A quantitative intraday trading strategy based on regression algorithms

My thesis is a quantitative intraday trading strategy based on regression algorithms. The main strategies used to generate trading signals for the stock market are buying and holding, intraday trading, scalping trading, weekly trading and so on. These trading strategies can be addressed by using technical analysis or fundamental analysis. Fundamental analysis attempts to use data such as revenue, expenses, growth prospects and competitive landscape to calculate the intrinsic value of stocks, while technical analysis uses past market activities and stock price trends to predict future activities. The fundamental uses a long-term approach of market analysis. Therefore, long-term investors often use fundamental analysis because it helps them to choose assets that increase in value over time. Technical analysis uses a relatively short-term method to analyze the market and use it for weeks, days or even minutes. So it is more commonly used by daily traders because it aims to select assets that can be sold to others at higher prices in the short term. The thesis work focuses on investigating the use of regression algorithms to generate trading signals for intraday stock trading. The proposed strategy separately analyzes the historical prices of one stock at a time and discovers patterns relevant for predict the future price of the stock in the next day. The proposed strategy automatically detects the right direction of investment (long- or short-selling) and decides whether to open an intraday trade or not on a given stock based on the expected profit. Therefore, potentially one trading signal per day is generated in case the potential profit of the recommended trade is sufficiently high. In this thesis work I conducted a large campaign of experiments on data acquired from the main Italian stock market index (FTSE MIB). To get significant results, we tested our trading strategies in different years (2011, 2013, and 2015), which correspond to different market conditions. We tested a variety of different algorithms in order to look for profitable trend forecasting models by analyzing the prediction performance and financial performance of the proposed models based on different algorithms, such as support vector machines, RepTree, Linear regression, NeuralNet and Random Forest. We tuned the configuration setting of each algorithm in order to achieve better prediction results. We tested also two baseline strategies: random signal generation and follow the trend of the last market day. For random strategy, I generate in Excel a new column where a random value (up, down, or no-action) is stored. Then, I compute the performance according to these values, not close price. The random value is the target that I evaluated. Baseline strategy is an simple strategy for regression that we developed. Specifically, the simple strategy is used to forecast the same direction as the one happened in the last day. The strategy is operated by constructing a sliding window, then the window will advance one day every time. For each sliding window, if you want to predict the trend (up, down or no action) of every stock on specific time point, you just need to evaluate the trend in the previous time point, then your prediction will be the same to the trend of the previous time point. By running experiments in different models and strategies, we found that, it is successful for day ahead forecasting of daily stock price movements by using these techniques. In addition, the prediction performance and the financial performance of the proposed models were verified and compared. The results show that The RepTree-based prediction model has higher prediction accuracy and financial performance than other models. We can conclude that prediction methods based on regression models in intraday trading can produce more efficient prediction systems than naive approaches.