Forward-testing mineral prospectivity models with discovery dates: Quebec graphite and nickel

Mineral prospectivity models are usually evaluated with spatial cross-validation, but known occurrences cluster around past exploration, so this may measure performance within the explored footprint rather than the ability to rank later discoveries. We trained seven prospectivity models on graphite and nickel occurrences known in Quebec, Canada, before 2000 and tested how they ranked cells with occurrences first recorded from 2000 onward. Spatial block cross-validation was compared with the cluster-weighted share of these discoveries captured in the highest-ranked 5% of eligible cells, across eight related settings per commodity (cutoff year, grid resolution, geophysical coverage, predictor dates and label definition). For graphite, whose later discoveries lay a median 88 km from earlier ones, cross-validation overestimated forward capture for at least six of seven models in every setting and in ten repeated spatial partitions. The proximity-to-known-occurrences baseline lost the most: 0.68 in cross-validation against 0.26 in the forward test, with non-overlapping intervals. For nickel, whose discoveries stayed near established camps (median 14 km), cross-validation and forward rankings agreed more closely. Which model cross-validation selected depended on whether never-worked showings were counted as discoveries. Fixing exploration effort at a common value when scoring never clearly improved graphite capture and consistently reduced nickel capture. Across primary-label settings, no model captured more than 18% of discoveries in cells without recorded prior exploration, a rate not clearly different from random selection. Where discovery dates exist, forward tests should accompany spatial cross-validation.