Hua Tang

Hua Tang

Undergraduate @Industrial Engineering, Shanghai Jiao Tong University

Empirical Strategies for High-frequency Quantitative Trading

Jul.2022 - May.2023: Empirical Strategies for High-frequency Quantitative Trading, advised by Tongxin Ren, Assistant Professor, Student Innovation Center, Shanghai Jiao Tong University(SJTU).

Source code is available at link

Key Takeaways:

  • The LSTM model performed better in predicting the direction of stock price movements. The model achieved high accuracy in predicting up or down trends on the test set. This indicates LSTM networks can effectively capture nonlinear relationships and long-term dependencies in stock price time series data.
  • The ARIMA model was suitable for extracting trends and periodicities from stable time series data. The study used a novel strategy of incorporating momentum and reversal factors. This combined method comprehensively considered the complexity of the stock market while balancing accuracy and simplicity.
  • According to comparative analysis, the LSTM model is more suitable for predicting stock price movements but may overfit noisy data. The ARIMA model better handles stable time series but has poor performance on nonlinear relationships. Overall, the two models have their own applicable ranges and should be selected based on the specific situation.

Outline:

  • Background: High-frequency trading has developed rapidly in recent years. This research aims to explore the application of HFT techniques in quantitative hedge funds.
  • Purpose: Through theoretical research and empirical analysis, this study investigates the application effects and suitable application scope of HFT techniques. It also compares different models to make more accurate predictions.
  • Methods: This study employs the LSTM model and ARIMA model for prediction and comparative analysis based on real market data. Quantitative trading strategies are designed and backtested.
  • Conclusions: According to comparative analysis, the LSTM model performs better in predicting stock price movements while the ARIMA model is more suitable for stable time series data. The two models have their own advantages and limitations and should be selected based on the practical situation.

Relation to the Previous Project:

The preceding project “Investment Cognitive Reasoning and Intelligent Model” delved into third-generation AI, aiming to deconstruct intricate behaviors into fundamental operations for potential recombination. This concept closely aligns with the quantitative trading strategies scrutinized in hedge fund research. Furthermore, it offers valuable insights into the evolving landscape of machine learning techniques within the realm of finance and investment.

While previous works have affirmed the efficacy of machine learning methods, there exists a notable gap in the exploration of a comparative analysis between traditional time-series models and deep learning models, as well as a scarcity of research delineating their respective practical applications. This paper is dedicated to addressing this void, with the primary objective of investigating the nuances and potential use cases within this domain.

Reflection:

Despite the absence of papers or patents, this phase allowed me the invaluable freedom to explore and engage in fruitful discussions with mentors and collaborators. I remain appreciative of this experience, with the only regret being that I believe I could have maximized it further—perhaps by condensing corporate experience and conducting independent experiments. Perhaps, it’s a necessary stage for development.