Price movements in financial markets are well known to be very noisy. As a result, even if there are, on occasion, exploitable patterns that could be picked up by machine-learning algorithms, these are obscured by feature and label noise rendering the predictions less useful, and risky in practice. Traditional rule-learning techniques developed for noisy data, such as CN2, would seek only high precision rules and refrain from making predictions where their antecedents did not apply. We apply a similar approach, where a model abstains from making a prediction on data points that it is uncertain on. During training, a cascade of such models are learned in sequence, similar to rule lists, with each model being trained only on data on which the previous model(s) were uncertain. Similar pruning of data takes place at test-time, with (higher accuracy) predictions being made albeit only on a fraction (support) of test-time data. In a financial prediction setting, such an approach allows decisions to be taken only when the ensemble model is confident, thereby reducing risk. We present results using traditional MLPs as well as differentiable decision trees, on synthetic data as well as real financial market data, to predict fixed-term returns using commonly used features. We submit that our approach is likely to result in better overall returns at a lower level of risk. In this context we introduce an utility metric to measure the average gain per trade, as well as the return adjusted for downside risk, both of which are improved significantly by our approach.