This thesis proposes a novel forecasting method that elaborates on the capability of integrating information measured at different frequencies and at irregular time intervals in financial markets. A data compression process is developed to take a whole range of financial time series and analyze their temporal information through multivariate dynamic kernels within a statistical machine learning algorithm, namely support vector machines. A number of dynamic kernels are designed to make the computational process more tractable without sacrifice on accuracy. Unlike most publications in the field, a complete analytical framework directly from the training data is provided for tuning hyperparameters. Experiments, based on predicting the S&P500 market, show promising results. Other potential applications of dynamic kernels are envisioned in such diverse areas as risk measurement, bioinformatics and industrial processes.