Bayesian Networks Models for Equity Market

Financial markets evolve quickly due to the continuous innovation of investment tool and investors need to take the best decisions in the shortest time possible. This is why we propose in my thesis an innovative approach based on graphical models in order to provide to practitioners buy or sell indications on the American equity market (S&P 500). Generally, investors observe the market and consequently make a decision but this procedure is generally time consuming and do not always lead to a gain. This is why algorithmic trading is spreading through financial industry in the last years. In this work, we first introduced graphical models theory in order to explain how to read a directed or an undirected graph. Then, when Directed Acyclic Graphs (DAGs) have been introduced, we used Bayesian Networks (BNs) and Object Oriented Bayesian Networks (OOBNs) in order to build models model that are able to provide a recommendation in a mouse-click time. The software used for running our simulations is Hugin (www.hugin.com). The reason why we have chosen BNs and OOBN is that they allow showing clearly and intuitively dependence and independence relations among variables. Moreover, they deal efficiently with decision under uncertainty situations. In the first experiment used BNs for S&P 500 buy/sell signals detection. The results obtained underline the potentiality of BNs in the analysis of financial markets. Our results show that market efficiency does not only depend on financial news but also on information coming from other areas. Using a BN allows us to reveal dependences that otherwise would not be evidenced by the common tools used every day by financial operators. Furthermore, the results obtained show that the market equilibrium and drivers are changed across the last 21 years. In the last experiment we adopted OOBNs for detecting market signals. This decision allow us to overcome the limit showed by BNs in dealing with large and complex problems. OOBNs are able to combine at the same time different objects and sources of information and they exploits some of the basic elements of object oriented programming (i.e. objects, hierarchy…). Moreover, they allow to simplify a complex problem by dividing it in several blocks called instance nodes. These instance nodes communicate among each other through to the input and output nodes. Thanks to these features, a graphical representation appear less cluttered. The results obtained by using the OOBNs add more details to the ones obtained with the standard BN. We can immediately observe that there is not a single variable that triggers the buy/sell signal on the American equity market. By observing the results obtained with our model, we notice that they are consistent with the common financial knowledge. However, in some cases the network allows us to discover hidden dynamics and relations among variables that are not catched by the tools generally adopted. Thanks to these results, we can underline the importance of introducing a rigorous algorithmic approach when we analyze a complex framework such as the financial markets.