Deploying agentic systems into real decision environments — capital markets and classrooms — where correctness, cost, and accountability are measured against outcomes rather than benchmarks. The recurring question is whether a system can improve on the cases it gets wrong without quietly breaking the cases it already got right.
Finance AI
Where this research started, and the longest-running thread in my work: how to act on a number you already know is uncertain. It runs from econometric risk measurement, through derivative pricing and factor models, to the routing and control problems I work on now. Four threads below, oldest to newest — each project names the setting it was done in, and the full chronology sits on the Experience page.
Market risk and systemic spillover
The starting point: measuring risk and tracing how it propagates between markets. Undergraduate research in econometric time series, and the origin of the regime-identification problem that the volatility work later turns into a control problem.
Digital Finance and Risk in Traditional Financial Markets
CompletedUndergraduate research · Central University of Finance and Economics · 2024
Computed VaR-based risk indices for five Chinese financial markets over 2013–2023, then estimated a joint vector autoregression linking those indices to a digital-finance development index. Variance decomposition and impulse response functions gave the counter-intuitive result the paper is built around: the growth of digital finance is associated with *lower*, not higher, risk in the traditional markets it competes with.
Long-Horizon Forecasting of the Dow Jones Index with ARMA Models
CompletedUndergraduate research · Central University of Finance and Economics · 2024
Fitted ARMA models to 1,685 weekly observations of the Dow Jones U.S. index from 1992 to 2024 — stationarity testing, differencing, order selection by information criteria, residual diagnostics — and projected the series forward ten years. A deliberately plain baseline, and a useful reference for how little a well-specified linear model concedes to heavier machinery on index-level data.
Derivatives pricing and surface calibration
Learning where neural networks genuinely beat classical numerical methods in pricing, and where they do not. This is also where the world-model intuition behind my current research first became concrete: a calibrated surface is a model of consequences you can query instead of re-simulating.
Neural-Network De-Americanization for American Option Pricing
CompletedQuantitative finance research · 2024
Cleaned SPX index option data from 2011 to 2023, fitted a power-law surface SSVI model to the raw implied volatility surface, and implemented binomial-tree de-Americanization as the benchmark. Synthetic prices generated by finite-difference pricing then let the neural de-Americanization of the early exercise premium be stress-tested in simulated extreme high- and low-rate regimes, where it held its accuracy at computational cost comparable to the binomial method.
SPX–VIX Joint Calibration with a Neural SDE
CompletedQuantitative finance research · 2025
Built a two-factor stochastic local volatility model in JAX whose drift, diffusion, and correlation are supplied by neural networks and trained through Euler–Maruyama simulation, so that SPX and VIX smiles are calibrated jointly and remain arbitrage-free across both markets. Quasi-Monte Carlo via Sobol sequences cut the deviation ratio from 5.02/0.764 (SPX/VIX) to 3.8/0.28; the calibration was extended from 8 maturities to all 38 SPX and 13 VIX maturities under a warm-start then next-day protocol, training roughly 7× faster than the reference implementation (0.9 vs 7.86 minutes per 100 epochs). Short-dated, deep out-of-the-money SPX calibration stayed unstable — small Vega makes the implied volatility gradients overshoot — which is the model's practical limit.
Neural-Network Optimization for Pricing and Local-Volatility Surfaces
In progressQuantitative finance research · 2025 – present
Continuing the pricing line: comparing SSVI parameterisations, local volatility extraction, American binomial trees, and finite-difference solvers against neural surrogates, with the goal of pinning down which part of the pricing pipeline the network should actually replace.
Factor models and LLMs for alpha
Cross-sectional prediction, and the point at which language models entered the pipeline. Replication came first — establishing what a careful baseline actually delivers — before adding architecture or modality on top of it.
Man versus Machine Learning Revisited — Replication and Extension
CompletedResearch assistant · Central University of Finance and Economics · 2025
Rebuilt the replication package for Zhang, Zhu and Linnainmaa (RFS 2025) end to end on WRDS data: look-ahead-bias-free earnings forecasts, a random-forest replication benchmarked against the linear forecasts of So (2013) and Hughes et al. (2008), then single sorts and Fama–MacBeth regressions over FF49 industries, with alternative random-forest and machine-learning specifications probed for robustness. The exercise is the reason I treat “the baseline, implemented properly” as the hardest number to beat.
Deep-Learning Factor Mining for A-Shares
CompletedHantak Investment Management · Jan – Apr 2025
A factor-mining pipeline built around the liquidity–volatility anomaly, adding ProbSparse attention and TFT variable selection to lift monthly IC from 0.05 to 0.13. Size and industry neutralisation plus AdamW with cyclic learning rates kept the gain from being an overfitting artefact; backtests from 2010 to 2023 showed a 15% Sharpe improvement and persistent out-of-sample alpha against the CSI 300.
Financial LLMs and Multimodal Alpha Signals
CompletedREI-Tech, AI Lab · May – Jul 2025
The first project where a language model carried the signal rather than the analysis. A financial RAG prototype over 10k+ instructions with chain-of-thought prompting for extracting structure from earnings reports, alongside a pipeline turning YOLO-detected K-line patterns and Whisper audio tags into embeddings; a LoRA-tuned backbone delivered a sentiment factor at 0.038 peak RankIC, ahead of Alpha101. The failure modes here — ungrounded retrieval, confident wrong extraction — are what BeliefRAG and the reliability work later attack directly.
Volatility forecasting and control
Where the thread arrives: once you accept that no single volatility estimator is reliably best, the decision stops being a forecasting problem and becomes a routing problem — which is the same question my agent research asks about models and memories.
Beyond Forecasting: Recasting Volatility Control as a Routing Problem
Under reviewACM ICAIF 2026 · arXiv:2608.10375
Instead of chasing a single best volatility forecast, the controller routes between estimators and scaling rules according to the prevailing regime, so exposure decisions degrade gracefully when any one estimator is wrong.
Volatility-Targeted Strategies for S&P 500 Futures
CompletedJPMorgan Chase & Co., Quantitative Research · Jan – Apr 2026
An end-to-end volatility-targeting pipeline — preprocessing, forecasting, exposure allocation, backtesting — comparing EWMA, GJR-GARCH, HAR-RV, and Random Forest estimators against regime-aware controllers. HAR-RV with regime scaling reached a 0.602 Sharpe versus 0.575 for naive scaling, while cutting maximum drawdown from 18.71% to 17.58%. The empirical basis for the routing paper above.
Cross-Asset Volatility: From the VIX Surface to Bitcoin
CompletedPositive Research Ltd. · Dec 2025 – Jan 2026
Validated volatility spillover from the S&P 500 into crypto, then built a hybrid model joining VIX surface features to BTC level-one data, which held up best precisely where single-asset models fail — around macro events. Arbitrage-free implied volatility surfaces calibrated by SSVI with differential evolution (WRMSE < 0.006) supplied ATM variance and skew as leading indicators; 30+ alpha signals from 60,000+ hours of tick data fed an Optuna-tuned LightGBM to a 0.46 out-of-sample IC on 24-hour volatility.
BTC Realized Volatility: GARCH-Family and LSTM Baselines
CompletedIndependent research · 2025
The groundwork for the cross-asset model above. Forecast seven-day-ahead average realized volatility for BTC-USD with the GARCH family — GARCH, TARCH and variants over interval windows — then an LSTM on the same features, scored against naive and rolling baselines. Framed throughout as a trade: forecast realized volatility against market implied volatility, and the sign of the gap tells you whether the option is cheap or rich.
Education AI
Assessment and feedback systems that stay calibrated, auditable, and stable as their grading standards are revised.
RubricWorld: World Model-Guided Rubric Optimization for LLM Graders
Under reviewARR, October 2026 cycle · Michigan State University, DSE Lab
Every edit to a grading rubric can correct some mistakes while degrading decisions that were already right — and the same edit behaves differently depending on which revisions came before it. Existing prompt and rubric optimizers pick edits on aggregate validation performance and never learn those state-dependent effects. RubricWorld learns them from paired grading decisions taken before and after each edit, predicting correction and degradation rates at individual ordinal score boundaries to guide candidate selection; an independent certification step then uses held-out measurements and finite-sample confidence bounds to decide whether an edit is committed at all. Across EIR (National), ASAP 2.0, and ASAP-SAS with three frozen LLM graders it achieves the best result in 20 of 27 reported metric comparisons, raising gpt-5-mini accuracy on ASAP 2.0 from 18.20% to 46.70%. Ablations isolate why both halves matter: removing rubric-state information drops average accuracy from 40.18% to 35.42%, and removing the learned edit-effect model drops it to 27.31%.