Stock forecasting with multi-modality graph neural network

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:

  • Financial time series forecasting with multi-modality graph neural network

  • Macroeconomic Announcement and Machine Learning for Asset Pricing

  • One Factor to Bind the Cross-Section of Returns

• 💊 AI Essentials: The section on top AI & Quant Finance learning resources: Today, I'm thrilled to share an excellent video from the Diginomics Brownbag Seminar by Prof. Dr. Markus Pelger: "Deep Learning in Asset Pricing."

• 🥐 Asset Pricing Insights: In this edition, I recommend the study "Universal Portfolio Shrinkage," which addresses the question, "How to Optimize Portfolios in High-Dimensional Settings?" This study proposes an innovative idea that enhances out-of-sample (OOS) performance.

Today’s Sponsor

Learn How AI Impacts Strategy with MIT

As AI technology continues to advance, businesses are facing new challenges and opportunities across the board. Stay ahead of the curve by understanding how AI can impact your business strategy.

In the MIT Artificial Intelligence: Implications for Business Strategy online short course you’ll gain:

• Practical knowledge and a foundational understanding of AI's current state

• The ability to identify and leverage AI opportunities for organizational growth

• A focus on the managerial rather than technical aspects of AI to prepare you for strategic decision making

LEARN MORE

AI-Finance Insights

“Financial time series forecasting with multi-modality graph neural network”

The paper introduces a novel model called "MAGNN" that integrates various data sources—such as historical prices, news, and events—into a unified framework using graph neural networks.

The main contribution of the paper is the development of this model, which can interpret and predict financial market trends by leveraging the lead-lag relationships in the data.

Numerical results demonstrate the superior performance of MAGNN. Specifically, in terms of asset return (A Return), the model achieves 0.8571 compared to 0.3002 for Stock-LSTM, 0.3960 for News-ATT, 0.5035 for Stock-GAT, and 0.5507 for Event-NTN.

The average daily return (D Return) for MAGNN is 0.0027, significantly higher than the baselines. Additionally, the Sharpe ratio, a measure of risk-adjusted return, is 3.7619 for MAGNN, outperforming all baseline models.

“Macroeconomic Announcement and Machine Learning for Asset Pricing”

💡Can macro announcements and ML unify the complexities of asset pricing? Clever use of ensemble models and time-series variation, keep reading! 👂

The paper leverages machine learning to enhance return predictions by distinguishing between announcement (A-days) and non-announcement (N-days) trading days.

Utilizing separate models for A-days and N-days, and combining them into an ensemble, the study achieves remarkable results.

The ensemble model outperforms single-period models, delivering an average quarterly return of 5.66% and the highest Sharpe ratio of 1.03. This approach reveals time-varying predictability in asset returns.

Incorporating bond-related characteristics further improves the model, yielding a long-short spread portfolio return of 6.41%, surpassing traditional models. The model effectively captures market mispricing, especially on N-days, indicating its robustness.

“One Factor to Bind the Cross-Section of Returns”

Can a single-factor model simplify and unify the complexities of asset pricing? Nice application of the Kolmogorov-Arnold theorem! 💡 

The paper introduces a non-linear single-factor asset pricing model, leveraging the Kolmogorov-Arnold representation theorem to consolidate various multi-factor models into a single, simple framework.

The study employs a dataset encompassing 171 assets across major classes such as U.S. equity portfolios, international bonds, commodities, and foreign currencies.

Utilizing a sieve-based estimation method, the model achieves an impressive adjusted R-squared value of 89%, demonstrating its superior ability to explain cross-sectional differences in asset returns. The pricing error remains statistically insignificant, underscoring the model's accuracy.

When benchmarked against traditional models like CAPM, Fama-French three-factor and five-factor models, and PCA-based models, the new model consistently outperforms, even when additional factors are included.

The robust performance across various datasets and asset classes highlights its potential to significantly enhance our understanding and prediction of asset returns, offering a powerful tool for financial analysts and portfolio managers.

Sponsor 👇

Investing in the next household name

RYSE has just launched in 100+ Best Buy stores, and you're in luck. You can still invest at only $1.50/share before their name becomes known nationwide. They’ve patented the only mass market shade automation device and their exclusive Best Buy deal resembles that which led to other past billion-dollar buyouts like Ring and Nest. Learn more + earn bonus shares →

AI-Essentials

📹 Check out this insightful video from the Diginomics Brownbag Seminar by Prof. Dr. Markus Pelger: "Deep Learning in Asset Pricing." Enjoy over an hour of inspiring discussion! 👇

Asset Pricing Insights

“Universal Portfolio Shrinkage”

💡How to optimize portfolios in high-dimensional settings? The paper, proposes an idea that enhances OOS performance.

Leveraging the Universal Portfolio Shrinkage Approximator (UPSA), this study optimally reweights low-variance principal components (PCs) of returns instead of eliminating them.

This approach complements the traditional Markowitz portfolio with a complexity correction, delivering remarkable results.

UPSA outperforms existing methods, achieving an impressive out-of-sample Sharpe ratio of 1.23, compared to 0.98 for Ridge shrinkage and 0.88 for the classic Markowitz portfolio. By effectively reweighting PCs, UPSA addresses the estimation errors prevalent in high-complexity environments.

Incorporating low-variance PCs, the method proves robust, with outperformance increasing monotonically with the number of PCs used. Empirical tests on various stock size groups demonstrate that CUPSA, the constrained version of UPSA, maintains superior Sharpe ratios, particularly for mega stocks.

🔢 Key Results:

Sharpe Ratio: CUPSA achieves 1.23 versus 0.88 (Markowitz) and 0.98 (Ridge).

Portfolio Return: The CUPSA portfolio yields a quarterly return of 5.85%, outperforming traditional methods.

Stability: Performance remains stable when including up to 30 PCs, highlighting robustness to low-variance components.

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! 🌟🤖💼