◆ yatawara.com / research
A single research program along four pillars that share one core: modeling complex temporal and spatial patterns — how systems encode, persist, and forget information — to produce actionable insight.
How do markets remember? My primary line of work develops new econometric models for volatility persistence — measuring the shape of market memory directly from data instead of assuming it. The central finding: that memory decays sub-exponentially — fast at first, then a long, faint tail.
$ kernel --plot stretched-exponential --alpha 0.27 --live
Volatility memory kernels across 100+ assets — equities, FX, commodities, fixed income, crypto, and the VIX — follow sub-exponential (stretched-exponential) decay, fitting markedly better than GARCH and FIGARCH. Introduces the GMARCH and SEARCH modeling frameworks.
Augments Conrad & Engle's (2025) Multiplicative Factor Multi-Frequency GARCH with smoothed trading volume. Evaluated on 15 US assets with rolling out-of-sample forecasts; ships with a MATLAB toolbox.
A sign-sensitive long-run component, giving the long-run variance an asymmetric response to positive versus negative shocks — bad news and good news leave different marks on market memory.
An exponential-GARCH version of the multiplicative-factor multi-frequency framework, including asymmetric MEM and logarithmic MEM long-term variants.
Multiple Regime Hyperbolic GARCH (MR-HYGARCH), Fractionally Integrated GJR-GARCH (FI-GJRGARCH), and Smooth-Transition FI-GJRGARCH — extending long-memory and asymmetric volatility modeling.
MF2V-GARCH Toolbox for MATLAB — open-source implementation of the multiplicative-factor multi-frequency framework with smoothed trading volume.
View on GitHub ↗Explainable machine-learning forecasting, grounded in statistics — pairing the structure of classical time-series models with the flexibility of deep learning, without giving up interpretability.
A hybrid LSTM architecture with Yeo-Johnson scaling for forecasting influenza-like illness — combining recurrent memory and convolutional feature extraction for more accurate, explainable public-health forecasts.
Statistical frameworks for measuring who breathes what — quantifying multi-pollutant exposure, its fine-scale spatial structure, and how it tracks with income across communities.
A county-level, multi-pollutant framework (PM2.5, NO₂, O₃, SO₂, CO) quantifying income-based air-quality disparities across California.
Evaluating fine-scale air-quality heterogeneity using a network of 45 low-cost multi-pollutant sensors across Minneapolis–St. Paul.
An ongoing study of the relationship between air pollution and fertility patterns across California counties.
Building and studying AI-powered tools that help students learn statistics — turning large language models from answer generators into thinking partners. Supported by the California Learning Lab ELEVATE grant.
Lead author. "From Answer Generators to Thinking Partners" — a Custom GPT framework for tutoring, prompting, and ethics in statistics education.
Co-author. A work-in-progress study of structured AI tutoring in engineering education.
A cognitive-apprenticeship approach to integrating large language models into instruction for K–12 teachers.
| Year | Project | Role | Source |
|---|---|---|---|
| 2025 | Air Pollution Disparities in California 2000–2025 | PI | CES Mini-Grant, CSUB |
| 2025–26 | ELEVATE: Enhancing Learning Experiences Via AI Techniques | Co-PI | California Learning Lab AI FAST Challenge |