Note: This content is accessible to all versions of every browser. However, this browser does not seem to support current Web standards, preventing the display of our site's design details.


Model predictive control of a stock portfolio


E. Zvizdic

Semester Thesis, HS13 (10321)

The goal of this semester project is to investigate the predictive power of machine learning methods, analyze the statistical relevance and significance of different types of features for prediction, and the possibilities and limitations of model predictive control when applied as an automatic portfolio investment strategy. Developing a successful algorithm required the implementation of methods for feature extraction and selection, a prediction algorithm and a receding horizon optimal investment procedure.

Firstly, we have formed a portfolio out of four major US indices, using the historical data from a publicly available database. The feature set we have selected comprises mostly of various technical indicators we generated from the data and external factors indirectly affecting the prices. Through feature selection and model selection, using cross validation for time series, we have obtained the optimal set of hyperparameters for our problem. The goal of the prediction step was to predict the rates of change of the stocks included in our portfolio. We have decided to solve a classification rather than a regression problem, as it proved to be more efficient and accurate. After prediction, we have employed a receding horizon controller to reallocate daily the portfolio wealth in an optimal fashion given risk constraints.

We have managed to predict the rates of return with a mean squared error ranging from 20-50% of the mean rate of change in our portfolio throughout the trading period. Oscillative technical indicators have proved to be the statistically most relevant features for the rate of change prediction. However, in high volatility conditions, we have shown that the inclusion of external factors like inflation, commodities prices and treasury bill rates is absolutely necessary in order to make profit. Model predictive control has outperformed the optimization based non-receding horizon predictor in terms of profit, demonstrating the benefits of future dynamics knowledge. In terms of earnings, our algorithm has managed to earn a 14.20% profit for 2011 (when most of the hedge funds experienced losses), and a 42.70% profit in the following year, thus topping the annual profits made by an average day trader, a random sampler that allocates wealth randomly, some prominent hedge funds and Swiss bank investment saving accounts. The most profitable strategy has turned out to be the one that emphasizes profit maximization over risk limitation, confirming that it is necessary to invest risky in order to earn money.

In light of these results, we believe it is worth trying the algorithm on the real daily trading market. However, in order to do so, it is necessary to invest further efforts into accuracy improvement and modifications to the algorithm that would make it suitable for successful day trading.

Supervisors: Michel Baes, Manfred Morari


Type of Publication:

(13)Semester/Bachelor Thesis

M. Morari

File Download:

Request a copy of this publication.
(Uses JavaScript)
% Autogenerated BibTeX entry
@PhdThesis { Xxx:2015:IFA_5141
Permanent link