Optimizing Risk and Return: Statistical & Volatility Modeling of Trading Strategies

Spring 2025

Researcher

Vaishnavi Sampath

Faculty Advisor

Professor Zhenyu Cui

Overview

This project develops and evaluates trading strategies for major energy-sector stocks by combining statistical forecasting, volatility modeling, macroeconomic analysis, and machine learning. The main objective is to identify strategies that can improve returns while controlling risk, with special attention to how volatility influences performance.

The analysis focuses on Exxon, Chevron, Shell, British Petroleum, and National Fuel Gas, using benchmark data from WTI, Brent crude, Henry Hub natural gas, the S&P 500, and macroeconomic indicators. The project applies models such as ARIMA, GARCH, SVAR, CAPM, logistic regression, KNN, and Random Forest, then compares multiple trading strategies including trend following, pairs trading, momentum, and mean reversion.

The overall goal is to determine whether these models and strategies can outperform traditional approaches in terms of risk-adjusted returns, forecasting reliability, and practical trading performance.

Key Findings

1. Market and Model Validation.

• The statistical analysis confirmed that the stock and benchmark return series were generally stationary, making them suitable for time-series modeling.
• ARIMA models were able to capture short-term return behavior for several assets, though many price forecasts remained relatively flat, showing the limits of traditional forecasting in directional prediction.
• GARCH models were effective in identifying volatility clustering, especially for major oil-linked equities and benchmarks.
• The analysis showed that oil-related stocks were strongly influenced by broader energy market movements, with stocks like Exxon and Chevron reacting closely to WTI, and Shell and BP reacting more to Brent.

2. Impact of Volatility and Energy Shocks

• Volatility played a major role in determining the effectiveness of both forecasting models and trading strategies.
• The SVAR analysis showed that shocks in WTI, Brent, and Henry Hub affect macroeconomic indicators, industrial production, price indices, and carbon-related variables, though many of these effects were short-lived or moderate.
• WTI and Brent had stronger and more persistent spillover effects across economic and industrial variables than Henry Hub, which tended to have more localized influence.
• Energy shocks were shown to matter not only for stock returns, but also for production, inflation-related indices, and sector-level forecast uncertainty.

3. Trading Strategy Methodology and Performance

• SMA-based Trend Following performed best on Exxon and National Fuel Gas, where stronger persistent trends allowed the strategy to generate better cumulative returns and Sharpe ratios.
• Pairs Trading between Exxon/Chevron and Shell/BP produced moderate results, but the lack of statistically strong cointegration weakened the long-term mean-reversion assumption behind these trades.
• The KNN-based Momentum Strategy performed poorly overall, with low win rates, frequent false signals, and weak Sharpe ratios across most stocks.
• The Random Forest Mean Reversion Strategy produced mixed outcomes, with Exxon and Shell showing the best relative performance, while Chevron and BP struggled with weak risk-adjusted returns and larger drawdowns.

4. Comparative Insights

• Traditional statistical models such as ARIMA and GARCH were useful for understanding return structure and volatility, but they had limited ability to produce strong directional forecasts on their own.
• Rule-based strategies like SMA trend following were more successful when the underlying asset exhibited clear trends and manageable volatility.
• Machine learning models did not consistently outperform simpler methods; in some cases, they generated excessive trades, low win rates, and unstable performance.
• Across the project, the most important lesson was that risk control, volatility awareness, and model-asset fit matter more than model complexity alone.

Conclusion

This project shows that profitable trading is not simply about using advanced models, but about selecting strategies that align with the behavior, volatility profile, and structural characteristics of each asset. Statistical models provided a strong analytical foundation, while volatility modeling helped explain why some strategies worked under certain market regimes and failed under others.

The findings suggest that trend-following approaches can work well in persistent markets, while mean-reversion and machine learning strategies require stronger assumptions, better feature design, and more robust validation to be reliable in practice.

Future research should:

• incorporate transaction costs, slippage, and real-world trading frictions
• use dynamic signal thresholds and stronger risk management tools
• include additional macroeconomic and fundamental variables
• apply walk-forward optimization more extensively
• explore news, sentiment, and NLP-based event signals
• further refine machine learning models to better adapt to changing market regimes

Significance

This project is significant because it connects quantitative finance, volatility modeling, macroeconomic analysis, and algorithmic trading in one integrated framework. It demonstrates that strong trading performance depends not just on prediction accuracy, but on how well models respond to risk, changing market conditions, and asset-specific dynamics.

By comparing classical statistical models, economic structure models, and machine learning strategies, the project provides a practical view of what works, what fails, and why. It offers useful insight for designing more robust trading systems that aim to improve risk-adjusted returns rather than raw returns alone.