University of Twente Student Theses
ML techniques in portfolio management
Herbert, G.C. (2024) ML techniques in portfolio management.
|
PDF
3MB |
| Abstract: | It is generally assumed that with efficient markets, it is not possible to make accurate stock price predictions. With the rise and rapid development of ML models in recent years, this assumption might be proven wrong. ML models are known for their ability to find patterns in large amounts of data. Therefore, this pattern finding ability might be able to identify patterns in stock price data to predict future trends. The purpose of this thesis is to research whether Machine Learning (ML) models can be used to actively trade a portfolio of equities and bonds. Such an ML portfolio would allow for reduced management fees compared to those that financial institutions charge for actively managed portfolios as well as portfolios personalised to an investor’s preferences. To determine if ML models can be used for the aforementioned purpose, the ML portfolio should be able to outperform passive portfolios. To examine whether the ML portfolio is able to outperform passive portfolios, the risk-adjusted performance was compared against two benchmark portfolios consisting of a minimum variance portfolio and the market index tracker (following the NASDAQ Composite index). All three portfolios were constructed using the same stocks, as the NASDAQ Composite index was assumed to be the general market for the purposes of this thesis. During the experiment of this thesis, there were three testing periods to allow for more generalised conclusions on the performance of the ML portfolio. These testing periods were from January to December of the years 2000, 2010, and 2020. The minimum variance portfolio was constructed prior to the start of the testing period based on the ten-year prior stock price data and kept in its original form for the entire duration of the testing period. The market index tracker price was followed during the testing period without any adjustments. The ML portfolio was actively traded on a monthly basis during each testing period. Actively trading the ML portfolio meant that the ML portfolio was reconstructed for every month in every testing period. The reconstruction happened based on the Black-Litterman model that determined the optimal portfolio weights for the fifty companies with the highest expected returns. An Artificial Neural Network (ANN) was used as the ML model to predict the stock prices. These stock price predictions were then used to calculate the expected returns for the Black-Litterman model. The ANN made the stock price predictions based on the training dataset that consisted of ten years of prior stock price information. As the input variable, the market index tracker price from 31-trading days prior to the prediction date. After the minimum variance and ML portfolios were constructed, the performance was compared against the market index tracker based on (risk-adjusted) performance metrics, including the diversification index, Sharpe ratio, beta, Treynor ratio, and Jensen’s alpha. For the diversification index, the minimum variance portfolio outperformed the ML portfolio, both for the individual stock diversification as well as for sector diversification (respectively 95.25 compared to 92.51, and 83.05 compared to 78.06). This showed that the minimum variance portfolio construction methods created more diversified portfolios compared to the Black- Litterman model. For all the remaining risk-adjusted performance metrics, the minimum variance portfolio outperformed both the ML portfolio as well as the market index tracker. In particular, the Treynor ratio was notable as the minimum variance portfolio outperformed the market index tracker by having a five times higher ratio (0.21 compared to 0.04). This is notable as this shows that the market does not proportionally compensate risk by higher returns, as is one of the assumptions of the CAPM theory. This finding was further displayed by the Jensen’s alpha 3 where the minimum variance portfolio, on average, had an excess return of 5.58 percent. compared to the CAPM expected return. From the portfolio comparison it can be concluded that the ML portfolio was unable to outperform the benchmark portfolio based on the risk- adjusted performance. The performance of the ANN itself was measured based on the accuracy, R2, and error-based metrics. The directional accuracy of the ANN was averaged at 55.04 percent during the three testing periods. The level of accuracy that the ANN achieved during the experiment seems to be too low to construct a benchmark outperforming portfolio, as was observed from the portfolio comparison. The reason why the accuracy was low might be due to the independent variable as the R2 scores were close to zero in the last two testing periods, showing the chosen input had little ability to explain the variation in the stock prices. The final ML performance metric were the error-based metrics. Between the testing periods, there were differing orders of magnitude for the size of the errors (with median RMSE scores ranging between 2.70 in 2010 and 9.47 in 2020), which made it complicated to generate generalised conclusions. An interesting finding was in relation with the accuracy metric, during the 2010 testing period, the accuracy was at the lowest level of the experiment. In the same 2010 testing period, the error-based metrics were also at the lowest level, showing the lowest difference between the predicted and observed values. This showed that the accuracy might have been misrepresenting the performance of the ANN as lower errors between predicted and observed values better represent the required performance of the ANN. Concluding, the ML performance metrics showed that the ML model needs to be improved to increase the performance of the ML portfolio. One way to improve the performance of the ANN is by determining what the best independent variable would be. The current independent variable showed too little explaining ability for the stock prices, leading to low accuracy and higher errors. |
| Item Type: | Essay (Master) |
| Faculty: | BMS: Behavioural, Management and Social Sciences |
| Subject: | 85 business administration, organizational science |
| Programme: | Business Administration MSc (60644) |
| Link to this item: | https://purl.utwente.nl/essays/103853 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page