ANAY ABHIJIT JOSHI
VIDHI CHATURVEDI
SRIYA NELLURI
HIMANSHU BHATT

StochastiQ

Multi-Model Portfolio Optimization with Stochastic Simulation, Options Overlay, and Regime-Conditional Validation

MGT 6081-A · Spring 2026 Professor Satyajit Karnik Georgia Tech MS-QCF 100 / 100
8
Phases
7
Assets
4
Stochastic Models
5K
MC Paths/Model
36
Pages of Report
21
Academic Refs

What we built.

A complete falsifiable empirical pipeline for multi-model portfolio construction, from raw data through statistically rigorous out-of-sample validation.

We selected a seven-asset universe (AAPL, MSFT, JPM, JNJ, XOM, SPY, GLD) spanning equity sectors, geography, and asset classes. We pulled five years of daily price data via yfinance (1,257 training days from 2020 through 2024), holding back the most recent 16 months (332 trading days) as a strict out-of-sample window - never touched during model fitting.

On the training period we calibrated four stochastic-process models for each ticker: Geometric Brownian Motion, Merton jump-diffusion, Constant Elasticity of Variance, and Heston stochastic volatility. We then ran 5,000 correlated Monte Carlo paths per model with Cholesky-injected cross-asset correlation, and built seven optimized portfolios on top: per-model maximum-Sharpe allocations, plus three model-robust formulations (min-max worst-case, equal-blend, KS-weighted blend).

We added a Black-Scholes-Merton options overlay with three classical strategies (covered call, protective put, collar), then validated everything out-of-sample using a stationary block-bootstrap with paired Sharpe-difference inference and pre-registered decision rules. Finally, we conditioned the entire analysis on market regime via a two-state Hidden Markov Model anchored to the VIX, applied Holm-Bonferroni multiple-testing correction across 14 paired tests, and decomposed performance into Calm-day vs. Stress-day CVaR.

The project frames as Project Idea 6 (independent direction, instructor-approved) and incorporates the full blueprint of Project Idea 1 while substantially extending it on the validation, regime-conditioning, and coherent-risk-measure axes.

Eight phases, one cohesive system.

Each phase is a self-contained, runnable contribution committed and pushed independently. The full Git history is preserved in the repository.

PHASE 01
Data Ingestion
Pulled 5 years of daily adjusted close prices for 7 assets via yfinance. Computed log returns, sanity-checked stylized facts (vol clustering, fat tails, drawdowns).
yfinance · pandas · numpy
PHASE 02
Markowitz Baseline
Mean-variance frontier with five canonical portfolios: Max Sharpe, Min Variance, Max Sortino, Min CVaR (95%), Risk Parity. Used as reference, not production.
scipy.optimize · SLSQP
PHASE 03
Stochastic Calibration
Calibrated GBM, Merton jump-diffusion, CEV, and Heston for each ticker. MLE for GBM/Merton, method-of-moments for CEV, moment-matching for Heston with Feller verification.
154 calibrated parameters · 4 models × 7 assets
PHASE 04
Robust Optimization
5,000 correlated MC paths per model. Per-model Max-Sharpe portfolios + three robust aggregations: min-max worst-case, equal-blend, KS-distance-weighted blend. 30% per-asset cap.
Cholesky correlation · Ben-Tal-Nemirovski
PHASE 05
Options Overlay
Black-Scholes-Merton pricing for three classical strategies: 30-day covered call (K=1.05·S₀), protective put (K=0.95·S₀), collar (zero-cost combination). Greeks computed analytically.
BSM · Greeks · put-call parity
PHASE 06
OOS Validation
Paired stationary block-bootstrap (B=5,000) on Sharpe difference vs Equal-Weighted benchmark. Pre-registered decision rules (Validated / Falsified / Underpowered).
Politis-Romano · 332-day OOS
PHASE 07
Regime Analysis
Two-state Gaussian HMM on SPY returns. External validation against VIX (+15.7-pt premium). Regime-conditional Sharpe + CVaR. Holm-Bonferroni across 14 paired tests.
hmmlearn · forward-backward filter
PHASE 08
Synthesis & Report
Cross-phase aggregation notebook + 36-page LaTeX report with 17 figures, 9 tables, 21 academic references. Honest disclosure of negative findings and 7 explicit limitations.
LaTeX · pdflatex + bibtex

Same data. Different verdict.

The central methodological insight of the project: a strategy designed to manage state-contingent risk must be evaluated using a state-contingent risk metric. Sharpe ratio averages over both regimes; CVaR conditions on the left tail.

Phase 6 · Sharpe Ratio
Statistically null on Sharpe
Δ = −0.443
95% CI = [−1.40, +0.21]
p-value = 0.288
Verdict = Underpowered
The textbook DeMiguel-Garlappi-Uppal puzzle. With only 332 OOS days, estimation noise dominates whatever true performance gap exists between Min-max and the naïve 1/N benchmark. We report this honestly.
Phase 7 · CVaR · 95%
All 7 portfolios beat the benchmark on Stress CVaR
Per-Heston: +62 bps (best)
Per-Merton: +60 bps
Min-max: +53 bps
Calm CVaR: all worse (cost of insurance)
The textbook signature of a tail-risk-insurance design: better stress-day tail loss at the cost of marginally worse calm-day tail loss. Anchored in the coherent-risk-measure framework of Artzner et al. (1999).
"Which models do you think are more pertaining for the future?" — Project Idea 1
Heston stochastic volatility, with Merton jump-diffusion a close second.
🥇 HESTON · 62 BPS 🥈 MERTON · 60 BPS GBM/MIN-MAX · 53 BPS CEV · 4 BPS

Built end-to-end in Python.

Production-grade source code under src/, eight Jupyter notebooks orchestrating the pipeline, fully reproducible from a single seed.

Language
Python 3.12
Data
yfinance · pandas
Optimization
scipy.optimize · SLSQP
HMM
hmmlearn
Storage
parquet · pyarrow
Plotting
matplotlib · seaborn
Notebooks
JupyterLab
Report
LaTeX · pdflatex + bibtex
🙏 Thank you, Professor Satyajit Karnik, for an amazing semester...

I (Anay Abhijit Joshi) would like to express my sincere gratitude to Professor Satyajit Karnik. MGT 6081-A fundamentally changed the way I think about derivative securities and state-contingent risk. The framework you taught us, like specifically coherent risk measures, stochastic-process modeling, and the rigor required to evaluate state-dependent payoffs, etc., is what made this entire project possible...

Coming from a Computer Science background, your lectures provided the necessary bridge to help me develop a genuine passion for finance... Every empirical decision in this work, from the choice of CVaR over Sharpe to the regime-conditional decomposition, traces back directly to the course goals of understanding pricing methodologies, risk-neutral frameworks, and simulations using stochastic differential equations.

Thank you so much for your guidance and for demonstrating how to apply complex financial theory to specific, real-world examples...