Cum laude for thesis 'Combining Strategies Efficiently' from Wouter Koolen

Computer programs advise on stock market investments

Computer programs advise on stock market investments

Machine learning algorithms are of crucial importance in our daily lives. They manage our financial portfolios and provide us with personalized movie and book recommendations. Wouter Koolen from the Centrum Wiskunde & Informatica in Amsterdam (CWI) studied in his dissertation the theory of online learning, a branch of machine learning, that focuses on high quality decisions from conflicting expert advice. On 27 January 2011 he defended his PhD thesis ‘Combining Strategies Efficiently’ at the University of Amsterdam and graduated with honours - 'cum laude'. The results of his research are for instance of potential interest for financial and climate experts and could be applied in industrial processes.

Computer programs that make the best decision

Every day we base many practical decisions and actions on opinions of experts. We rely on the weather forecast for planning a weekend trip and we consult financial experts before investing on the stock market. Also devices, like hybrid cars, integrate several ‘expert policies’ to switch between the electric motor and gas motor. Unfortunately, experts often disagree and it is hard to find the most reliable forecast or prediction. In his research Koolen developed computer programs that automatically learn to make the best decisions by combining the opinions of all available experts. Moreover, his programs learn when to switch when the relative performance of experts changes over time.

Stock market investments

One of the practical strategies Koolen developed in his thesis recommends on investments in a simplified model of the stock market. Based on past prices, the strategy learns what the best financial advice is: either to buy or sell shares. An interesting feature is that Koolen’s strategy recommends to sell part of the shares when the price increases and, vice versa, recommends to buy shares when the price decreases. Koolen also proved mathematically that the final outcome of the decisions of his computer programs are as successful as what one could have earned with inside information, minus small learning costs. As a result, it is proved that the strategy earns lots of money when the share price fluctuates strongly.

Royal Holloway
Koolen is going to pursue his research at the Computer Learning Research Centre at Royal Holloway, University of London, with support of his Rubicon grant that he got recently awarded by NWO. He will focus in his future research on developing computer programs that learn when to switch between different experts over time, especially in case of taking ‘hard decisions’ (as in the case of hybrid cars). The Centre is internationally recognized for its machine learning research.


Interviews of Wouter Koolen about his dissertation:

Automatisering Gids, 21 January 2011

Folia, Edition 18, Volume 64

Intermediair, Edition 4, 2011

Listen here to the radio interview of Wouter Koolen in BNR DENKTANK! (BNR THINK TANK!)


Illustration: apbart