Seasonal-reference and regularization sensitivity with weak annual chlorophyll prediction

Luigi Caputi

Do bloom-phase heatwave exposures improve annual surface-chlorophyll anomaly prediction beyond
a nutrient-aware baseline? We compare two historical seasonal-reference constructions on fixed lateryear
evaluation cases, outcomes and inputs. The best pooled error on the original model and penalty
grid is less than 4% below zero-anomaly prediction, which predicts the calibrated seasonal-reference
integral; primary conditional bands for benchmark gains include zero, with coverage unvalidated.
Under fixed historical geometry, an apparent selected-procedure advantage depends on nearly tied
coarse-grid tuning, while full and baseline model errors are nearly identical at penalty 1000. A
retrospective one-standard-error rule selects 1000 for every pooled model under both constructions
and removes the large reversal; it is not a validated remedy. Competing troughs exchange rank and
move annual boundaries. Joint response and predictor substitution for ten extreme historical windows
is conditionally sufficient to reproduce the penalty switch; substituting their whole complement is
not. Cross-scoring shows prediction changes and target changes each reverse the tested penalty
pair. With original targets fixed, both errors increase; the weaker penalty deteriorates more. Target
substitution alone yields a near-zero reversal. Both constructions share a dense-grid minimum and
broad near-optimal region. Report calibrated benchmarks and paired selection losses, and justify
the scientific target before freezing preprocessing and selection for independent evaluation.