Date of Degree


Document Type


Degree Name





Christos Giannikos

Committee Members

Thom Thurston

Chun Wang

Subject Categories



stock market, machine learning, yield curve, trading


This dissertation consists of three chapters.

Chapter 1: Behavioral heterogeneity among investors has been shown to explain the volatile nature of stock markets. In this chapter, I investigate the different behaviors of investors by proposing a heterogeneous agent model based on Chiarella et al. (2012) which involves fundamentalists, chartists, and noise traders with two-state hidden-Markov regime switching expectations. By applying the S&P 500 and CPI data from January 1990 to December 2020, the model shows strong evidence of behavioral heterogeneity among different groups of traders. After an in-sample backtesting and out-of-sample forecasting which further evaluate the capability of the model, two simple trading strategies are designed, both of which imply that the trading performance is better than the S&P 500 index.

Chapter 2: Stock price movement prediction is challenging among researchers because of non-stationary nature of the data. In recent years, machine learning models have become increasingly popular in predicting stock markets. In this chapter, using training data from 01/2012 to 12/2017 and test data from 01/2018 to 12/2018, I predict the daily and weekly average price movement of S&P 500 constituents and compare the prediction accuracy using five machine learning models: Artificial neural network, Naive Bayes classifier, Support vector classifier ensemble, Random forest, and Boosted decision trees. For the input features of each stock, LASSO penalized logistic regression is performed to extract the top 5 features from all the 65 technical and macroeconomic indicators. Experimental results show that Naive Bayes classifier outperforms other models, and weekly average price is more predictable than daily price.

Chapter 3: Yield curve modelling and forecasting is crucial for investment management. Using monthly yield curve data from 2003 to 2021, I investigate the forecasting performance of the yield curve using dynamic Nelson-Siegel models and compare with several alternative models. Next, I use the best performing model to evaluate the impact of training sample size on forecasting accuracy. Finally, after selecting the best model and the best sample size, I design two trading strategies using four different stock selection methods and compare the returns with two benchmarks: S&P 500 index and buy and hold strategy. Results show that, using the best sample size and Nelson-Siegel factors state space model which performs best in forecasting, the trading signals generated by yield curve prediction can be used as strategies to achieve higher average returns but lower Sharpe ratio than their benchmarks.

Included in

Finance Commons