Research

Problem

Graph neural networks combined with meta-learning are increasingly proposed for financial forecasting and often reported to achieve strong risk-adjusted returns. Such results are extremely sensitive to evaluation methodology, and standard practice in this literature routinely violates temporal, informational, and accounting boundaries.

Method

MetaGraph couples a causal temporal encoder with a Meta-Relational Prior Network that conditions a biased graph-attention layer on time-varying correlation graphs of S&P 500 constituents, and adapts to the prevailing regime through first-order MAML. Training and evaluation target the cross-sectional Information Coefficient. A leakage-free protocol combines strict temporal splits with an embargo, past-only graphs, inference-time adaptation that mirrors training, non-overlapping holding-period accounting, and significance testing via stationary block bootstrap and deflated Sharpe ratio.

Contributions

• Leakage-free evaluation protocol for meta-learning GNNs on financial time series, with four specific safeguards.
• Documentation of four common evaluation errors (non-temporal splitting, overlapping-return accounting, ill-posed objectives, train/inference adaptation mismatch) that inflated an earlier version of this pipeline to an implausible Sharpe above 10.
• Honest test-set result under the corrected protocol, with full audit trail and reproducible pipeline.
• Co-authored with Sumitaa S Deshbhandari (Independent Researcher).

Problem

Classical allocators minimize variance and treat drawdown depth and recovery time as constraints rather than first-class objectives.

Method

Models drawdown geometry and recovery stability directly. Allocates capital to maximize resilience rather than minimize variance. Evaluated on 31 years of multi-asset data with out-of-sample validation across independent splits.

Contributions

• Sharpe 1.54 and max drawdown 9.11% versus buy-and-hold 0.52 and 55.19%.
• Resilience-first objective that admits nonlinear drawdown penalties.
• Out-of-sample validation across independent splits.

Problem

Blink-phase segmentation from vertical EOG signals supports dry eye assessment but is sensitive to morphological variation across cohorts.

Method

A CNN-BiGRU sequence model trained on vertical EOG traces to segment blink phases across normal and dry-eye cohorts. Strict leave-one-subject-out validation across 43 subjects.

Contributions

• CNN-BiGRU sequence model that generalises across normal and dry-eye morphologies.
• Leave-one-subject-out validation across 43 subjects.
• First-author work conducted during the Indiana University Bloomington research internship.

Problem

Policy-gradient methods degrade under non-stationary reward landscapes due to limited exploration diversity.

Method

A swarm-style exploration loop on top of policy-gradient RL. The agent maintains a population of candidate directions instead of relying on a single gradient step. Robustness tested under shifting reward structures.

Contributions

• Swarm-augmented exploration robust to non-stationary rewards.
• Drop-in compatible with standard policy-gradient pipelines.
• Empirical gains in regimes where vanilla policy gradient methods fail.

Problem

Top-k retrieval over-fetches documents, increasing latency and compute cost. The overhead is particularly costly on edge devices.

Method

A PPO-trained selector replaces top-k with adaptive retrieval. The reward is shaped by downstream generation quality and on-device latency. Drop-in compatible with existing LLMs and retrievers.

Contributions

• Document selection in RAG framed as a reinforcement learning problem with sparse, cost-aware rewards.
• Edge-feasible deployment without modification to underlying models.
• Selected as Best Paper Candidate at ICEdge 2025 (lecture track).

Problem

Early-stage startups have access to AI tooling but lack frameworks to convert it into strategic advantage.

Method

Mixed-methods analysis across healthcare, finance, retail, and education case studies, synthesised with recent AI entrepreneurship literature into a single framework for AI adoption in small teams.

Contributions

• A framework linking AI/ML adoption to strategic agility, operational efficiency, and innovation capacity.
• Identification of technical, ethical, and regulatory barriers with proposed mitigations.
• Best Paper Award at ICEI 2025.

Problem

Several hypercube variants lack closed-form expressions for the Wiener index.

Method

Combinatorial derivation of exact Wiener index expressions for hypercubes and selected variants.

Contributions

• Closed-form Wiener index expressions for hypercubes and variants.
• Verifiable proof structure with explicit bookkeeping.
• Connections to network distance metrics in graph learning.

Problem

Imitating black-box trading strategies without recovering the underlying objective fails under distribution shift, and the field's headline improvements are often confounded with turnover reduction rather than strategy quality.

Method

Two-stage pipeline. Stage one performs behavioural cloning into a sparse linear surrogate, achieving return correlation of 0.92 to 1.00 across two 24-asset universes. Stage two applies cost, volatility, and drawdown re-engineering on the recovered surrogate. Evaluated on historical equity and derivatives data with controlled baselines that separate policy quality from turnover effects.

Contributions

• Transparent decoding of opaque strategies via sparse linear surrogates with near-perfect return correlation.
• Honest attribution: cost improvements are shown to stem from turnover reduction rather than the method itself.
• Identifiability characterisation for reward recovery under this class of setups.

Problem

Strategy crowding is discussed qualitatively in finance but rarely measured directly in a controlled reinforcement-learning setting. The question of whether population-level correlation actually erodes realized returns is largely empirical, and existing evidence is fragmented.

Method

Population of 50 RL trading agents evaluated in a dual-engine testbed combining synthetic markets and block-bootstrapped S&P 100 data. Introduces a diversification gain metric that links population-level pairwise correlation to realized pooled returns.

Contributions

• A correlated population (0.48 mean pairwise correlation) realizes Sharpe of -0.47 despite individual agents earning Sharpe 0.40, quantifying crowding-driven erosion directly.
• Shows that reducing correlation does not monotonically improve pooled returns, complicating naive diversification arguments.
• Reproducible dual-engine testbed for population-level RL trading experiments.

Problem

The original REBOUND framework treats resilience as an allocation objective but does not explicitly govern exposure through regime transitions, leaving the framework silent on when to scale down aggregate risk.

Method

Extends REBOUND with regime-conditional exposure governance and additional ablations beyond the WFC presentation. Retains the drawdown-geometry allocation core; adds a resilience-driven exposure controller that adjusts aggregate risk taking rather than relative asset weights.

Contributions

• Regime-conditional exposure governance built on top of the resilience allocation framework.
• Extended ablation study clarifying which parts of REBOUND drive the reported drawdown reduction.
• Direct follow-up to REBOUND (WFC 2026).

Problem

Basel traffic light backtests are the regulatory standard for VaR model validation, but their coverage guarantees do not carry over to machine-learning VaR models selected from many candidates, and to models with clustered violations.

Method

Empirical demonstration that Basel traffic light backtests pass LSTM VaR models that fail a selection-aware copula backtest. The copula backtest is drawn from a broader selection-aware backtesting codebase and applied to a range of ML VaR candidates.

Contributions

• Basel cherry-picking is shown to inflate apparent pass rates by 12 to 17 percentage points relative to selection-aware alternatives.
• LSTM clustering-aware pass rates fall 33 to 42 percentage points below Basel pass rates.
• Practical case that regulatory backtests need supplementary selection-aware and clustering-aware coverage tests when ML VaR is in use.

Problem

Human portfolio managers carry well-documented cognitive biases that large language models do not share, motivating the substitution of an LLM for the biased human interpreter inside a portfolio agent.

Method

ORACLE fuses an LLM observer, a sentiment block, and Hidden Markov regime estimates into a single belief vector consumed by a regime-conditioned PPO executor. Evaluated with walk-forward analysis on a multi-asset universe against a fair classical baseline.

Contributions

• A negative result: the LLM-augmented agent fails to beat a fair classical baseline, with an ablation isolating where each component helps or hurts.
• A causal audit showing the LLM signal is heeded but miscalibrated, which suggests some human biases are adaptive priors a general-purpose LLM does not carry.
• A reproducible benchmark and evaluation protocol for LLM-augmented portfolio agents.

Problem

Most edge RAG systems apply a single fixed retrieval policy to every query, paying latency and memory costs a given generator may not convert into better answers. Retrieval value depends on the deployed generator's parametric coverage, faithfulness to retrieved evidence, and noise tolerance.

Method

ATTUNE profiles an unknown, black-box generator at deployment time along four behavioural dimensions (retrieval faithfulness, noise robustness, context-window utilisation, parametric coverage) and uses that profile with lightweight query features to route each query over a discrete action space. Actions span a two-tier index (product-quantised and flat), variable depth, and optional reranking. A contextual-bandit controller learns on-device when retrieval is worth its cost.

Contributions

• First RAG framework to condition retrieval on a deployment-time behavioural profile of a black-box generator.
• Two-tier index with a 50x compressed tier and a high-recall flat tier, and a contextual-bandit controller over routing actions.
• Across four instruction-tuned generators (1B to 3.8B parameters), two quantisations, and four QA benchmarks: a single untuned controller matches the strongest fixed retrieval baseline; profiles transfer across corpora and calibrate within about 50 probes.

Problem

The conference version of ARISE established empirical robustness. The journal extension requires theoretical analysis of when and why swarm-perturbed policy gradients converge.

Method

Convergence and exploration-exploitation trade-off analysis for swarm-based perturbation in policy-gradient RL, paired with expanded empirical evaluation across non-stationary benchmarks.

Contributions

• Convergence and exploration-exploitation analysis for swarm-perturbed policy gradients.
• Expanded empirical sweep on non-stationary benchmarks.
• Journal extension of ARISE (WCSC 2026), in preparation.