Portfolio Management via Empirical Asset Pricing Powered by Machine Learning

In this paper we propose an alternative machine learning
forecasting technique for the canonical problem of predicting expected
stock returns. The final goal is enhancing the financial performance of
the investment product, which in our case refers to a portfolio of equities. We adopt a combination of algorithms, capable of hand-ling highlevel abstraction, to study short- and long-term patterns emerging from
the analysis of financial factors and market signals. The core of the
model adopted to perform the prediction is composed of two independent
entities, analyzing short-term dynamics and capturing long-term trends
respectively. This adjustment helps us improve the predictive ability of
the model in a dynamic environment, where high volatility and noise
are intrinsic features. Lastly, we employ an ensemble algorithm that performs an intelligent weighting of each agent’s output. This method allows
us to identify the best stocks in terms of performance and to successfully implement quarter-long hold strategies that outperform the selected
universe’s equities return benchmark.