Dataset for "Demand Observability of Critical Minerals under Statistical Discontinuity: Evidence from Japan's Special Steel Industry"

This dataset supports the manuscript “Demand Observability of Critical Minerals under Statistical Discontinuity: Evidence from Japan’s Special Steel Industry.”

The study examines whether continuously available industrial production statistics can serve as monitoring-oriented proxy indicators for use-specific critical-mineral demand after the corresponding official demand statistics are no longer continuously available. The empirical case covers special-steel-use demand in Japan for six mineral or statistical categories: antimony (Sb), nickel (Ni), Ni alloy, cobalt (Co), tungsten (W), and molybdenum (Mo). “Ni alloy” is retained as a legacy statistical category from the original Japanese statistics and should not be interpreted as an independent policy-defined critical mineral.

The dataset contains processed annual mineral-demand series and hot-rolled special steel production series for eight steel-category groups: MCS, LAS, SUP, SUJ, SUS, SUM, SWP, and HTS. The source series cover FY1987–FY2019. The principal analysis links mineral demand in fiscal year t to steel-category production in fiscal year t−2, resulting in 31 paired annual observations for demand FY1989–FY2019 and production FY1987–FY2017.

The workbook includes:

- processed annual mineral-demand data;
- processed annual hot-rolled special steel production data;
- the main two-year-lag-aligned analysis dataset;
- Lag 0–3 screening results for all 48 mineral–steel-category relationships;
- ordinary least squares t-values, Newey–West HAC t-values, adjusted R-squared values, Durbin–Watson statistics, and Tier classifications;
- full Lag-2 screening results;
- expanding-window rolling-origin evaluations for 17 proxy models carried forward from the screening stage over 11 target years, FY2009–FY2019;
- annual forecast errors and model-level MAPE, MASE, and RMSPE results;
- centered-linear-trend sensitivity analyses; and
- Benjamini–Hochberg multiple-testing sensitivity analyses.

The Tier classification is used as an exploratory screening procedure rather than as a confirmatory hypothesis-testing hierarchy. Candidate relationships identified using the full-period data are evaluated retrospectively through an expanding-window rolling-origin procedure. The out-of-sample evaluation is therefore conditional on prior full-period screening and should not be interpreted as a fully independent pseudo-real-time model-selection exercise. The estimated relationships are intended as statistical monitoring signals and should not be interpreted as causal effects or direct physical mineral-input coefficients.

Version 3.1 updates Version 3.0 in the same Zenodo version series. The previous published version (Version 3.0) is available at https://doi.org/10.5281/zenodo.21553180.

Zenodo Version 3.1 — Correction note

Date
----
2026-09-06

Reason for update
-----------------
Version 3.1 corrects a predictor-reference inconsistency in the Co–SUJ
expanding-window rolling-origin evaluation contained in Version 3.0.

Affected results
----------------
The correction affects only:
- Sheet 8 (Yearly_forecast): Co–SUJ target FY2009–FY2019 predictions and the
  corresponding APE, ASE, and SPE values; and
- Sheet 9 (Model_summary): Co–SUJ MAPE, MASE, RMSPE, model-level assessment,
  and interpretation.

Corrected Co–SUJ model-level results
------------------------------------
Version 3.0:
- MAPE = 24.9797503352086%
- MASE = 1.927921073204356
- RMSPE = 28.518841016629498%
- Assessment = Weak

Version 3.1:
- MAPE = 18.52147079398997%
- MASE = 1.320979129469377
- RMSPE = 23.661664122736568%
- Assessment = Limited

Verification
------------

The correction was verified by independently recalculating the rolling-origin forecasts from the lag-aligned data in Sheet 4 without using the stored forecast results in Sheets 8 and 9. The recalculated predictions for all 17 evaluated models and 11 target years reproduced the corrected workbook results within numerical precision. The Co–SUJ MAPE, MASE, RMSPE, and model-level assessment were also independently reproduced. The Benjamini–Hochberg adjusted p-values for the 48 Lag-2 tests were independently recalculated and reproduced within numerical precision.

The correction does not alter:

• the processed mineral-demand and special-steel production input series;
• the Lag 0–3 screening results;
• the main Lag-2 screening and Tier classifications;
• the other 16 evaluated rolling-origin models;
• the Table S5 centered-linear-trend sensitivity results; or
• the Table S6 Benjamini–Hochberg results.

The mineral-level interpretation is unchanged. Co–SUM remains the best-performing
Co proxy by MAPE (17.28974740000374%), and Co remains classified as having limited
monitoring performance in the study framework.

Files for Version 3.1
---------------------
- analysis_dataset_and_results_v3.1.xlsx
- README_v3.1.txt
- Data_sources_and_processing_notes_v3.1.txt