Kexin Deng

Hi! I am a graduate of the Master of Engineering in Financial Engineering program at Cornell University. My work focuses on systematic equity research, machine learning for financial prediction, and quantitative portfolio construction. At Cornell, I conducted research on market microstructure, execution modeling, and regime-aware portfolio strategies. I have also built large-scale machine learning pipelines for cross-sectional equity prediction, including feature engineering, signal evaluation, and portfolio optimization. My broader interests lie in applying machine learning and data systems to financial markets, including alternative data, large-scale factor modeling, and AI-driven research workflows like RAG.

Email  /  CV  /  SSRN  /  Instagram  /  LinkedIn  /  Github

Research

My research interests lie at the intersection of systematic investing, machine learning, and quantitative finance. I work on building data-driven models for financial markets, including cross-sectional equity prediction, portfolio construction, and regime-aware risk modeling. My recent work includes large-scale machine learning pipelines for equity prediction, retrieval-augmented generation systems for financial knowledge retrieval, and systematic trading strategies informed by market microstructure signals.

This project studies retrieval-augmented generation systems for financial knowledge retrieval. We evaluate multiple RAG architectures including vector search, hierarchical retrieval, and graph-based retrieval pipelines over enterprise-scale financial document corpora.

Developed a large-scale cross-sectional equity prediction system combining domain-driven feature engineering with machine learning and deep learning models. The pipeline expands 400 base alpha signals into 1,947 engineered features, integrates LightGBM, IC-oriented neural networks, and Ridge regression into a Model Zoo ensemble optimized using a Correlation-Optimized Ensemble (COE). The final system achieved a public leaderboard score of 0.07066, ranking Top 2 publicly and Top 3 privately in the Trexquant competition.

This survey studies the emergence of agentic artificial intelligence in financial markets. We analyze architectures for autonomous financial agents capable of reasoning, planning, and adaptive decision-making across trading, portfolio management, and market infrastructure. The paper reviews system design, multi-agent coordination, regulatory considerations, and the implications of agentic AI for market efficiency, liquidity provision, and systemic risk.

This study discusses how recent innovations in cryptography can help bridge the divide between the optimality of now-traditional dark pools and the advances in the latest digital ledger technologies. We argue that recent innovations such as Merkle tree compression enable efficient decentralized electronic matching in a traditional dark pool environment. This development can help bring about higher market efficiencies in markets around the world.

Developed a machine learning framework to forecast extreme spikes in the CBOE Volatility Index (VIX) using cross-asset market signals and macroeconomic indicators. The system integrates equity market features, options-implied volatility metrics, macro factors, and market microstructure signals to detect early-warning patterns preceding volatility explosions. The feature pipeline incorporates realized volatility, term structure dynamics of VIX futures, equity index momentum, credit spreads, liquidity indicators, and macroeconomic variables from the FRED database. Multiple classification and regression models including Gradient Boosting, Random Forest, and neural networks are trained to predict the probability of large VIX spikes and regime transitions. The model is evaluated on historical market stress events such as the COVID crash, inflation shocks, and volatility regime shifts, demonstrating improved early detection of volatility surges and enabling systematic portfolio protection strategies for risk-managed equity portfolios.

Built an end-to-end platform for extracting structured investment signals from financial news. The system ingests company-specific news articles, applies FinBERT to generate article-level sentiment scores, and aggregates them into daily stock-level sentiment time series for S&P 500 constituents. The pipeline combines sentiment signals with price dynamics to visualize sentiment–return co-movements and generates simple Buy/Hold/Sell regimes based on smoothed sentiment indicators. An LLM-based agent summarizes daily news flows into concise investor-friendly insights, enabling faster interpretation of market-moving events. The platform integrates data ingestion, NLP-based sentiment modeling, signal aggregation, and interactive dashboards, demonstrating how large-scale financial text data can be transformed into actionable signals for quantitative asset management workflows.

Developed a machine-learning–driven portfolio construction framework combining clustering-based stock selection with shrinkage covariance estimation and CVaR optimization. Backtests on S&P 500 equities (2020–2023) show improved stability and downside risk control, achieving Sharpe ratios above 1.6 and significantly outperforming the S&P 500 benchmark.

Score-based high-frequency execution model using L2 limit order book signals including momentum, order flow imbalance, volatility-to-spread ratio, bid–ask spread, and time pressure. Rolling optimization and queue-aware execution improve trade timing and reduce buy–sell spread by 40–125% relative to a TWAP benchmark.

Quantitative sector analysis of U.S. auto manufacturers (Tesla, GM, Ford) implemented in R (quantmod). Constructed a tangency portfolio using Mean–Variance Optimization with Treasury-bill risk-free rates and benchmarked against the industry index. Extended analysis using CAPM regressions and tail-risk metrics (VaR, Expected Shortfall) to evaluate systematic exposure and downside risk.

This project addresses a relational prediction task where the objective is not to classify individual inputs, but to model the interaction between two inputs. Using a Siamese architecture with a shared ResNet50 backbone adapted for grayscale images, each input is encoded into a comparable feature representation, which is then combined through concatenation and passed into a fully connected classifier to capture pairwise dynamics. Beyond architecture design, performance gains were driven primarily by training strategy, including mixed precision for stability in deep networks, balanced binary cross-entropy to address class imbalance, cosine warm restarts for optimization, and F1-based threshold tuning to better align training with evaluation metrics. The final model achieved 0.9048 accuracy on the private leaderboard (Top 4 out of 91), significantly outperforming CNN and ResNet18 baselines. This project highlights that for relational tasks, explicitly modeling interaction structure is often more critical than increasing model complexity, and demonstrates how training methodology can be as important as model architecture in achieving state-of-the-art performance.

Miscellanea