Most concept induction algorithms process concept instances described in terms of properties that remain constant over time. In temporal domains, instances are best described in terms of properties whose values vary with time. Data engineering is called upon in temporal domains to transform the raw data into an appropriate form for concept induction. I investigate a method for inducing features suitable for classifying finite, univariate, time series that are governed by unknown deterministic processes contaminated by noise. In a supervised setting, I induce piecewise polynomials of appropriate complexity to characterize the data in each class, using Bayesian model induction principles. In this study, I evaluate the proposed method empirically in a semi-deterministic domain: the waveform classification problem, originally presented in the CART book. I compared the classification accuracy of the proposed algorithm to the accuracy attained by C4.5 under various noise levels. Feature induction improved the classification accuracy in noisy situations, but degraded it when there was no noise. The results demonstrate the value of the proposed method in the presence of noise, and reveal a weakness shared by all classifiers using {\em generative\/} rather than {\em discriminative\/} models: sensitivity to model inaccuracies.