PhD Oral Exam - Soroush Shahsafi, Information and Systems Engineering

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

Financial markets are noisy, non-stationary, and driven by heterogeneous information sources, which makes it difficult to build trading systems that are both profitable and robust. This thesis develops a set of deep learning and reinforcement learning frameworks that improve the full pipeline from data preprocessing and signal generation to portfolio construction and position sizing.

First, the Deep Q-Network (DQN) trading under noisy price observations is studied, and it is shown that advanced denoising techniques such as attention-based Temporal Attention Network (TAN) preprocessing consistently improves return-risk performance and reduces drawdowns relative to common technical-indicator benchmarks. Second, for multi-asset trading, a hybrid architecture is proposed that decomposes discrete asset selection (via DQN buy/sell decisions) from continuous capital allocation (via portfolio optimization). Across emerging-market Exchange-Traded Funds (ETFs) and U.S. equities, coupling DQN with portfolio optimizers improves the result, with Sharpe-ratio maximization providing the most consistent gains.

Also, a multimodal forecasting approach is developed such that it converts open, high, low, and close (OHLC) time series into Gramian Angular Field images and extracts visual features using a pre-trained ResNet, and feeds them as covariates to a Temporal Fusion Transformer (TFT), yielding substantial accuracy improvements over Long Short-Term Memory (LSTM) baselines.

In addition, candlestick trading as an information-fusion problem is developed by combining pattern voting, interpretable confirmation rules, and Convolutional Neural Network (CNN)-based chart learning. Finally, this is extended into a two-stage system in which candlestick-derived signals are sized using short-horizon forecasts, optionally augmented with aligned news sentiment, with the strongest results being delivered by LSTM-based sizing.