| October 1, 2021
Project overview
The project explores financial time series modeling for automated trading systems, focusing on analyzing the non-stationary nature of financial data. Since market movements depend on partially observable environments, identifying patterns and relationships between financial assets requires advanced modeling techniques. Despite extensive research, no consensus has been reached on the best approach, and the field remains highly active.
Traditional methods for financial series modeling rely on Gradient Boosted Models (GBM) such as XGBoost and LightGBM, as well as Hawkes processes, which are widely used in the industry for event-driven modeling. However, the recent dominance of deep learning models in various fields has led to the emergence of hybrid approaches that combine neural networks with traditional statistical methods.
The goal of this project is to compare hybrid neural networks with traditional approaches such as GBM, stochastic models (Hawkes processes), prediction intervals, and deep learning architectures like 1D-CNNs and LSTMs. The evaluation will consider different data formats (tabular vs. sequential) and use diverse metrics, including cross-entropy for trend analysis, RMSE for regression accuracy, and domain-specific indicators like the Sortino and Omega ratios.
The study involves implementing state-of-the-art financial modeling methods [1][2][3][4], developing a testbench to compare their performance, and evaluating models such as Neural Hawkes Processes (NHP), Transformer Hawkes Processes (THP), CT-GRU, and Intensity Processes (IP). The final deliverable will be a scientific article for arXiv, summarizing the findings and providing insights into the strengths and limitations of different modeling approaches.
References:
[1] Visualizing and Understanding Atari Agents, Greydanus, 2018.
[2] Uncovering Surprising Behaviors In Reinforcement Learning Via Worst-case Analysis, Ruderman, 2019.
[3] TA-Lib, python wrapper to the ta-lib library for technical analysis.