Should you even train a neural network?

Neural networks frequently dominate unstructured learning tasks but often fail to outperform simpler, non-neural baselines on tabular data—where ensembles such as Gradient Boosted Trees remain superior under equal budgets. This work intro duces ShouldYouEvenNN, a forecasting-based decision framework that determines whether continued neural training is rational given early evidence and computa tional constraints. Instead of performing exhaustive Neural Architecture Search (NAS), the method reframes model exploration as a continuation decision prob lem. The framework integrates three core mechanisms: (i) a shape-aware morphing forecaster that extrapolates early batch-level validation signals through a smooth blend of rational, sigmoid, and linear functions; (ii) a probabilistic scoring layer estimating the likelihood that a neural candidate will surpass a strong non-neural baseline; and (iii) a cost-aware expected-utility criterion that decides whether train ing should proceed. Experiments across seventy-one OpenML datasets show that ShouldYouEvenNN correctly identified neural-worthy tasks in over half of the cases, achieving validation accuracies within 5–6 percentage points of exhaustive NAS results (LCBench) while reducing runtime significantly. These findings demonstrate that early learning dynamics can provide sufficient statistical evidence to govern NAS decisions, transforming neural exploration from blind optimization into rational inference. ShouldYouEvenNN thus offers a prin cipled step toward evidence-based automation, enabling resource-efficient model discovery that decides not only how to train, but whether training is worth it.