Quantitative Researcher - ML-Driven Market Modeling Expert

Yuan Zhao

*****@****.*** 862-***-**** Harrison, NJ (Open to Relocate) Linkedin Github

OBJECTIVE

Ph.D. candidate specializing in signal processing, statistical inference, convex optimization, and time-series modeling, with hands-on experience applying machine learning and large language models to solve complex real-world problems. Proven track record in likelihood-based modeling, information-theoretic evaluation, and ML/LLM-driven system design across financial microstructure and NLP applications. Seeking Quantitative Researcher roles focused on systematic strategy development and quantitative system modeling. EDUCATION

New Jersey Institute of Technology Newark, NJ

Ph.D., Electrical Engineering, GPA: 3.9/4.0 Expected 06/26 M.S., Electrical Engineering, GPA: 4.0/4.0

Graduate Courses: Mathematical Statistics; Bayesian Statistics; Regression Analysis; Statistical Learning / Statistical ML; LLM Engineering and Applications; Statistical Inference; Optimization; Random Signal Analysis; Linear Algebra Chongqing University of Posts and Telecommunications Chongqing, China B.S., Electrical Engineering 06/19

RESEARCH EXPERIENCE

Quantitative Data Processing on Market Microstructure (Market Analysis and Time-Series Modeling) Y. Zhao, A. Abdi, and A. Abdi, “An information transmission approach to study informed stock trading models and data,” submitted, 2026.

• Large-Scale Data Analysis & Modeling: Applied maximum-likelihood estimation to latent-information microstructure models (Easley–Kiefer–O’Hara–Paperman, EKOP; Venter–de Jongh, VdJ) on daily aggregated buy/sell order counts. Built scalable estimation pipelines over 11 years of high-frequency trade records aggregated to daily counts across 500+ NASDAQ stocks, extracting the PIN (Probability of Informed Trading) index as an interpretable factor of information asymmetry.

• Latent-State Model Evaluation: Developed a label-free evaluation framework for latent-state models. For each day, computed class-conditional likelihoods under the three event regimes and assigned the event label via Maximum a Posteriori (MAP) decoding. Validated reliability using mutual information (prior to posterior information gain) and correlation diagnostics linking inferred states to raw order-flow patterns.

• Backtesting & Business Impact: Conducted historical backtests to quantify the link between inferred event states and subsequent returns (same-day/next-day), reporting signal hit-rate and stability across windows. Packaged outputs into actionable signals and monitoring metrics—supporting event-risk detection, portfolio risk controls, and alpha/risk feature engineering (e.g., market-quality indicators around corporate disclosures). Rational Inattention and Noisy Data (Optimization Theory and Decision Science) Y. Zhao, A. Abdi, M. Dean, and A. Abdi, “Optimal decision making with rational inattention using noisy data,” in Proc. Conf. Inform. Sci. Syst., Johns Hopkins University, Baltimore, MD, 2023, pp. 1-5. Y. Zhao, A. Abdi, M. Dean, and A. Abdi, "A framework for noise mitigation with rational inattention to noisy data," submitted, 2026.

• Theoretical Modeling via Convex Optimization: Proposed the Noisy Rational Inattention (Noisy-RI) model to capture the reward–cost tradeoff in decision making. Extended classical RI theory by introducing a noise parameter p and solving the model via constrained convex optimization (Lagrangian duality/KKT conditions), deriving closed-form results for decision accuracy under limited attention resources.

• Large-Scale Validation & Goodness-of-Fit: Validated the theoretical framework on both synthetic datasets

(20,000+ samples) and experimental datasets (200+ subjects). Demonstrated consistent theory-to-data agreement and achieved a ~92% reduction in Sum of Squared Errors (SSE) compared to basic Rational Inattention baselines in high-noise regimes.

• Closed-Form Robustness to Noise Shifts: Derived mathematically tractable closed-form asymptotic results for the mitigation rule, allowing rigorous analysis of how decision accuracy changes under noise-level shifts (including p drops) and identifying stable operating regions.

Analyzing Large Language Model for Interpretability (Black-box Behavior Analysis) Y. Zhao, A. Abdi, “Explainability of Large Language Models Using the Rational Inattention Theory: A Case Study in Hate-Speech Detection,” accepted for presentation at the 190th Meeting of the Acoustical Society of America, 2026. Y. Zhao, A. Abdi, “Interpretability of LLM Classifiers via the Rational Inattention Theory with Application to Hate Speech Detection,” submitted, 2026.

• Noisy-RI Numerical Interpretability Framework: Proposed a novel “Noisy Rational Inattention” (Noisy-RI) interpretability framework for LLM classification that relies solely on input–output behavior without requiring computationally expensive Chain-of-Thought (CoT) generation. This enables a low-cost, evaluation-driven self-diagnosis pipeline that significantly outperforms traditional token-based explanation methods in efficiency.

• Agent-Based Modeling & Parameter Estimation: Modeled the LLM as an information-constrained decision-making agent. Fit the extended RI model to estimate interpretable parameters, successfully decoupling performance drivers into internal factors (internal decision strategy) and external factors (noise sensitivity and distortion level).

• Multi-Environment Adversarial Benchmarking: Constructed a rigorous evaluation dataset by injecting realistic text perturbations, including in-word character distortion to mimic homophones, and adversarial misspellings. Benchmarked across 11 distinct environments ( 1,000 samples each) via consistent model fitting. Results indicated that OpenAI models (GPT-3.5 and GPT-5.2) consistently exhibited more efficient internal decision strategies than Google models (Gemini-2.5 and Gemini-2.0) under text distortion.

• Issue Localization Framework: Established a structured parameter-based debugging protocol that serves as a systematic framework for localizing root causes of LLM accuracy drops, enabling targeted fixes across model-side and data-pipeline components.

PROFESSIONAL EXPERIENCE

Finz New York, NY

AI Researcher Intern 05/25 – 07/25

• Product Backend Platform: Built an “AI CFO” backend platform enabling small businesses (restaurants, convenience stores) to securely connect financial and supply-chain data and ask questions on cash flow, balances, and expenses—reducing spreadsheet work and improving visibility.

• Modular LLM Orchestration: Refactored the chatbot into a modular, multi-stage LangChain pipeline with prompt engineering and multi-model routing (GPT-4, Gemini 2.5): query rewriting intent routing data fetching/RAG response aggregation sanitization, improving maintainability and end-to-end financial Q&A reliability. Containerized and deployed services via AWS/Docker.

• Performance Optimization & Observability: Introduced Redis caching and end-to-end profiling/monitoring

(latency + memory dashboards) to reduce redundant fetches and optimize system performance - improving p95 latency from ~7s to ~1s and cutting LLM costs by ~30%. Skills: Python, FastAPI, MongoDB, LangChain, Redis, Docker, AWS New Jersey Institute of Technology Newark, NJ

Lab Assistant, LLM-Based Financial Q&A System for Company Annual Report 01/25 – 05/25

• Corpus-Scale Platform: Architected an LLM-based financial Q&A system over 10,000+ A-share annual reports

(~70GB), enabling high-precision fact-checking and complex financial/statistical analysis via multi-model orchestration.

• Data Extraction & Normalization: Built a robust PDF parsing pipeline using xpdf + Camelot to extract high-fidelity text and tables from complex layouts; automated filtering/normalization to convert heterogeneous financial statements into analysis-ready datasets stored in SQLite.

• Hybrid QA Engine: Fine-tuned Qwen models for intent classification, keyword extraction, and NL2SQL to enable deterministic SQL answers with calculation and statistical aggregation; paired with a retrieval/analysis path for ambiguous questions. reaching 87% QA accuracy.

• High-Throughput Serving and Latency Optimization: Implemented vLLM to accelerate the inference; reduced GPU resource consumption by 50% and achieved 2.7s average response time. Skills: Python, SFT (Supervised Fine-Tuning), vLLM, RAG, Qwen, LangChain, SQLite, xpdf, Camelot Lab Instructor and Teaching Assistant Microprocessor Laboratory (ECE 395), Circuit Analysis DC and AC (ECET 201)

01/25 – 05/25

• Instructed and mentored 100+ students across four years in microprocessor systems and analog/digital circuit analysis; designed practical lab challenges and theoretical overviews, and guided student teams through industry-level technical projects, fostering hands-on problem-solving and engineering rigor. PROFESSIONAL MEMBERSHIPS AND ACTIVITIES

• Student Member, Institute of Electrical and Electronics Engineers (IEEE)

• Student Member Acoustical Society of America (ASA)

• President, Chinese Students and Scholars Association, New Jersey Institute of Technology 09/23 – 09/24

• Host and Organizing Volunteer, 40th IEEE Sarnoff Symposium, New Jersey Institute of Technology 09/19 AWARDS

• Selected as the Ph.D. Student of the Month, Electrical and Computer Engineering Dept. 8/23

• Ross Memorial Fellowship, New Jersey Institute of Technology 9/24 – 05/25

• Hashimoto Fellowship, New Jersey Institute of Technology 09/25 - Present

• ASA Workshop 2026 accommodation financial support (competitive selection)06/26

Contact this candidate