Anru Joshua Colmenar

Quantitative Risk Management | Robust Portfolio Construction | Infrastructure

I am an MSc Quantitative Finance candidate at Rutgers Business School (May 2026), building upon a BSc in Finance and Economics from Kean University. My academic and professional journey is driven by a deep respect for the unpredictability of markets. I focus on building end-to-end risk pipelines that address the reality of heavy-tailed market returns, rather than relying on fragile predictive models. I am actively seeking roles in the Greater NYC/NJ area where I can continue to learn, refine my mathematical modeling, and contribute to scalable, production-ready systems.

Quantitative Risk & ML Pipelines

My core research philosophy is that classical Markowitz optimization is dangerously unstable with finite data. Below is a breakdown of my recent architecture for a "Tail-Risk Parity" portfolio. This project taught me a critical lesson: predicting returns is incredibly noisy, but systematically managing downside risk provides tangible value.

Read Technical Whitepaper View PyTorch Source Code
Neural Network Architecture

KRNN Return Predictor & Scalable Data Engineering

The Method: I engineered a K-parallel GRU encoder (KRNN) trained via Gaussian Negative Log-Likelihood to output next-day return $\mu$ and volatility $\sigma$. By mean-pooling across $K$ independent GRUs, I aimed to reduce the variance in the network's predictions.

The Math: The network optimizes the per-sample GNLL loss:

$$ \mathcal{L}(\mu,\sigma;y) = \frac{1}{2} \log(\sigma^2) + \frac{(y-\mu)^2}{2\sigma^2} $$
Engineering & Reality: To ensure code scalability and avoid "quiet" data leaks, I built a modular PyTorch pipeline utilizing Parquet for I/O efficiency, strictly separating the StandardScaler fit to the training chronologies. While the predictive alpha proved negligible ($R^{2} \approx 0$), isolating the standardized residuals ($Z_t$) provided the exact heteroscedastic signals needed for the downstream tail-risk engines.

Risk Optimization

DCMP Worst-Case Bounds & Mean-CVaR Optimization

The Method: Because financial tails are fatter than Gaussian models assume, I replaced traditional variance constraints with a Rockafellar-Uryasev Mean-CVaR Linear Program. To prevent the optimizer from exploiting sampling errors, I anchored the constraints using a Discrete Moment Problem (DCMP), which outputs the highest mathematically possible CVaR consistent with the asset's empirical moments.

The Math: By standardizing residuals via

$Z_t = \frac{y_t - \mu_t}{\sigma_t + \epsilon}$
I solved the following objective to minimize the portfolio's tail loss:
$$ \min \phi + \frac{1}{(1-\alpha)T} \sum_{t=1}^T z_t $$
The Result: The optimization naturally favored assets with lower tail indices ($\gamma$). Testing out-of-sample, this heavy-tail-aware allocation achieved a lower realized volatility (20.56%) and maintained a competitive Sharpe ratio without relying on spurious return predictions.

Fiscal Operations & Systems Infrastructure

A mathematically sound model is only as reliable as the infrastructure supporting it, and alpha is meaningless without stringent fiscal oversight. My background spans both enterprise-level budget management and hands-on Linux architecture.

  • Enterprise Fiscal Management: Operating as an Administrative Analyst for the NJ Department of Health (PHEL), I utilized Business Intelligence platforms to forecast and maintain financial reports across a $100M+ portfolio of State and Federal accounts.
  • Low-Latency Systems Architecture: Engineered and currently maintain a self-hosted, high-performance communications server running on a dedicated Linux environment.
  • Network Routing & CI/CD: Implemented strict UFW firewall protocols, dual-stack IPv4/IPv6 tunneling, and automated edge deployments utilizing Cloudflare’s global network and Git integration.
  • Certifications: Oracle Cloud Infrastructure, Business Intelligence Data Analytics, Statistics with Python.

Get In Touch

I am actively exploring full-time quantitative and risk-focused opportunities for post-graduation in May 2026. I am always looking to learn from experienced practitioners. Please feel free to reach out via email or LinkedIn to discuss market infrastructure, review my GitHub repositories, critique my research, or to simply just connect!