Forecasting returns on financial assets has been an important task throughout the history of the financial economy. This study employs machine learning (ML) techniques to predict portfolio returns based on the size factor, aiming to not only improve predictions but also understand the underlying source of predictability. Amid the challenge of identifying relevant predictors in noisy data, this research employs a rolling window approach, incorporating three lags of stock returns as candidate predictors to project returns one minute ahead. Benchmark models, including in-sample averaging and autoregressive approaches, are explored alongside ML techniques such as Ridge, LASSO, AdaLASSO, and Random Forest. We consistently identify the superiority of ML models over benchmark models in terms of predictability, with the Random Forest model standing out as the most effective. Furthermore, analysis of the predictors selected by the models revealed that they are predominantly unexpected, short-lived and sparse.