Timing Factor Models

Hi! Here's Iván with this week's exciting newsletter, brimming with insights and discoveries on building robust investment strategies and risk models using Machine Learning.

In this edition, I am presenting the following sections:

• 🕹️ AI-Finance Insights: I summarize three must-read academic papers that mix cutting-edge ML/DL with Asset Pricing & Quant Finance:

  • Timing Factor Models

  • Kernel Trick for the Cross-Section

  • Complexity in Factor Pricing Models

• 💊 AI Essentials: The section on top AI & Quant Finance learning resources: Today, I'm thrilled to share an excellent video from the Center for Financial Markets and Policy's (CFMP) Global Virtual Seminar Series on Fintech. In this session, recorded on January 29, CFMP brought together leading scholars to present their research on AI in Asset Management.

• 🥐 Asset Pricing Insights: In this edition, I recommend the study "Constructing Time-Series Momentum Portfolios with Deep Multi-Task Learning," which addresses the question, "How to construct time-series momentum portfolios with better risk-adjusted performance?" This study proposes a novel approach using Multi-Task Learning (MTL).

Today’s Sponsor

Missed out on Ring and Nest? Don’t let RYSE slip away!

Ring 一 Acquired by Amazon for $1.2B

Nest 一 Acquired by Google for $3.2B

If you missed out on these spectacular early investments in the Smart Home space, here’s your chance to grab hold of the next one.

RYSE is a tech firm poised to dominate the Smart Shades market (growing at an astonishing 55% annually), and their public offering of shares priced at just $1.50 has opened. 

They have generated over 20X growth in share price for early shareholders, with significant upside remaining as they just launched in over 100 Best Buy stores.

Retail distribution was the main driver behind the acquisitions of both Ring and Nest, and their exclusive deal with Best Buy puts them in pole position to dominate this burgeoning industry.

Secure your stake before the offer closes.

AI-Finance Insights

“Timing Factor Models”

💡 How to Time and Interpret Factor Models Effectively? The paper provides must-read practical insights. Keep reading!

The study provides some important results:

👉 Principal Component Analysis (PCA) Factors: Using 5 PC factors, the model significantly reduces the maximum (absolute) alpha, indicating a better fit. It achieves a maximum squared Sharpe ratio (SR) of 1.77 for anomaly portfolios, compared to 4.23 without interactions.

👉 Comparison with Fama-French Factors: The PC factor model performs comparably to the Fama-French model in explaining average returns for size and book-to-market (B/M) portfolios. The PC model achieves a squared SR of 0.69, slightly higher than the Fama-French model's 0.64.

👉 Pseudo Out-of-Sample Tests: In-sample Sharpe ratios (SRs) are higher than out-of-sample SRs, indicating potential overfitting. For anomaly portfolios, the in-sample SRs for the first 3 PCs are significantly higher than the out-of-sample SRs.

👉 Impact of Sentiment-Driven Demand: The model shows that even when deviations from the CAPM are driven by sentiment, the stochastic discount factor (SDF) can be approximated by a low-dimensional factor model with a few PC factors. This implies that a few dominant factors can explain most of the variation.

In summary, the paper highlights that the absence of near-arbitrage opportunities ensures factor models with a few dominant factors can explain the cross-section of expected returns.

This holds even in the presence of investor sentiment, challenging traditional distinctions in asset pricing models. The empirical results demonstrate that a few PC factors can capture the cross-sectional variation in expected returns of anomaly portfolios, performing comparably to well-known models like Fama-French.

“Kernel Trick for the Cross-Section”

📢 How can we enhance the efficiency of asset pricing models using ML? The paper proposes a novel approach using the kernel trick from ML.

Leveraging this technique, the study extends the set of characteristics to an arbitrary—potentially infinite—dimension, enhancing the modeling of non-linearities and interactions in characteristics.

The proposed kernel-based SDF (Stochastic Discount Factor) model delivers substantial improvements:

👉 Out-of-Sample Sharpe Ratio: 3.0, compared to 1.65 with no interactions.

👉 Simulated Factor Recovery: Correlation of 0.98 using kernel-based PCA, versus 0.13 with traditional PCA.

👉 Empirical Analysis: Cross-validated Sharpe ratios improve from 0.31 to 0.75 with the Fama-French-Carhart factors, and from 1.5 to above 3 with forty anomaly characteristics.

Additionally, the model maintains robust performance in out-of-sample tests, with the out-of-sample Sharpe ratio doubling compared to linear methods. The kernel method's advantage persists despite the overall deterioration of anomaly returns in recent years.

👉 Monthly Predictive R-Squared for Individual Stocks: 0.8%, compared to a benchmark of 0.2%.

👉 Daily Predictive R-Squared: 0.045%, compared to a benchmark of 0.01%.

The kernel-based approach significantly enhances the efficiency and predictive power of the SDF, offering a more robust framework for understanding the cross-section of expected returns.

This method demonstrates the importance of incorporating non-linearities and interactions in asset pricing models, resulting in better risk-adjusted performance and predictive capabilities.

“Complexity in Factor Pricing Models”

💡 How does complexity affect factor pricing models? The paper "Complexity in Factor Pricing Models" shows how highly parameterized ML/DL models lead to improvements in out-of-sample performance.

The study makes several key contributions:

👉 Virtue of Complexity: The paper demonstrates that asset pricing models with an extremely large number of factors (even more than the number of training observations) significantly outperform simpler models. This challenges the traditional belief that a small number of factors should capture the risk-return tradeoff.

👉 Empirical Results: Models with many factors achieve higher out-of-sample Sharpe ratios and lower pricing errors. High-complexity models (with more parameters than observations) consistently outperform traditional low-dimensional models, such as the Fama-French model.

👉 Sharpe Ratios: The out-of-sample Sharpe ratios of high-complexity models exceed those of simpler models. For example, the empirical "VoC curves" indicate that increasing the number of model parameters enhances the out-of-sample SDF Sharpe ratio and reduces pricing errors.

👉 Robustness: The improved performance of complex models is robust across different subsets of the stock universe, including large and liquid US stocks.

In summary, the paper highlights the significant advantages of using highly parameterized models (ML/DL) in asset pricing, showing that these models can capture more information and provide better out-of-sample performance compared to traditional, simpler models.

Sponsor 👇

Transform How You Work and Organize Information

Flot is your all-in-one and everywhere AI assistant, working across any websites and apps to help you write better, read faster, and remember information well. Everything is just one click away.（Available on Windows and macOS）

🎁 Lifetime Deal just $67 – Grab It Now! →

AI-Essentials

📹 Check out this insightful session from the Center for Financial Markets and Policy's (CFMP) Global Virtual Seminar Series on Fintech. CFMP was delighted to bring together scholars to present their research for AI in Asset Management Day. Enjoy over an hour of inspiring discussion! 👇

Asset Pricing Insights

“Constructing Time-Series Momentum Portfolios with Deep Multi-Task Learning”

💡 How to construct time-series momentum portfolios with better risk-adjusted performance? The paper "Constructing Time-Series Momentum Portfolios with Deep Multi-Task Learning" proposes a novel approach using Multi-Task Learning (MTL).

Leveraging a deep neural network architecture, this study jointly learns portfolio construction and auxiliary tasks related to volatility forecasting. This integrated approach enhances both return generation and risk management.

The proposed MTL-TSMOM model delivers substantial improvements:

👉 Annualized Return: 7.90%, compared to 5.54% for TSMOM and 1.66% for CTA-MOM.

👉Sharpe Ratio: 0.81, outperforming 0.59 for TSMOM and 0.21 for CTA-MOM.

👉Sortino Ratio: 1.20, versus 0.83 for TSMOM and 0.31 for CTA-MOM.

Additionally, the model maintains a lower maximum drawdown of -21.00% compared to CTA-MOM’s -41.63%, though slightly higher than TSMOM’s -19.88%. Importantly, the MTL-TSMOM model significantly reduces drawdown recovery time, enhancing its attractiveness to investors.

Ablation studies show that including all auxiliary tasks yields the best performance. The MTL-TSMOM model’s ability to maintain a low correlation with equity markets (-0.24 to 0.28) makes it an excellent diversifier in investment portfolios, especially during market stress.

If you're enjoying our newsletter and want to support us, please recommend it to anyone you know who's interested in AI and Finance. Your referrals are the biggest compliment and help us grow! 🌟🤖💼