Accuracy alone may not show which model is more useful. It can conceal poor performance on a minority class or differences in the kinds of errors models make. Choose metrics that fit your research question, the outcome and the balance of your dataset.
Compare models under the same evaluation protocol: use consistent data preparation, splits, validation strategy and metrics. For a paired comparison, the predictions or scores should come from comparable observations, such as the same held-out test cases or matching validation folds. This makes the comparison fairer and helps avoid treating unrelated results as if they were paired.
Comparing predictive performance is not the same as establishing statistical significance. A score difference describes observed performance; a significance test assesses how compatible that difference is with chance under a particular design and its assumptions. The suitable test depends on the experimental design, number of models and evaluation setup, so no single test applies to every dissertation.
Interpret significance alongside practical importance. Where appropriate, report effect sizes, uncertainty and limitations, including the dataset size, validation choices and variability across runs. A small difference can be statistically detectable without making one model meaningfully better for the intended use.