Advith's Calendar

Room 1 (4269, Beckman)

Student Name

Mentor

Title

Abstract

10:10-10:30

Benjamin Leidig

Hyoeun Lee (seesion chair)

Evaluating Hybrid Deep Learning Architectures for Day-Ahead Electricity Price Forecasting in the German Spot Market

Forecasting day-ahead electricity prices is a longstanding focus of econometric research and energy market participants. Recent studies indicate that contemporary approaches, including hybrid and deep learning techniques, are increasingly used. However, the complexity and resource requirements of advanced techniques make custom hybrid learning architectures challenging. For example, inconsistent adherence to temporal constraints is a recurring limitation. Additionally, inappropriate model interpretations and comparisons can be misleading, often due to the black-box nature of neural networks. To address these limitations, a comprehensive selection of feature engineering and machine learning techniques was used to forecast hourly day-ahead electricity prices on the German spot market. This selection included architectures incorporating Variational Mode Decomposition, convolutional and recurrent neural networks, multi-head attention, and transformers, all under strict temporal training and forecasting constraints. The supervised models and signal decomposition techniques were tuned using Bayesian optimization and frequency separation analyses, respectively. To compare real-time performance, these frameworks were tested on a year of hourly data. The results revealed that novel hybrid deep learning architectures demonstrate superior performance during regime changes and volatility spikes inherent to energy markets compared with traditional statistical methods.

10:30-10:50

Jackson Fleege

Kit Clement

Predicting Use of Narrative in Probability Modeling from Students’ Narrative Identity

11:00-11:20

Shiding Huang (M)

Joshua Agterberg

Shared Subspace Estimation in the AJIVE Model via Spectral Initialization and Refinement

In data integration, one is provided with multiple data matrices posited to share some common structure.  The Angle-Based Joint and Individual Variation Explained (AJIVE) model is a statistical model for data integration that assumes each matrix shares a common signal subspace while containing unique individual signals.  This project explores a novel two-stage approach to recover the shared signal subspace. First, we estimate an initial shared signal subspace via existing spectral methods from the literature. Second, we apply a refinement step to improve the recovery based upon the initial estimate. We compare the results through simulations. The results indicate that the refinement step improves over different initializations. Furthermore, the results also clarify how initialization quality, signal strength, and structure of unique components affect the recovery of shared signal subspace.

11:20-11:40

Molin Yang, Mengxuan Wei, Rongsheng Zhang

Hyoeun Lee (session chair)

Cross-Exchange Limit Order Book Signals for Ultra-Short-Term Cryptocurrency Return Prediction

The rapid growth of cryptocurrency derivatives markets and the simultaneous operation of multiple trading venues create microstructure dynamics that single-exchange models fail to capture. This study investigates ultra-short-horizon return prediction in multi-exchange cryptocurrency markets using limit order book data, focusing on both numerical forecasting and directional classification of 5-second returns. For classification, we formulate the task as a multi-class prediction problem using causal and regime-aware label design. We compare sign-based, fixed-threshold, quantile-based, and volatility-scaled labeling schemes, and find that volatility-scaled labels provide the most stable and economically meaningful formulation across changing market conditions. For regression, we compare raw-return and volatility-normalized targets. Using data from four major perpetual futures exchanges—Binance, Bybit, OKX, and GateIO—aligned to a common 200-millisecond grid, we construct features combining single-exchange microstructure signals with cross-exchange signals such as relative prices, order flow differentials, and market-wide principal components. We evaluate tree-based models, Hawkes-process-based models, and neural networks under a strict chronological split. Our results show that cross-exchange microstructure information provides incremental predictive power for ultra-short-term crypto return forecasting, highlighting the value of multi-venue signals beyond what any single exchange can provide.

11:40-12:00

Idrees Kudaimi, Christopher Ye, and Amy Kodama

Daniel Eck

Expanding The Interactiveness of Era-Adjusted Baseball Statistics

Full House Modeling (Yan et al., Annals of Applied Statistics, 2025) produces era-adjusted baseball statistics by balancing a player's performance against their peers with the size of the MLB talent pool in their era. The methodology is peer-reviewed, but its reach depends on whether audiences outside academic statistics can engage with the results. This talk surveys three student-driven efforts to push era-adjusted statistics into wider circulation. The Era Curveball Substack, launched in 2025, publishes player profiles and case studies on figures including Juan Soto, Freddie Freeman, Cole Hamels, and Mark Buehrle, alongside literary pieces that explore what era-adjustment means for how baseball greatness is understood. Christopher Ye's DataDugout YouTube channel translates the same statistics into video-essay form, with case studies on Lefty Grove, Willie Stargell, Joey Votto, and the 2026 Hall of Fame ballot. A preliminary effort to convert era-adjusted statistics into era-adjusted player ratings for Out of the Park Baseball (OOTP) is also discussed. This work has not yet yielded results, but the exploration clarifies what a statistic-to-rating conversion would require. Taken together, these efforts illustrate how a peer-reviewed methodology can be extended into a public-facing ecosystem where the same underlying model supports rigorous research, fan-friendly writing, video content, and the beginnings of interactive simulation.

1:00-1:20

Junan Mao (M), Chaerin Kim, Chahak Gupta

Matthew Singh

Embedding Neural Machine-Learning Models in Anatomical Space

1:20-1:40

Rebecca Shi

Matthew Singh

Network Embedding of Cognitive Aging

1:40-2:00

Ning Xu

Susu Zhang (session chair)

Performance of Estimators for IRT Estimation Under Varying Conditions

This study investigates the performance of Marginal Maximum Likelihood Estimation (MMLE) and Constrained Joint Maximum Likelihood Estimation (CJMLE) for the two-parameter logistic (2PL) item response theory (IRT) model. MMLE assumes latent traits follow a normal distribution (N(0,1)) and integrates them out of the likelihood. However, this numerical integration becomes slow as the number of items grows. CJMLE avoids integration entirely by treating all parameters as fixed unknowns. Standard JML is known to be inconsistent due to the incidental parameter problem, but CJMLE adds constraints that restore consistency when both the number of items and sample size grow. A Monte Carlo simulation is conducted under varying numbers of items (J) and examinees (N).  Nine simulation conditions cross J = 40, 100, 500 with N = 500, 1000, 2000, each replicated 100 times. Performance is measured by bias, RMSE and response probability recovery. When J is small, MMLE performs better. As J increases, CJMLE's bias decreases significantly. At J = 500, both methods achieve the same level of accuracy in parameter and probability recovery. These results suggest that MMLE remains the preferred choice when J is small while CJMLE is a practical alternative when J is large and numerical integration becomes slow.

2:00-2:20

Bingqing Li

Gökçe Dayanıklı

Deep Learning for Epidemic Control with Vaccination via Mean Field Games

Traditional epidemic models, such as the SIR (Susceptible-Infected-Recovered)     model, typically treat individuals as passive agents and fail to capture the influence between individual behavior and population-level outcomes. To model the individual-level decision-making, mean field games (MFG) framework has been recently utilized to model the spread of infectious diseases. In these models, individuals may balance infection risk, vaccination cost, and behavioral costs into a Nash Equilibrium by choosing social-distancing and vaccination levels. Then, finding this Nash equilibrium can be reduced to solving a coupled forward-backward ordinary differential equation (FBODE) system. In our work, inspired by Liu et al (2025) which models the vaccination and social-distancing decisions together, we will extend the deep learning model that is introduced in Aurell et al (2022) to incorporate vaccination decisions. We will implement a neural network based numerical approach to solve the FBODE system that characterizes the Nash equilibrium, and discuss our convergence and experimental results.

2:20-2:40

Cheng Ai

Kelly Findley (session chair)

When AI Writes the Code: Barriers and Affordances to Productive Computing for Introductory Statistics Students

Generative AI is widely recognized as being highly effective at writing code quickly and efficiently. This reality may have implications for the common workflows and coding knowledge needed for learners of statistics and data science. Little is known, however, about whether novice learners with little to no coding background can effectively use AI to complete coding tasks in the context of data analysis. This study examines how introductory course students use generative AI while completing data visualization tasks in R. We completed task-based interviews with six students in an introduction to biostatistics course. Using thematic analysis, we document three barriers to productive task engagement that occurred while students used AI tools. We next consider two affordances that AI tools enabled in the tasks. These themes are contextualized through student examples observed in task-based interviews. Through our observations of affordances and barriers, we consider implications for instructors in how productive engagement might best be supported if introductory students use AI to code.

2:40-3:00

Room 2 (5602, Beckman)

Student Name

Mentor

Title

10:10-10:30

Han Jiang

Lelys Bravo (session chair)

Bayesian Modeling of SoyFACE Multi-Year Data on Soybean Yield Response to Elevated CO₂ and O₃

This project reanalyzes soybean data from the UIUC SoyFACE experiment (2020–2022) using a Bayesian hierarchical modeling framework. The original study applied frequentist mixed models to assess the effects of elevated ozone and altered precipitation on soybean productivity.Here, we model yield-related outcomes (e.g., seed mass and seed number) as functions of ozone, precipitation, year, and their interactions. The hierarchical structure enables partial pooling across years and treatments, improving estimation under limited sample sizes.Posterior distributions provide probabilistic interpretation of treatment effects and their interactions, demonstrating how Bayesian methods can complement traditional approaches and offer deeper insight into agricultural experiments.

10:30-10:50

Ruichen Tang (M)

Lelys Bravo

CCM-Based Causal Predictor Selection for Drought Forecasting

This project uses Convergent Cross Mapping (CCM) to select effective predictor–lagcombinations for drought forecasting. Traditional methods such as the cross-correlation function (CCF) are widely used in drought forecasting studies, but they canbe affected by autocorrelation, lagged correlation bias, and the assumption of lineardependence. To overcome these limitations, this project uses CCM as a nonlinear,trajectory-based method for lag selection. Drought indices, including SPEI and SPI, aretreated as target variables, while macroclimatic and local hydroclimatic variables areexamined as potential predictors. Through detecting the embedding dimensions, CCMtests whether the lagged predictors can reconstruct the dynamics of the droughtindices. A predictor lag is selected potentially if its cross-map skill improves withincreasing library size and satisfies significance criteria. The results suggest that CCMprovides a more flexible and robust framework for drought predictor selection,especially when climate–drought relationships are nonlinear or delayed. Compared withCCF, CCM helps to reduce the risk of selecting variables based only on shared trends orautocorrelation. Overall, this project demonstrates that CCM can improve droughtforecasting by selecting lags based on dynamical relevance rather than simplecorrelation.

11:00-11:20

Zihan Wang (M)

Jingbo Liu

When Does In-Context Learning Behave Like Bayesian Inference?

In-context learning (ICL) has recently been interpreted as performing implicit Bayesian inference over latent concepts or tasks. While prior work shows that transformer predictions can match Bayesian optimal predictors in controlled settings, it remains unclear when this interpretation truly holds and what mechanisms underlie such behavior. This work examines the Bayesian perspective on ICL through a combination of recent theoretical work and controlled synthetic experiments. We review settings in which transformer predictions align closely with Bayesian posterior predictive estimates, and investigate how this alignment depends on task structure, including task separability and sufficient statistics. The analysis highlights that agreement with Bayesian predictions does not necessarily imply correct task inference. We further identify a collapse phenomenon in larger models, where predictions degenerate to near-constant outputs and deviate significantly from Bayesian behavior. Additional experiments suggest that this effect is strongly influenced by optimization dynamics, particularly the choice of learning rate. Overall, the results indicate that while ICL can approximate Bayesian inference in certain regimes, this behavior is not robust and depends critically on both data structure and optimization.

11:20-11:40

Cainan Gu (M)

Jingbo Liu

Memorization-Generalization transition in diffusion and flow models

Prior works have revealed that diffusion models avoid memorization due to a special regularization mechanism arising during training, characterized by a separation between generation and memorization timescales. While it has been suggested that this phenomenon may extend to score-based generative models such as Flow Models, this hypothesis has not been empirically validated. In this work, we empirically investigate this question using Flow Models trained on CIFAR-10 subsets. By adapting the memorization metric based on nearest-neighbor gap ratios, we track the evolution of memorization during training. We observe that Flow Models exhibit a similar delayed onset of memorization, with a distinct generalization phase followed by a memorization phase under certain conditions. Moreover, by varying both dataset size n and model size p, we observe a similar structural pattern of memorization behavior. In particular, the gap between generalization and memorization phases scales with the dataset size n, consistent with the findings in previous works. These results suggest that the memorization–generalization transition is not unique to Diffusion Models , but may extend to score-based generative models such as Flow Models.

11:40-12:00

Rishita Mylavarapu (M)

Lelys Bravo (session chair)

Extreme Value Modeling of Heatwave Intensity and Heat-Related Mortality in India Using a Peak-Over-Threshold Framework

In recent decades, India has witnessed a steady escalation in the frequency, duration, and severity of heatwaves. These events are no longer isolated climatic anomalies but recurring extremes with measurable public health consequences. Between 1990 and 2019, India accounted for nearly one-fifth of global heatwave-related deaths, translating to tens of thousands of excess deaths annually .Existing research has established a clear association between prolonged heat exposure and excess mortality using regression-based approaches such as Poisson models, distributed lag nonlinear models, and percentile-based threshold methods. However, most of these approaches rely on predefined temperature cutoffs and deterministic definitions of heatwaves. While effective for identifying associations, they do not explicitly model the statistical behavior of extremes themselves.Exploratory analysis conducted for this project reveals strongly right-skewed and fat-tailed distributions in both temperature exceedances and excess mortality . In particular, heatwaves lasting five or more consecutive days produce mortality levels that are substantially higher than short-duration events, suggesting nonlinear escalation of risk. These empirical patterns indicate that a probabilistic extreme value framework may provide deeper insight into the true nature of heat risk in India.This study therefore proposes to apply Extreme Value Theory (EVT), specifically a Peak-Over-Threshold (POT) approach using the Generalized Pareto Distribution (GPD), to characterize the tail behavior of extreme temperature events at the city level. By shifting the focus from fixed percentile thresholds to probabilistic modeling of extremes, the study aims to provide a more rigorous understanding of rare but high-impact heat events.

1:00-1:20

1:20-1:40

Samhita Periyanayaham

V.N. Vimal Rao (session chair)

Assessing Dispositions Towards Uncertainty in Statistics

1:40-2:00

Reed Haas

V.N. Vimal Rao

Situated Measurement of Sports Fan Identification on a College Campus

2:00-2:20

Tailei Liu

Ruoqi Yu

Causal Inference in Observational Factorial Studies with Multi-Level Factors

We investigate causal inference in observational factorial studies with multi-level factors, focusing on interaction effects in the presence of incomplete treatment combinations. Using a contrast-based framework, we define interactions via tensor products of factor-level contrasts and show how to estimate factorial effects in the case of incomplete treatment combinations, under a negligibility assumption on higher-order effects. We further incorporate a bootstrap procedure to estimate the variance of effect estimators, enabling practical uncertainty quantification. Together, these contributions provide a unified framework for the identification and inference of factorial effects in multi-level settings.

2:20-2:40

Kaixuan Zhang

Ruoqi Yu

Does Early Retention Help? A Causal Analysis of Dynamic Retention Strategies in Early Elementary Education

Grade retention is widely used to support struggling students, but its effects are difficult to evaluate due to non-random assignment. Students are retained based on prior achievement, behavior, and school policies, leading to strong confounding. This project analyzes retention as a sequence of time-varying treatment decisions using longitudinal data from the ECLS-K cohort. Dynamic retention policies based on academic thresholds are evaluated with inverse probability weighting to adjust for time-varying confounders. Results estimate fifth-grade outcomes under alternative policies, providing evidence on the long-term impact of dynamic retention strategies.

2:40-3:00

Xiaoxue Fan and Yangfei Chen

Lelys Bravo (session chair)

Spatial Modeling of Hazard Incidence in South Carolina

This project investigates the spatial distribution of hazard incidence across counties in South Carolina using spatial statistical models implemented in INLA. We use hazard event data to model the frequency of different hazard types and incorporate population-based offsets to account for variation in exposure. To capture spatial dependence between neighboring counties, we compare several models, including non-spatial Poisson, Besag, and BYM2 models.Model performance is evaluated using multiple criteria, including WAIC, PIT, and CRPS, to assess goodness-of-fit and predictive accuracy. The results reveal clear spatial patterns in hazard incidence and highlight the importance of incorporating spatial structure in modeling. While we also explored damage-related measures such as damage per capita, a full integration of hazard and damage components is left for future work toward comprehensive risk assessment.