Reproducible pipeline, code, and supporting outputs for "Systematic Evaluation of Record Selection and Validation Reveals Sources of Uncertainty in Coral-Based ENSO Reconstructions"

This repository contains the reproducible analysis pipeline, source code, configuration files, supporting outputs, and Tables S1–S19 for the study “Systematic Evaluation of Record Selection and Validation Reveals Sources of Uncertainty in Coral-Based ENSO Reconstructions.” The materials support coral-record inventory and eligibility assessment, split-period validation, reconstruction experiments, sensitivity analyses, ENSO event diagnostics, generalized extreme-value analyses, and reproduction of the reported results. Table S1. Complete inventory of the 130 tropical Pacific CoralHydro2k δ18O and Sr/Ca series, including site and dataset metadata, source-record start and end years, median sampling interval, observed-month counts, within-year monthly-coverage diagnostics, separate indicators for the sampling-interval and within-year coverage criteria, missing-data fraction, temporal-eligibility status, and reasons for exclusion. Table S2. Sensitivity of temporal eligibility and leading-record selection to minimum full-period coverage requirements of 120, 240, 360, 480, and 600 observed months and minimum coverage requirements of 120, 180, and 240 observed months in each split period. The median-sampling-interval criterion of no more than 45 days and the requirement that at least 30% of years contain 10 or more observed months were held fixed. The table reports eligible-record counts by proxy type and the highest-ranked δ18O and Sr/Ca records and their minimum split-period verification correlations. Table S3. Temporally eligible δ18O records ranked by the prespecified hierarchy led by the minimum of the two split-period verification correlations. The table reports site metadata, record coverage, full-window skill, skill in both calibration–verification directions, and the minimum verification correlation. Table S4. Temporally eligible Sr/Ca records ranked by the prespecified hierarchy led by the minimum of the two split-period verification correlations. The table reports site metadata, record coverage, full-window skill, skill in both calibration–verification directions, and the minimum verification correlation. Table S5. Moving-block bootstrap comparisons of the minimum split-period verification correlation for Tarawa CO93TAR01 δ18O versus the next 11 evaluable δ18O records. The table reports the reference and candidate correlations, observed reference-minus-candidate differences, percentile 95% confidence intervals, whether each interval excludes zero, and common-month sample sizes in the two calibration and verification directions. Results are based on 2,000 replicates using blocks of 60 chronologically consecutive entries from each common-month series, approximately representing five years. Table S6. Moving-block bootstrap comparisons of the minimum split-period verification correlation for Palmyra NU11PAL01 Sr/Ca versus the next 11 evaluable Sr/Ca records. The table reports the reference and candidate correlations, observed reference-minus-candidate differences, percentile 95% confidence intervals, whether each interval excludes zero, and common-month sample sizes in the two calibration and verification directions. Results are based on 2,000 replicates using blocks of 60 chronologically consecutive entries from each common-month series, approximately representing five years. Table S7. ICOADS 2° sea-surface-temperature observation density in the Niño3.4 region during 1900–1949 and 1950–1990, including observation counts, occupied-grid-cell coverage, and observations per occupied cell. These values contextualize historical data availability and are not the exact observations retained or weighted by ERSSTv5. Table S8. Post-1950 sensitivity of the configurations selected in the main 1900–1990 analysis and those selected under the 1950–1990 ranking. Full-period correlation and root-mean-square error and bidirectional split-period correlation and root-mean-square error are reported over 1950–1990, using 1950–1970 and 1971–1990 as alternating calibration and evaluation intervals. The table includes the original Tarawa δ18O and Palmyra Sr/Ca records and pair, the highest-ranked post-1950 δ18O and Sr/Ca records, and the corresponding post-1950-ranked pair. Table S9. Sensitivity of record selection and reconstruction skill to the main anomaly definitions and to a common 1961–1990 anomaly baseline for both the coral records and Niño3.4. The eligible record set was held fixed while anomaly calculation, ranking, record selection, and reconstruction were repeated. The table reports the selected δ18O and Sr/Ca records, whether they match the main analysis, and full-window and bidirectional split-period skill. Table S10. Moving-block bootstrap comparison of the Tarawa δ18O + Palmyra Sr/Ca reconstruction with the Tarawa δ18O-only reconstruction. Observed two-record-minus-Tarawa differences and percentile 95% confidence intervals are reported for the minimum split-period verification correlation and maximum split-period root-mean-square error. The two models were fitted and evaluated on identical common months and refitted in each of 2,000 bootstrap replicates using blocks of 60 chronologically consecutive common-month entries, approximately representing five years; calibration and verification samples were resampled independently. Table S11. Sensitivity of two-record reconstruction skill to requiring at least one or both selected coral records in each month. Monthly correlation, root-mean-square error, and evaluation-month counts are reported for the full-overlap and bidirectional split-period experiments. Table S12. Bidirectional split-period skill of nested δ18O-only, Sr/Ca-only, and combined δ18O–Sr/Ca networks as successively lower-ranked records are added. The table lists network membership, skill in each verification direction, the minimum verification correlation, and the maximum verification root-mean-square error. Table S13. Common-month standard deviation and skewness of ERSSTv5 Niño3.4 and the variance-scaled Tarawa-only and two-record reconstructions during 1900–1945 and 1946–1990, with the number and fraction of expected months retained. Table S14. Comparison of ERSSTv5-derived seasonal Niño3.4 indices with the official NOAA Climate Prediction Center Oceanic Niño Index over 1950–1990. Correlation, error statistics, agreement after rounding to 0.1°C, and event-identification matches are reported for the exact implementation using 30-year base periods updated every five years and for a continuous 30-year running-baseline analogue. Table S15. CPC-style ENSO event catalog for 1950–1990 based on ERSSTv5 Niño3.4 and the raw and variance-scaled Tarawa-only and two-record reconstructions. The table reports the amplitude treatment and baseline method, event type, start and end seasons, number of consecutive overlapping seasons, approximate calendar duration, peak index, and peak center month. These events are retained as secondary reconstruction-sensitivity diagnostics rather than as primary estimates of historical event frequency. Table S16. Summary of CPC-style ENSO event diagnostics during 1950–1990 for ERSSTv5 Niño3.4 and the raw and variance-scaled Tarawa-only and two-record reconstructions. The table reports El Niño and La Niña event counts, median and maximum approximate durations, strongest peak indices, and counts and fractions of reconstructed and instrumental events with descriptive temporal overlap. Events were not paired one-to-one, and the overlap fractions are descriptive rather than formal classification-accuracy statistics. Table S17. Direct moving-block-bootstrap tests of late-minus-early generalized extreme value (GEV) return-level changes within each series for annual warm maxima and annual cold minima of the centered 3-month Niño3.4 mean. The table reports early- and late-period 5-, 10-, and 20-year return levels, observed late-minus-early changes, percentile 95% confidence intervals, two-sided bootstrap p values, and within-family Benjamini–Hochberg false-discovery-rate q values. Results are based on 2,000 replicates using 5-year blocks of common annual extrema; the same bootstrap indices were applied across the three series within each period, whereas the early and late periods were resampled independently. Table S18. Direct paired moving-block-bootstrap tests of between-series generalized extreme value (GEV) return-level differences within 1900–1945 and 1946–1990 for annual warm maxima and annual cold minima of the centered 3-month Niño3.4 mean. The table reports Series A minus Series B contrasts at 5-, 10-, and 20-year return periods, percentile 95% confidence intervals, two-sided bootstrap p values, and within-family Benjamini–Hochberg false-discovery-rate q values. The same 5-year-block bootstrap indices were applied to all series within each period to preserve covariance. Table S19. Direct moving-block-bootstrap tests of between-series differences in late-minus-early generalized extreme value (GEV) return-level change for annual warm maxima and annual cold minima of the centered 3-month Niño3.4 mean. The table reports Series A minus Series B contrasts in temporal change at 5-, 10-, and 20-year return periods, percentile 95% confidence intervals, two-sided bootstrap p values, and within-family Benjamini–Hochberg false-discovery-rate q values. Common annual extrema were resampled with shared indices across series within each period, whereas the early and late periods were resampled independently.