Now comes the critical moment: testing our model's performance against real-world data. We've deliberately withheld a portion of our dataset—the test data—to evaluate how well our model generalizes beyond the training examples it has already seen.
Think of this as administering a final exam. After teaching a student subtraction, we ask "What is 10 minus 6?" After training a model to distinguish cats from dogs, we present an image it has never encountered and ask for classification. This evaluation phase reveals whether our model has truly learned underlying patterns or simply memorized training examples—a crucial distinction that separates robust models from brittle ones.
Our test dataset contains 31 rows, manageable enough for manual inspection. Let's create a variable called `model_predictions` and assign it the output of `model.predict()`. This predict method, now available on our trained model object, represents the culmination of our training process.
Notice the critical difference in our approach here: we pass only the X_test features, deliberately withholding the Y_test target values. The model must make predictions based solely on input features, mimicking real-world scenarios where ground truth labels are unknown. This blind prediction process provides an unbiased assessment of model performance.
The resulting predictions certainly look promising at first glance. But appearances can be deceiving in machine learning—we need quantitative validation. Fortunately, we retained the actual Y_test values for precisely this comparison.
To facilitate side-by-side analysis, we'll convert the Y_test pandas series to a list format, matching the structure of our model predictions. This formatting consistency makes visual comparison more straightforward and reduces cognitive overhead when scanning results.
Initial inspection reveals a mixed performance profile typical of regression models. Consider the prediction of 26.6 versus the actual value of 31.39—a difference of roughly 15%, which falls within acceptable bounds for many business applications. Similarly, our model's guess of 16.6 against the true value of 19 represents reasonable accuracy.
However, not all predictions demonstrate such precision. The comparison of 14.69 to 22 reveals approximately 33% error—significant enough to warrant investigation. The fourth prediction shows even greater deviation, highlighting the inherent challenges in predictive modeling and the importance of comprehensive evaluation metrics.
Yet encouraging signals emerge from the data. The prediction of 39 compared to the actual 46 demonstrates solid directional accuracy, while 19.39 versus 19.58 achieves remarkable precision—less than 1% error. These variations underscore a fundamental truth: model performance rarely follows a uniform distribution across all test cases.
While manual inspection provides valuable intuition, professional model evaluation demands systematic measurement. Visual assessment, though instructive, introduces subjectivity and scales poorly with larger datasets. Fortunately, established statistical metrics can quantify prediction accuracy with mathematical precision, providing the objective foundation needed for confident model deployment.