<![CDATA[Purpose – This study re-evaluates a multivariate LSTM for forecasting oil palm fresh fruit bunch (FFB) yield at Marihat Plantation after an audit identified a data-export bug in the original dataset. The data were rebuilt and assessed using a leakage-safe chronological design, naive and moving-average baselines, and seed-sensitivity checks.Methods – Monthly FFB production, rainfall, and rainy-day records from plantation reports (2020–2023) were reconstructed and aggregated by planting-year cohort because block-level data were unavailable. Twelve-month windows across 11 cohorts yielded 240 sequences. With an 11-month purge gap, the chronological split comprised 10 training, 20 validation, and 10 test sequences. A two-layer LSTM (64 and 32 units; 20% dropout) was compared with five naive and moving-average baselines and retrained using 10 random seeds.Findings – The LSTM achieved MAPE = 29.73% (MAE = 0.27; RMSE = 0.34 ton/ha) on the test set, outperforming all baselines (MAPE = 74–118%). Across 10 seeds, mean MAPE was 31.79% (SD = 4.59%; worst = 40.60%), remaining below every baseline. However, all test sequences represented December 2023 across 10 cohorts. The observed advantage therefore indicates feasibility, not forecasting superiority across time.Research implications – The LSTM produced consistently lower errors than the baselines across cohorts for one target month, but the 10-sequence, single-month test set does not support general temporal claims. The prototype dashboard may support scenario exploration, although its 12-month forecasts depend on deterministic, unvalidated synthesized inputs.Originality/value – Prompted by a data-quality audit documented in Supplementary S1, this study transparently rebuilds the evidence base and reframes the model as a feasibility pilot rather than a validated solution. Longer, gap-free records covering multiple test months and a validated multi-step procedure are required to establish a dependable advantage.]]>
