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 Quant Finance:

  • Support for Stock Trend Prediction Using Transformers and Sentiment Analysis

  • Portfolio Construction Through Deep Reinforcement Learning and Interpretable AI

  • Missing Values Handling for Machine Learning Portfolios

• 💊 AI Essentials: The section on top AI & Quant Finance learning resources: Today, I'm thrilled to present a great video with hands-on guidance and complete code for using Deep Learning Momentum with TFT for trading. This must-watch video offers over an hour step-by-step instructions, helping you understand and apply these advanced techniques in your trading strategies.

• 🥐 Asset Pricing Insights: In this edition, I recommend the study "Machine Learning vs. Economic Restrictions: Evidence from Stock Return Predictability," which examines the effectiveness of machine learning models compared to traditional economic restrictions.

Today’s Sponsor

Advertise in newsletters, like this one, with the beehiiv Ad Network.

Wondering how you can sponsor hundreds of newsletters at once, without spending hours managing campaigns? 

The beehiiv Ad Network delivers 800 million impressions a month across hundreds of the world's top newsletters, paid on performance. 

Our managed service enables you to unlock newsletters as a growth channel without the added time. 

Connect with the team

AI-Finance Insights

“Support for Stock Trend Prediction Using Transformers and Sentiment Analysis”

How do Transformer architecture and sentiment analysis improve stock trend prediction?

The paper introduces StockFormer, a model that outperforms traditional RNN methods, offering enhanced accuracy and deeper insights for long-term trading strategies.

Main ideas in a nutshell 🔽 

✴ The paper introduces a novel model, StockFormer, which leverages Transformer architecture and sentiment analysis for stock trend prediction. This approach aims to overcome the limitations of Recurrent Neural Networks (RNNs), such as gradient vanishing and long-term dependency issues, particularly when dealing with long sequence data.

✴ StockFormer combines technical stock data and sentiment analysis from news headlines to predict stock trends over extended periods. This dual-input model provides a more comprehensive understanding of market conditions by incorporating both quantitative data and qualitative news sentiment.

✴ The dataset used in the study includes daily technical stock data and top news headlines for FAANG companies over nearly three years, from March 2019 to February 2022. Sentiment scores were extracted using FinBert, a financial text analysis tool.

✴ Experimental results show that StockFormer outperforms traditional RNN models like LSTM and GRU in terms of directional accuracy, especially as the prediction window lengthens. For instance, StockFormer's directional accuracy improves by up to 18.63% over a 30-business-day prediction window compared to RNNs.

✴ The study concludes that combining sentiment analysis with technical data using Transformer architecture can significantly enhance stock trend prediction, offering a robust tool for long-term trading strategies. The findings support the potential of advanced deep learning models in capturing long-term dependencies and improving predictive performance in financial markets.

“Portfolio Construction Through Deep Reinforcement Learning and Interpretable AI” 

Another interesting architecture leverages Transformer Encoders + Cross-Asset Attention Networks. I’ve seen several papers with this approach, and all have shown promising results. It might be worth a read!

Main paper takeaways ⤵

✴ The paper introduces a novel portfolio management approach, AlphaPortfolio, using deep reinforcement learning (RL) to directly optimize portfolio objectives without the need for traditional return distribution estimation.

✴ AlphaPortfolio leverages advanced sequence representation extraction models (SREM) such as Transformer Encoder (TE) and Long Short-Term Memory (LSTM) with historical attention, combined with Cross-Asset Attention Networks (CAAN), to capture complex asset interrelationships.

✴ The model addresses the "black-box" issue in AI by employing economic distillation techniques, utilizing polynomial-feature-sensitivity analysis and textual-factor analysis to interpret the key drivers of investment performance.

✴ Extensive out-of-sample performance evaluations demonstrate that AlphaPortfolio achieves Sharpe ratios consistently above 2.0 and an annualized risk-adjusted alpha exceeding 13% with monthly rebalancing, outperforming traditional machine learning and anomaly-based strategies.

✴ The robustness of AlphaPortfolio is evident across various market conditions and economic restrictions, maintaining superior performance metrics, including lower turnover and better maximum drawdown figures compared to high-frequency machine learning strategies and traditional trading anomalies.

✴ These results underscore the potential of deep RL in finance, showcasing significant enhancements in portfolio management and providing hedge funds, mutual funds, and proprietary trading firms with a powerful tool for superior investment performance while ensuring model interpretability and trustworthiness.

“Missing Values Handling for Machine Learning Portfolios”

The paper investigates the handling of missing values in machine learning-based portfolio construction using 159 cross-sectional return predictors.

Sum up of the main findings 👇

💊 The study finds that simple cross-sectional mean imputation performs comparably to more sophisticated methods like expectation-maximization (EM) due to the nature of the missing data, which often occurs in large, time-organized blocks with low cross-sectional correlation.

💊 Both mean and EM imputation methods yield similar results in terms of expected returns, with sophisticated imputations sometimes introducing noise that can lead to underperformance if not carefully managed.

💊 Simple mean imputation combined with machine learning models such as neural networks and principal component regressions can deliver high returns, up to 66% per year for equal-weighted portfolios and 39% for value-weighted portfolios.

💊 For practical applications, the paper recommends using simple mean imputation due to its transparency, tractability, and competitive performance compared to more complex methods. These insights are crucial for practitioners aiming to enhance the reliability and performance of machine learning-based trading strategies while efficiently handling missing data.

💊 Overall, the research provides valuable guidance for hedge funds, mutual funds, and proprietary trading firms on the effective use of imputation techniques in portfolio management, highlighting the benefits of simplicity and robustness in predictive modeling.

Sponsor 👇

A Newsletter You’ll Love: Stay ahead on the business of AI

Need ONE newsletter to stay on top of everything?

Prompts daily will make you more productive at work and help you look smarter in front of your friends (seriously!)

It’s also FREE to sign up!

In just ONE click, you can sign up for the daily newsletter (free)

AI-Essentials

Great video with hands-on guidance and complete code for using Deep Learning Momentum with TFT for trading. Over an hour of step-by-step instructions. 👇

Sponsor 👇

Get smarter on AI in 5 minutes a day.

• The world’s largest AI newsletter, read by over 600,000 AI professionals.

• One free email every morning on what’s new in AI and gives you “the rundown” of the most important developments.

• Allowing for readers to keep up with the insane pace of AI and why it actually matters

Subscribe with one click.

Asset Pricing Insights

“Machine Learning vs. Economic Restrictions: Evidence from Stock Return Predictability”

Can machine learning models outperform traditional methods in stock return prediction?

This paper by Avramov, Cheng, and Metzker (published in Management Science) examines the effectiveness of machine learning models compared to traditional economic restrictions.

The study uses techniques like neural networks, feed-forward networks, recurrent neural networks with LSTM cells, and generative adversarial networks to predict returns for a broad sample of U.S. stocks from 1987 to 2017.

The findings show that machine learning models can achieve high monthly returns without restrictions, with the NN3 model delivering an average value-weighted long-short portfolio return of 1.56% per month and an FF6-adjusted return of 0.92%.

However, when economic restrictions like excluding microcaps and distressed firms are applied, performance significantly drops. For example, the value-weighted FF6-adjusted return for the NN3 model decreases by 66% after excluding microcaps.

Additionally, high turnover rates exceeding 87% per month suggest substantial trading costs, affecting the strategies' economic viability.

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! 🌟🤖💼