04/01/2021
By Jia Wang

The Department of Computer Science announces the Doctoral Dissertation Defense of Jia Want on "Financial Markets Prediction with Deep Learning."

Ph.D. Candidate: Jia Wang
Date: Friday, April 16
Time: 8:30 – 10:30 a.m.
Location: Via Zoom: Meeting ID: 831 550 6238, Passcode: 756394

Thesis Committee:

  • Benyuan Liu (advisor), Professor, Dept. of Computer Science, UMass Lowell
  • Yu Cao, Professor, Dept. of Computer Science, UMass Lowell
  • Hongwei (Harry) Zhu, Professor, Chair, Dept. of Operations and Information Systems Acting, UMass Lowell 

Abstract:
Financial markets are a complex dynamical system. The complexity comes from the interaction between a market and its participants. The integrated outcome of activities of the entire participants determines the market trends, in turn, the market trends affect activities of participants. These interwoven interactions make financial markets keep evolving and thus extremely difficult to be predicted. In an effort to explain the "random walk" behavior of financial markets, the Efficient Markets Hypothesis (EMH), the idea that market prices reflect all information, has become one of the most influential and enduring idea in financial economics.

The main implication of EMH is that, markets are rational and market prices cannot deviate from the fair market value so that it is impossible to outperform the overall market through expert security selection or market timing. Proponents of EMH believe that investors benefit only from investing in a low-cost, passive portfolio, and "Warren Buffett is as abnormal as 3-sigma event''. However, the emerging discipline of behavioral economics and finance argues that markets are not rational in that as human beings, market participants cannot be consistently rational and always make optimal decisions. Some participants make choices based on their "optimal'' guess that they believe, but in fact mixed with subjective factors, such as loss aversion, overconfidence, herding, and overreaction; Meanwhile, some other participants who are truly rational at that moment search for these biased decisions to obtain high returns with low risks. We have to accept a brutal but true fact that someone's mistakes would be others' opportunities in financial markets.

What if a decision maker is purely rational? Inspired by the Adaptive Markets Hypothesis (AMH), which reconciles market efficiency with behavioral alternatives by applying the principles of evolution to financial interactions, this thesis explores predicting financial markets with deep learning approaches, which are designed to automatically capture generalized features from large-scale datasets and have been successfully applied in various computer vision and natural language processing tasks. We remodel multiple deep learning approaches to adapt them for the characteristics of financial markets, and in the meanwhile, organically combine them to address the aforementioned challenges of financial markets predictions. In particular, (1) We design customized 1-D convolution neural networks named Cross-Data-Type 1-D CNNs (CDT 1-D CNNs) to extract local features instead of technical indicators, which can prevent our approach from being affected by human biases; (2) We introduce sequence-to-sequence framework (Seq2Seq) with attention mechanisms to extract temporal features in order to promptly adopt to the environment changes of financial markets; (3) Aiming for organically combining local and temporal features, we replace the fully connected layers inside the sequence-to-sequence framework with CDT 1-D CNNs; (4) We introduce inter-attention mechanism to capture the context information for temporal feature extraction, meanwhile, we also introduce intra-attention mechanism to capture the daily regularity and then highlight informative spots within the span of a given trading day; (5) We also introduce Kullback-Leibler divergence as an extra regularizer to alleviate overfitting problems. To our best knowledge, the integrated system of all the above approaches, called Convolutional LSTM based Variational Sequence-to-Sequencemodel with Attention (CLVSA), is the first deep learning approach without technical indicators to predict financial markets.

There have been many attempts to predict the market movement using machine learning algorithms. However, most of these attempts are evaluated by machine learning criteria such as accuracy and F1 score. To approach the real-time trading scenario, we design backtesting strategies to evaluate the practical performance of the models in this thesis. We can assess the viability of our approach by verifying whether or not backtesting experiments can yield positive and robust results. To avoid over-optimistic results, we apply high transaction cost in backtesting experiments. We also backtest all the rivals of our approach for comparison. Our experiments show that our approaches, especially CLVSA, outperform the baseline methods, such as support vector machine (SVM), Feed-forward networks, regular CNNs, and vanilla Seq2Seq on both financial and machine learning criteria.

Although it is the common sense in finance and economics that prices reflect all information, investigating the sentiment data is still informative for traders. we thus introduce TRMI data to investigate whether or not the sentiment data provides signals that are more directional than price movements. A series of experiments indicate that sentiment data does not only provide informative features to prediction systems, but it also contains the extra information which prices and volume do not reflect. With a subtle method to fuse TRMI data and historical trading data, we upgrade CLVSA to dual-CLVSA, which outperforms CLVSA 9.3% by average annual return and 0.91 by Sharpe ratio on SPDR S&P 500 ETF Trust. We also reveal the details about how TRMI data working in dual-CLVSA with the analysis of two cases.