This thesis investigates the use of path signatures as feature representations for derivative pricing within a regression-based framework. Motivated by the universal approximation theorem for signatures, which ensures that any continuous path functional can be approximated by a linear functional of its signature, the study evaluates the practical performance of this approach for a range of path-dependent derivatives. The analysis focuses on two research questions: how does signature-based regression compare with standard Monte Carlo simulation in terms of accuracy and stability, and whether more compact signature representations can alleviate the exponential growth in the signature dimension with respect to the number of assets. To address these questions, the work introduces an alternative representation of the signature features, the filtered signature, which uses the signatures of individual time-augmented asset paths rather than the full joint multi-asset signature, with the aim of reducing dimensionality while retaining essential path information. The effectiveness of this representation, along with the overall performance of the regression framework, is then evaluated through a series of numerical experiments. For the financial products analyzed, numerical results indicate that, when combined with analytically derived expected signatures, the price estimator obtained through signature-based regression exhibits markedly lower variance than standard Monte Carlo. As a result, the required number of simulated paths to achieve the same level of accuracy is reduced. The results further show that the filtered signature representation attains pricing accuracy and stability comparable to the full signature, while relying on significantly fewer terms. These findings suggest that essential and enough information about the path is encoded within the individual signatures of the time-augmented underlying assets, supporting the adoption of compact signature representations for multi-asset derivative pricing.