GDP per capita and investment drive India’s surging steel demand, study finds

Steel consumption is often described as the backbone of modern industry, and few materials track a nation’s macroeconomic health as faithfully. As China’s steel demand passes its plateau and the world pivots toward a low-carbon transition, attention is shifting to India, now the fastest-growing steel market on Earth and the world’s second-largest producer. A new study published in Frontiers in Earth Science by Xiaoying Zhang, Fengyan Fan and Li Li of the Chinese Academy of Geological Sciences and China University of Geosciences, Beijing, sets out to answer a deceptively simple question with unusually rigorous methods: what actually drives India’s steel demand? The answer matters not only for forecasters of the global steel market but for governments and corporations betting billions on the next industrial superpower.

The research team assembled annual time-series data spanning 1970 to 2025, covering per capita apparent steel consumption alongside GDP per capita, the gross fixed capital formation rate, industrial value added as a share of GDP, the urban population growth rate and international iron ore prices. Rather than relying on a single analytical tradition, the authors constructed an integrated framework that combines classical econometrics with machine learning. At its core sits the Autoregressive Distributed Lag-Error Correction Model, or ARDL-ECM, a technique prized for its ability to disentangle long-run equilibrium relationships from short-run fluctuations within the same estimation. Two machine learning models, a quadratic polynomial ridge regression and a dynamic Support Vector Regression, were then deployed to cross-validate the econometric findings from the perspective of predictive performance and relative variable importance.

The methodological care begins before any coefficients are estimated. Because several candidate variables, including construction value added and industrial value added, are highly correlated with one another, the team ran variance inflation factor tests to detect multicollinearity. When all variables were entered simultaneously, VIF values exceeded the critical threshold of 10, so the authors selected industrial value added including construction as the main industrial proxy and built an alternative specification, replacing it with the sum of manufacturing and construction value added, to test whether conclusions depended on a particular indicator. Unit root testing with the ADF and KPSS procedures, supplemented by the Zivot-Andrews test that allows for one endogenously determined structural break, confirmed that every variable was integrated of order zero or one, with no I(2) series, satisfying the preconditions for ARDL modeling.

The bounds test delivered the study’s first headline result. For both the main and robustness models, the computed F-statistics of 5.796 and 5.573 exceeded the upper critical bound of 4.871 at the 1 percent significance level, rejecting the null hypothesis of no cointegration. In plain terms, India’s per capita steel demand is not drifting randomly; it is stably linked to economic development, investment, industrialization, urbanization and price factors in a long-run equilibrium. This finding guards against spurious regression, a notorious pitfall when non-stationary time series are regressed directly on one another.

Within that equilibrium, one variable towered above the rest. GDP per capita carried a long-run coefficient of 2.180, significant at the 1 percent level, meaning that a 1 percent rise in per capita GDP is associated with an average long-run increase of roughly 2.18 percent in per capita steel demand. Because both variables are expressed in logarithms, this coefficient reads directly as an elasticity. The gross fixed capital formation rate also showed a significant positive long-run relationship, with a coefficient of 0.035 at the 5 percent level, consistent with the steel-hungry nature of roads, railways, energy projects, housing and industrial construction. Industrial value added as a share of GDP was positive but statistically insignificant in the long run, a result the authors attribute to the indicator’s breadth, which mixes sectors of very different steel intensity.

Two relationships stood out for their sign. The urban population growth rate showed a negative conditional relationship with per capita steel demand, marginally significant at the 10 percent level in the long run and significantly negative in the short run. The authors are careful to stress that this indicator measures the annual growth rate of the urban population, not the level of urbanization, so it cannot be read as evidence that urbanization suppresses steel demand. Meanwhile, international iron ore prices carried a negative but statistically insignificant coefficient in every specification, suggesting that upstream raw material prices, which do not capture domestic steel prices, exchange rates or trade policies, have limited explanatory power for India’s final steel consumption.

Perhaps the most elegant finding concerns the error correction term. Estimated at minus 0.496 in the baseline model and minus 0.495 in the robustness model, both significant at the 1 percent level, it reveals a remarkably consistent adjustment mechanism: when India’s steel demand deviates from its long-run equilibrium level in a given year, roughly 49.6 percent of that deviation is corrected in the following year. Short-run shocks, whether from investment swings or industrial fluctuations, may jolt demand temporarily, but the system gravitates back toward the trajectory dictated by economic fundamentals. Short-run coefficients told a parallel story, with GDP per capita changes showing a significant positive effect of 1.231 and annual changes in investment and industrial value added also significantly positive.

The machine learning validation added an independent layer of evidence. Hyperparameters were selected strictly within the training period using an expanding-window time-series cross-validation procedure, with observations from 1970 to 2011 for estimation and 2012 to 2025 reserved as an untouched out-of-sample test set. The dynamic SVR, equipped with a linear kernel and a lagged steel demand term, achieved the strongest out-of-sample performance, with an R-squared of 0.81, a root mean square error of 5.337 kilograms per person and a mean absolute percentage error of 5.87 percent. The quadratic polynomial ridge regression, which expands the inputs into squared and interaction terms to capture nonlinearities, performed respectably but trailed the SVR on most metrics.

Variable importance rankings from the ridge model echoed the econometrics with striking fidelity. GDP per capita ranked first with a relative importance of 41.1 percent, the gross fixed capital formation rate second at 25.8 percent, together explaining 66.9 percent of the variation in steel demand. Iron ore prices ranked last at just 1.5 percent, mirroring their statistical insignificance in the ARDL-ECM. In the dynamic SVR, the lagged steel demand level dominated with 74.3 percent relative importance, underscoring the strong annual persistence of steel consumption, which evolves along existing trends tied to sustained economic activity rather than changing drastically from one year to the next.

The authors frame their findings as conditional statistical relationships and predictive importance rather than strict causal effects, noting that economic development, capital formation and steel demand may all be shaped by reforms, industrial policies and macroeconomic cycles. The implications, however, are hard to ignore. With India’s steel demand projected by the World Steel Association to grow around 9 percent in 2025 and 2026, and to rise by 75 million tons by 2026 compared with 2020, the study suggests that the trajectory of the world’s next steel giant will be written primarily by economic growth and investment activity, not by raw material prices. As China restructures and the global supply-demand system rebalances, the numbers behind India’s ascent are now anchored by evidence that survives both econometric scrutiny and machine learning cross-validation.

Subject of Research: Science

Article Title: Core driving determinants of India’s steel demand: evidence from ARDL-ECM and machine learning cross-validation

Article References: Zhang, X., Fan, F., & Li, L. (2026). Core driving determinants of India’s steel demand: evidence from ARDL-ECM and machine learning cross-validation. Frontiers in Earth Science, 14, Article 1883390. https://doi.org/10.3389/feart.2026.1883390

Image Credits: AI Generated

DOI: 10.3389/feart.2026.1883390

Keywords: Core, driving, determinants, India, steel, demand, evidence, ARDL-ECM, machine, learning, cross-validation, scientific research