CS & Mathematics @ University of Maryland — Executive Board @ Maryland Quant Finance Club — Research Fellow @ Startup Shell
class PaulVega:
def __init__(self):
self.education = "Computer Science & Mathematics @ University of Maryland"
self.focus = "Quantitative research, time series Machine Learning, and research systems"
self.roles = [
"Executive Board Member @ Maryland Quant Finance Club",
"Research Fellow @ Startup Shell"
]
self.email = "vega.paul.pfv@gmail.com"
def research_interests(self):
return [
"Quantitative research and hypothesis testing",
"Time series classification and forecasting",
"Labeling and validation methodology",
"Statistical robustness and leakage control",
"Research-to-production infrastructure"
]
def current_goal(self):
return "Build research that survives contact with reality."I build research systems end to end: from hypothesis design and labeling methodology, through statistically validated ML experiments, to deployed software. My forte is quantitative research on sequential data, but the work spans open-source Python packages, desktop tooling, and cloud research infrastructure.
A model that looks good in a notebook is a hypothesis, not a result. The interesting problems are honest validation, leakage control, and knowing when a system is learning signal versus memorizing noise.
Original research on classification when the reward is a rule-constrained payout game rather than symmetric profit-and-loss. Shows why symmetric triple-barrier labels fail under convex payoff constraints, develops a label-health screening method that avoids economic overfitting, introduces an account-level bootstrap that treats the account itself as a second triple-barrier problem, and documents a forward-test failure that motivated a stricter eight-stage validation standard — same-count random baselines, probability of backtest overfitting, sealed quarantine holdouts, and sentiment/regime agreement gates.
Model-free Python validation framework born directly from the paper above. Audits whether a label beats same-count random paths, fits selectors inside purged combinatorial cross-validation with block bootstrap, scores sealed out-of-sample quarantine data, and stress-tests economics with barrier-based Monte Carlo account paths. Deterministic, seeded, and schema-agnostic via column role mapping. Published on PyPI.
End-to-end research infrastructure for sequential prediction: event-driven candidate generation (CUSUM filters, information-driven bars), gradient-boosted classifiers tuned with Optuna, hidden Markov regime models trained walk-forward, decision-time sentiment scoring with leakage audits, purged cross-validation at scale, and live inference with regime routing and feature-freshness monitoring. Backed by PostgreSQL/Supabase storage and AWS compute.
Desktop research tool (Python/Tkinter) that combines data loading, a modular auto-discovered strategy catalogue, discrete trade simulation, Monte Carlo risk analysis (VaR, CVaR, equity confidence bands), and grid-search parameter optimization in one interface. Engine validated against TradingView with near-identical equity curves and trade statistics.
Languages
Machine Learning & Data
Modeling toolkit: gradient boosting, feedforward/recurrent networks, hidden Markov models, bootstrap and Monte Carlo methods, purged combinatorial cross-validation, probability calibration, lexicon-based NLP sentiment.
Infrastructure & Tools
Open to conversations about quantitative research, time series machine learning, research tooling, and applied AI.