| Summary: | Predicting asset price trends is often posed as a classification problem, where trends are classified as positive or negative. Since asset price series are noisy and volatile, it is difficult to distinguish true trends from short-term fluctuations. To this end, several trend definitions have been proposed in the literature, but it is yet to be known how these trend definitions affect the performance of classification algorithms designed to learn such labels from historical data. In this paper, we define the robustness of the trend labeling algorithm as a measure of how well a classifier designed to learn such labels can withstand a change in the cumulative return considering the classifier’s generalization error. Moreover, we propose a noise model to simulate the desired accuracy score, which allows us to evaluate the robustness of a trend labeling algorithm without the need to train an actual classifier and consequently choose the optimal algorithm in terms of robustness. Experimental results confirm the adequacy of the proposed noise model and show that classification algorithms perform better when trained with such optimal labels.
|
|---|