Stock trend prediction is a challenging task due to the market's noise, and machine learning techniques have recently been successful in coping with this challenge. In this research, we create a novel framework for stock prediction, Dynamic Advisor-Based Ensemble (dynABE). dynABE explores domain-specific areas based on the companies of interest, diversifies the feature set by creating different "advisors" that each handles a different area, follows an effective model ensemble procedure for each advisor, and combines the advisors together in a second-level ensemble through an online update strategy we developed. dynABE is able to adapt to price pattern changes of the market during the active trading period robustly, without needing to retrain the entire model. We test dynABE on three cobalt-related companies, and it achieves the best-case misclassification error of 31.12% and an annualized absolute return of 359.55% with zero maximum drawdown. dynABE also consistently outperforms the baseline models of support vector machine, neural network, and random forest in all case studies.
Stock trend prediction is an area of interest to researchers and investors alike, due to the complex patterns underlying the price data and the high profitability of successful trading strategies. In recent years, machine learning has become a popular technique for modeling the stock market. There are three quantitative approaches to stock prediction in general, each exploring different areas related to the stock market. The most common approach is based on general indicators, specifically the historical price and technical indicators. Such an approach relies on the traditional chartist theory that price patterns in the past will reoccur in the future [1]. The second approach is based on sentiment analysis, using natural language processing techniques to interpret text-based data like news articles. It is based on financial research such as He et al. [2] who show how investor sentiment influences stock returns. The third approach is based on the intercorrelation of corporations that uses information of other companies to predict the stock trend of one company, such as the recent work by Chen and Wei in 2018 [3].
In this paper, we present a novel stock prediction model, Dynamic Advisor-Based Ensemble (dynABE). There are four main contributions of our research to current works: the exploration of domain-specific information for high-frequency predictions; the establishment of an effective, first-level ensemble learning framework; the proposal of “advisors” for a second-level ensemble; and an online update strategy for dynamic flexibility.
First of all, instead of the three common approaches to stock prediction, this work is one of the few that explores the direction of commodity-stock relationship by incorporating domainspecific information. By “domain-specific information,” we mean the information that is related to the specific industry of a certain company. For example, the automobile market would be a type of domain-specific information for automobile producers, clean energy technologies for oil mining companies, and the consumer electronics market for technology companies. While fundamental analysis often explores the industry of specific companies to estimate their intrinsic values for long-term investments [4], few works use quantitative and high-frequency domain-specific information for short-term stock prediction.
Moreover, as we will later show in the literature review, no single machine learning model has been established to be superior for stock prediction. This calls for the need of ensemble learning, which combines the strengths of different models to compensate for one another’s mistakes when no single model is guaranteed to be most effective [5]. Our work presents an effective model ensemble framework that has a hybrid feature selection method and uses stacking to combine the base models.
In addition, we propose the concept of “advisors,” which is especially effective for the stock market. Specifically, we first define a number of domain-specific areas we want to investigate for a certain company. For each area, we find a pool of related features that will go through the previously defined ensemble learning framework to form one advisor. The multiple advisors are then combined to form a second-level ensemble.
The last innovation lies in the method we designed for combining these advisors, an online update strategy performed during the active trading period. Most current methods for stock prediction are static after the initial training. Therefore, they lack the flexibility to update themselves during the active trading period, making them vulnerable to the stock market dynamics-the market’s changes in price patterns may render a previously effective prediction model suddenly less accurate. In contrast, dynABE uses an online update strategy to dynamically weigh the advisors during trading. We will later elaborate the details of the online update. Intuitively, the use of advisors and online update ensures that all factors of the stock market that we wish to investigate, as defined in the formation of advisors, are available at all times during trading. Therefore, even if the price pattern of the stock market changes, dynABE is still able to adapt to the new pattern by changing the weights of the advisors. We show in our experiments that this additional dynamic flexibility of dynABE effectively increases its accuracy. In addition, since we do not need to retrain the base models with new data but only update the weights of the advisors, dynABE’s online update method is robost with few parameters.
We compare the performance of dynABE to three baseline models commonly used for stock prediction, namely support vector machine, neural network, and random forest. We show that dynABE consistently outperforms all the baseline models in all our case studies. We further use the predicted stock trends as trading signals on a naïve trading strategy to illustrate dynABE’s high potential profitability. It
This content is AI-processed based on open access ArXiv data.
