical-precision bias: relevance was enforced ex ante through the co-presence requirement of the search equation rather than by manual title–abstract screening, so a residual set of records treating one of the three assets only marginally may remain in the corpus; the expected direction is dilution of the mapped structures rather than their creation, since weakly related records enter the networks as sparsely connected nodes. Third, language bias: the English-language filter (with one bilingual Turkish exception) may deepen the same regional asymmetry. Fourth, temporal-truncation bias: the search window (2010–2026) and the analytical window (2010–2025) are reported separately, and the four 2026 Early Access records are excluded from all statistics that depend on publication year, which avoids the artificial deflation that incomplete years introduce (Bornmann and Mutz, 2015); their removal from the remaining analyses would move no reported share by more than 0.4 percentage points. Fifth, deduplication and metadata bias: deduplication leaned mainly on exact DOI matching, backed by fuzzy title similarity for the 29 records without a DOI, an approach vulnerable to inconsistent DOIs across databases; residual country-assignment error from automated affiliation parsing was estimated below 2%. Sixth, citation-maturation bias: raw citation counts favor older documents, so normalized indicators (citations per document per year) and coupling analysis restricted to 2022–2025 were used to reduce the distortion. Seventh, algorithmic bias: community detection is sensitive to resolution and random seed, which was mitigated by nine sensitivity configurations and a Leiden replication (NMI = 0.74); for bibliographic coupling the corresponding sensitivity to the reference floor and the coupling threshold was not tested, and the fronts are reported as threshold-conditional. Taken together, the appraisal indicates that the principal residual risk is coverage- and language-driven under-representation of Global South output, a bias whose direction is known and bounded, rather than a threat to the internal validity of the mapped structures.

Summary of main findings

Between 2010 and 2025 the literature on spillovers and co-movements among crude oil, gold, and equity markets grew quickly and unevenly, and analyzing 1,081 documents lets us pin down that unevenness with some precision. The growth was steep, a 27.4% compound annual rate over 2010–2024, before flattening from 2023, with 77.0% of all output falling between 2021 and 2025. Impact is substantial: 31,996 total citations, 29.60 per document on average, a corpus h-index of 87, and a g-index of 140. Five Asian countries together record 53.3 participations per 100 corpus documents, while Latin America and the Caribbean contribute at most 1.76% of documents and Africa 17.95 participations per 100, most of them Tunisian. Triangulating co-occurrence (four thematic clusters), co-citation (three intellectual schools), and bibliographic coupling (nine emerging fronts) brings into view both the field's settled structure and its lines of escape, chief among them the emerging green bonds and energy-transition front.

Convergences and divergences with prior bibliometrics

Set against the eight bibliometrics from Section 1.1, our findings line up on three points and part ways on four. Where they converge: the centrality of Mensi, Bouri, and Diebold as field-shaping authors, which our co-citation analysis flags as bridges across the three schools, matches the influence rankings in Singh et al. (2023), Bashir (2022), and Anas (2025); that three reviews of partly different objects agree suggests these authors form the field's intellectual core in a broad sense. The dominance of the Diebold and Yilmaz (2012) framework as the reference method likewise appears in Singh et al. (2023) and Harikumar and Muthumeenakshi (2023). And the COVID-19 pandemic as a productive turning point shows up across every recent review covering the post-2020 years.

Where they diverge: the integrated oil–gold–stocks corpus (n = 1,081) is about 58% larger than the closest triad-focused comparator (Bashir, 2022; n = 684), while remaining smaller than dyad-focused corpora such as Liu et al. (2020; n = 1,342), which is the qualification that keeps the breadth claim comparable across objects. The longer window lets us document something no prior review in the comparison set reports: the field's leveling-off from 2023 onward; studies that close in 2021 (Singh et al., 2023) or 2022 (Anas, 2025) simply cannot see the plateau. The coupling analysis also surfaces an emerging front—green bonds and energy transition—that earlier bibliometrics do not register. Finally, the corrected geographic reading exposes an overlooked pattern: Tunisia's concentration (10.18 participations per 100 documents, or 56.7% of the field's African total) is a striking exception to Africa's general under-representation.

Reading the three network analyses together—co-occurrence, co-citation, and bibliographic coupling—points to a methodological lesson for future work in this area. Most prior bibliometrics rely on a single technique, which hides the intellectual structure when only co-occurrence is reported and hides the emerging fronts when only co-citation is. The three are complementary by necessity, not by preference, if the field's full shape is to be captured. Recent systematic bibliometric work on financial-market risk contagion likewise confirms the value of bibliometric methods for tracing knowledge bases and research fronts (Su et al., 2024).

Read as a whole, the descriptive indicators are not a separate layer of the analysis but the mechanism behind the conceptual pattern the maps reveal, and three of them interlock. Journal concentration−47.9% of output in fifteen titles, with a Bradford core of seven—means that a small number of editorial agendas in resource and energy economics determine what counts as a publishable question, which is one reason the field's methods converge faster than its theory. Author concentration around a few highly productive teams, combined with an MCP/SCP ratio of 6.19, means that method diffusion travels through dense co-authorship rather than through independent replication: a TVP-VAR refinement propagates across dozens of papers within two or three years, which explains both the speed of School B's rise and the low modularity of the co-citation network (Q = 0.104), since schools that share people cannot separate cleanly. Geographic concentration then bounds external validity, because the markets studied are largely those the producing countries care about; the near-absence of Latin American and, outside Tunisia, African production is at once a gap in who asks and a gap in what is asked, which is why questions native to commodity exporters—stabilization funds, fiscal rules through boom–bust cycles, sovereign exposure to price volatility—remain outside the dominant agenda and appear in no thematic cluster. The plateau from 2023 acquires its meaning in the same frame: a field whose output is concentrated in a few outlets and a few teams, and whose method has stabilized, saturates its existing questions well before it exhausts the phenomenon. That is what makes a 42-document green bonds front analytically more consequential than its size suggests—it is the point at which the field's question set, rather than its technique, begins to change.

Quantifying the Latin American gap

Geographically, Latin America and the Caribbean supply at most 1.76% of the field's output. That number takes on weight once it is set beside a measure of the region's weight in the commodities under study that is expressed on the same basis. Per UN Trade and Development (UNCTAD) (2025), the region exported an annual average of US$678 billion in commodities over 2021–2023—South America US$536 billion, Central America US$129 billion, and the Caribbean US$13 billion—against world commodity exports of US$7,105 billion, so the region accounts for 9.5% of world commodity exports. It is also among the most commodity-dependent regions in the world: all twelve South American economies are classified as commodity-dependent, eleven of them with commodity shares above 80% of merchandise exports, and the region is a leading supplier of copper, soybeans, coffee, silver, and oil products.

To gauge the mismatch we define a scientific representation ratio whose numerator and denominator are shares of the same kind of global total: the region's share of the field's documents divided by the region's share of world commodity exports. For Latin America and the Caribbean it comes to 0.19 (1.76% / 9.5%), meaning the region helps frame the questions at roughly one-fifth of what its weight in world commodity trade would predict, and 1.76% being an upper bound on the regional document share, 0.19 is an upper bound on the ratio. Dividing the regional publication share by the share of commodities within the region's own exports (68%) would not yield a comparable quantity, since the two are shares of different totals, and is therefore not used. Africa supplies the instructive contrast: 6.6% of world commodity exports (US$467 billion) against 17.95 participations per 100 corpus documents, a ratio near 2.7—yet remove Tunisia's single national cluster and African participation falls to 7.77 per 100, a ratio close to 1.2. Regional presence in this literature tracks the existence of a concentrated research group far more closely than it tracks exposure to commodity price risk, which is what makes the Latin American absence a structural rather than an incidental feature.

Three forces help account for the gap. The first is infrastructure: R&D spending runs near 0.7% of GDP across Latin America, well below emerging Asia (about 2.3% in China). The second is language and visibility, since regional work on financial commodities in Spanish or Portuguese remains thinly indexed in Web of Science and Scopus (Pranckute, 2021). The third is the agenda itself: the dominant literature answers the concerns of global financial actors rather than those that arise from an exporting economy's structural commodity dependence, such as macroeconomic stabilization, countercyclical sovereign funds, or fiscal policy through boom–bust cycles. Beyond R&D spending, institutional capabilities for investment facilitation and digital government—highlighted by UN Trade and Development (UNCTAD) (2024)—may also condition the development of stronger regional research and data infrastructures.

Implications of the study

The implications of these results run along four connected lines—theoretical, methodological, practical, and specific to energy economics and green finance—and are taken in that order.

The field rests on three intellectual schools, but no shared theory binds them. The tension among the financialization hypothesis (Tang and Xiong, 2012), the safe-haven hypothesis (Baur and Lucey, 2010), and the directional-connectedness hypothesis (Diebold and Yilmaz, 2012) remains unresolved. Bibliometric mapping makes that integration vacuum visible, and the low network modularity reflects it (Q = 0.231 for co-occurrence; Q = 0.104 for co-citation): the field is densely connected in practice yet not integrated in its conceptual architecture.

Methodologically, the same complementarity yields a concrete recommendation for future financial bibliometrics: triangulating co-occurrence (thematic structure), co-citation (intellectual structure), and bibliographic coupling (emerging fronts) is what a complete reading of a field requires. Replicability, in turn, depends on spelling out the Louvain configuration (γ, seed, modularity, sensitivity analysis, and Leiden comparison) and on lemmatizing keywords systematically.

Practically, for portfolio managers, the coupled evidence around Front F2 (Bitcoin, safe haven, and cryptocurrencies) suggests that digital assets substitute for gold only conditionally: they behave like havens under moderate macroeconomic uncertainty but shed that property in systemic crises (Smales, 2019; Conlon and McGee, 2020). For regulators, Front F3 (green bonds and energy transition) signals a coming reconfiguration in which energy-transition assets interact with traditional commodities in still poorly understood ways; frameworks built around a commodity–equity architecture will need to absorb this third dimension. For Latin American science agencies, putting a number on the representation deficit supplies a case for dedicated commodity-finance programs aimed at regionally relevant questions.

The results also carry implications specific to energy economics. Crude oil is the corpus's energy anchor: the two leading outlets are the flagship journals of resource and energy economics (Resources Policy and Energy Economics; Figure 3A), and the field's central question—how energy-price shocks transmit to equity markets and haven assets—is at once a financial-stability question for energy importers and a fiscal one for energy exporters. The rise of Front F3 signals that this transmission mechanism is acquiring a third dimension: instruments tied to the low-carbon transition (SDG 7 and SDG 13). How transition risk reprices fossil-energy assets, and whether green bonds inherit, dampen, or amplify oil-driven volatility, are questions this literature is only beginning to pose.

Researchers moving into that front could draw on three tools the oil–gold–stocks literature has barely used: time-varying asymmetric copulas for tail dependence under climate- and market-stress events; quantile-connectedness and TVP-VAR frequency decompositions (Antonakakis et al., 2020; Baruník and Krehlík, 2018) restricted to left tails, to measure transition-risk spillovers from oil to green bonds; and multivariate GARCH with leverage effects to estimate the greenium and its covariance with gold as a parallel haven. Casting these models in accounting terms, under IFRS 9 (fair-value measurement) and IFRS S2 (climate-exposure disclosure), ties the empirical questions to corporate reporting decisions.

Identified research gaps and future directions

The analysis points to five gaps that frame a forward agenda: (i) the missing theoretical integration of the safe-haven, connectedness, and financialization hypotheses; (ii) the quantified Latin American under-representation (Section 4.3); (iii) the thin use of complementary mapping techniques; (iv) the loose articulation between traditional commodities and energy-transition assets (Front F3); and (v) the still-unexplained post-stabilization dynamics of 2023–2025, whose drivers—field maturity, the spent pandemic impulse, or reallocation toward other domains—warrant dedicated study.

Rather than a flat list of directions, the agenda that follows is organized into three priority programs, ordered by how much each would change what the field is able to claim.

Program 1—theoretical integration. This is the first priority because it addresses the constraint the maps show to be binding: a field densely connected in practice and unintegrated in its conceptual architecture. The task is to build a framework in which financialization supplies the structural coupling, connectedness supplies the transmission mechanics, and the safe-haven property appears as a state-dependent outcome of both, rather than three parallel literatures citing one another. It extends naturally to the 2023–2025 post-stabilization regime, whose drivers—field maturity, a spent pandemic impulse, or reallocation toward adjacent domains—bear directly on whether the existing framework has exhausted its questions.

Program 2—geographic and contextual diversification. This means building Latin American research around questions native to commodity-exporting economies (macroeconomic stabilization, countercyclical sovereign funds, fiscal policy through boom–bust cycles, sovereign exposure to price volatility) instead of importing the allocator's problem, and cross-checking against Global South databases (SciELO, Redalyc, Latindex) to establish how much of the measured gap is real and how much is a coverage artifact. Its payoff is not representational but substantive: it adds the state-contingent questions the dominant agenda does not pose.

Program 3—new methodological and thematic frontiers. This means developing the green bonds and energy-transition front through empirical work on its interaction with traditional commodities, confirming it first through a sensitivity replication of the coupling solution across alternative reference floors and coupling thresholds and through a focused qualitative review of its 42 documents; extending the toolkit to interpretable machine-learning methods and to the tail-dependence and quantile-connectedness tools identified in Section 4.4; and replicating the whole mapping on an extended corpus with the triangulated design used here. Within this program the sensitivity replication comes first, since the standing of the front conditions the value of everything built on it.

Limitations

A few limitations bound these conclusions. On coverage, the search ran only on Web of Science and Scopus, leaving out Dimensions, Google Scholar, and regional databases (SciELO, Redalyc, Latindex); part of the Latin American shortfall may therefore be an artifact of coverage rather than a true deficit. The English-language filter (with one Turkish exception) deepens the same asymmetry, and the four 2026 Early Access records are excluded from every statistic that depends on publication year, so the analytical window is 2010–2025 throughout while the corpus total of 1,081 refers to the 2010–2026 search window.

Second, on method, deduplication leaned mainly on exact DOI matching, backed by fuzzy title similarity for the 29 records without a DOI, an approach that is vulnerable to inconsistent DOIs across databases. The Louvain partition (Q = 0.231) reflects a densely interconnected field rather than a flaw in the analysis, and it holds up under nine sensitivity configurations and a Leiden replication (NMI = 0.74). Lemmatization ran through a verified 16-mapping dictionary; deeper conceptual synonyms were kept as separate terms, out of respect for authors' own lexical choices.

Two further method-side limitations bear directly on the results reported here. First, topical relevance was enforced entirely ex ante, through the co-presence requirement of the search equation and the database filters; no manual title–abstract screening was applied to the retrieved set, so a residual number of records in which one of the three assets is treated only marginally may remain in the corpus. Their expected effect is dilution rather than distortion—they enter the networks as weakly connected nodes that attenuate clusters instead of creating them—but a manual precision audit on a random sample would put a figure on that residual and is a straightforward extension. Second, the bibliographic coupling solution is conditional on two thresholds, a floor of 10 cited references and 8 shared references, whose effect on the number and size of the detected fronts was not tested. The modularity of the solution (Q = 0.404) indicates a well-separated partition, but the smaller fronts—the 42-document green bonds and energy-transition front among them—should be treated as threshold-conditional until a sensitivity analysis across alternative floors and thresholds confirms them.

On indicators, citation counts favor older documents; we report citations per document per year and lean on bibliographic coupling for 2022–2025, which is less exposed to such lags. Country assignment from automated parsing of the affiliation field was corrected for known special cases, with residual error estimated below 2%. More broadly, bibliometrics maps a field's formal structure from metadata, not the substance of its documents, so causal claims fall outside a descriptive design (Bornmann and Mutz, 2015); validating the emerging fronts identified here would call for a complementary qualitative review.

Conclusions

This systematic review shows that the oil–gold–stocks literature is at once quantitatively mature and conceptually fragmented: it grew at a 27.4% compound annual rate before flattening from 2023, it is anchored by three intellectual schools that share methods but not a common theory, and it is opening a new front on green bonds and the energy transition that earlier reviews did not register. The study's contributions are sixfold, each stated within the scope of the comparison assembled in Table 1. Within the strictly defined oil–gold–stocks triad it assembles the most extensive bibliometric corpus published so far, and among the reviews compared it is the first to treat the three assets as one interdependent system—claims about scope within a defined object, not about corpus size in energy-finance bibliometrics at large, where larger dyad-focused corpora exist. It documents a post-COVID stabilization of the field that reviews closing in 2021 or 2022 could not observe. It detects an emerging green bonds front through methodological triangulation, a front whose threshold-conditional status is set out in Section 4.6. It puts a comparable number on the Latin American representation deficit, a ratio of 0.19 against the region's 9.5% share of world commodity exports. And it sets transparency standards (PRISMA 2020 selection flow, documented lemmatization, sensitivity analysis, and Leiden comparison) that earlier work in the field has met only unevenly. For energy economics and green finance, the central open question is now how transition risk reprices fossil-energy assets and whether green instruments inherit, dampen, or amplify oil-driven volatility—a question this literature is only beginning to pose.