Building an investment strategy based on stock prediction | |
| Author | Pongraphe Kangsanarak |
| Call Number | AIT RSPR no.CS-25-02 |
| Subject(s) | Stock price forecasting--Data processing Deep learning (Machine learning) |
| Note | A research study submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Computer Science |
| Publisher | Asian Institute of Technology |
| Abstract | Short-term financial forecasting remains a central challenge in quantitative finance due to the noisy, nonlinear, and regime-dependent behavior of market data. Traditional sta tistical models and many early machine learning approaches often fail to capture the rapid changes, volatility clustering, and structural shifts that characterize real-world markets. These difficulties become more pronounced when comparing markets with distinct characteristics, such as the highly liquid and well-integrated U.S. market versus the more volatile and slower-adjusting Thai market.This research investigates these challenges by evaluating a range of deep learning architectures—including LSTM, CNN–LSTM hybrids, and ensemble methods—against the more recent N-HiTS framework in the context of short-horizon volatility forecasting. Instead of modeling daily returns, which exhibited near-random variation and produced consistently negative R2 scores, the study focuses on predicting changes in log-volatility. This target provides stronger theoretical grounding and demonstrates greater forecastability. Model outputs are incorporated into two rule-based strategies: a Moving Average–based volatility switching strategy and a Bollinger Band mean-reversion strategy, allowing direct assessment of whether improved prediction accuracy translates into better portfolio outcomes.The empirical findings reveal that N-HiTS delivers higher accuracy than LSTM-based models, particularly in the U.S. market where volatility regimes are more clearly de f ined. When integrated into trading strategies, N-HiTS enhances the performance of the Moving Average approach, although its impact on Bollinger Band–based trading is more limited. Differences in results across markets emphasize that structural proper ties—such as liquidity, volatility persistence, and speed of information diffusion—play a critical role in determining model effectiveness.Overall, the study shows that volatility is a more stable and informative forecasting target than returns, that the alignment between model outputs and trading rules is essential for generating practical value, and that market-specific dynamics significantly influence the success of deep learning–based forecasting systems. |
| Year | 2025 |
| Type | Research Study Project Report (RSPR) |
| School | School of Engineering and Technology |
| Department | Department of Information and Communications Technologies (DICT) |
| Academic Program/FoS | Computer Science (CS) |
| Chairperson(s) | Chaklam Silpasuwanchai; |
| Examination Committee(s) | Chantri Polprasert;Attaphongse Taparugssanagorn; |
| Scholarship Donor(s) | AIT Scholarships; |
| Degree | Research Studies Project Report (M. Eng.) - Asian Institute of Technology, 2025 |