Developing a Machine Learning based Systematic Investment Startegy: A case study for the Construction Industry

In this research work, an end-to-end systematic investment strategy based on machine learning models and leveraging the construction industry operational and management practices knowledge, is implemented. First, a literature research in the field of behavioral finance is done, presenting the current state of the knowledge and trends in the industry. A suitable investment opportunity exploiting prevailing market inefficiencies around earnings announcements is identified. Second, an extensive literature research is performed identifying the most relevant characteristics of construction companies’ operations and major risk factors they are exposed to. These insights are used to engineer a set of relevant variables. Third, advanced statistical techniques are used to select the most relevant subset of features, which includes market and analysts’ expectation data, macroeconomic indicators, the delay in reporting earnings, and the most important financial dimensions for construction firms. Fourth, the earnings’ surprise classification problem is characterized by a class imbalance and asymmetric misclassification costs. These issues are a consequence of the desired business application, and are addressed by selecting an appropriate evaluation metric. Additionally, considerations on the temporal dimension and generative process of the data are made to select an appropriate validation scheme. Five different state-of-the-art machine learning algorithms are considered: a multinomial logistic regression, a bagging classifier, a random forest, an XGBoost and a linear Support Vector Machine. The multinomial logistic regression is found to be the most suitable model, exhibiting a bias towards predicting positive earnings’ surprises over the rest of classes. The firm size, and the profitability and valuation measures, portrayed by the Return on Assets and Enterprise Value multiples, are found to be the most important variables when predicting earnings surprises. To conclude, the systematic investment strategy based on the investment signals produced by the selected machine learning model is back-tested, being the performance of the long-short portfolio driven by the positive surprise one as a consequence of the selected model bias. Keywords: Quantitative Investing, Machine Learning, Behavioral Financ