Re(Visiting) Large Language Models in Finance

Hi! Here's Iván from Noax Capital 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:

• 📈 Noax Capital Updates : 📢 New Section! As our quantitative investment firm grows, I've decided to start sharing interesting updates through this newsletter. Today, I am sharing with you a summary of our main value proposition.

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

  • Re(Visiting) Large Language Models in Finance

  • Quantformer: from attention to profit with a quantitative transformer trading strategy

• 💊 AI Essentials: The section on top AI & Quant Finance learning resources: Today, I'm sharing an excellent free data sicence course offered by WorldQuant University (WQU) that guides students through six real-life, hands-on projects.

• 🥐 Asset Pricing Insights: In this edition, I introduce Statistical Arbitrage via Single-view and Multi-view Spectral Clustering on Mixed Frequency Data, a study that enhances statistical arbitrage strategies by using high-frequency intraday data. It combines realized estimators like volatility and beta with spectral clustering to more accurately cluster assets based on daily and intraday features.

NOAX CAPITAL: What we do?

We're launching a new section to discuss Noax Capital updates! In this case, I want to sum up our main value proposition: delivering mid-frequency, high-capacity strategies with Sharpe ratios above one and low market correlation for our institutional clients.

As proof of our models’ potential, the audited track record for Gestión Boutique VI / Noax Global has consistently shown Sharpe ratios above one since our systematic advisory began in January 2023, despite the strict retail investment fund restrictions in Europe.

For institutional mandates, please contact directly: [email protected]

“Re(Visiting) Large Language Models in Finance”

🔔 A look-ahead bias-free paper using LLMs in finance?👇

This paper presents FinText, a suite of specialized language models for accounting and finance, pre-trained on data from 2007 to 2023 to avoid look-ahead bias.

FinText outperforms larger models like LLaMA in trading applications, showing the value of domain-specific pre-training. It achieves a Sharpe ratio of 3.45 in an equal-weighted portfolio, surpassing other models, including LLaMA, FarmPredict, FinBERT, and LMD.

Key numerical results:

➡ FinText's base model (125 million parameters) achieves a Sharpe ratio of 3.45 in an equal-weighted portfolio.

➡The smaller version of FinText (51 million parameters) still performs strongly with a Sharpe ratio of 2.75.

➡FinText's superior performance remains valid even after accounting for transaction costs and adjusting for CAPM and Fama-French factor models.

➡FinText models consistently demonstrate strong performance across various trading sizes and yearly models.

➡While specific annual return figures are not provided in the given excerpt, the high Sharpe ratios indicate that FinText's performance is not only profitable but also risk-adjusted, outperforming larger models with approximately 50 times more parameters.

“Quantformer: from attention to profit with a quantitative transformer trading strategy”

The paper in a nutshell ⤵

▶The paper introduces "quantformer", an enhanced neural network architecture based on transformers, designed specifically for building investment factors in quantitative trading.

▶Quantformer adapts the transformer model to handle numerical inputs directly, allowing it to process financial data and predict future stock returns over a given period.

▶The study collected over 5,000,000 rolling data points from 4,601 stocks in the Chinese capital market from 2010 to 2019 for model training and testing.

▶Quantformer demonstrated superior performance in predicting stock trends compared to 100 other factor-based quantitative strategies.

▶The model innovatively combines transformer-like architecture with market sentiment information to enhance the accuracy of trading signals.

▶The researchers tested quantformer's performance across different data frequencies and training scales to comprehensively evaluate its capabilities.

▶The implementation details and code for quantformer are made available on GitHub, promoting reproducibility and further research in this area.

AI-Essentials

Excellent free data science course offered by WorldQuant University (WQU) that guides students through six real-life, hands-on projects. The course covers essential topics including database management, data cleaning and preprocessing, regression and classification modeling, data visualization, ethics in machine learning, and business insight & intelligence. It's a comprehensive resource for anyone looking to gain practical experience in applying data science to real-world business and financial problems.

Asset Pricing Insights

“Statistical Arbitrage via Single-view and Multi-view Spectral Clustering on Mixed Frequency Data”

👉 What’s New? This paper introduces a powerful approach that uses high-frequency intraday data to enhance statistical arbitrage strategies. By combining realized estimators (like volatility and beta) with single-view and multi-view spectral clustering, it clusters assets more accurately based on daily and intraday features.

👉 Assets are clustered based on various features, from closing prices to realized volatility and betas, creating “similar” asset groups. These clusters then guide trading strategies like the distance method, cointegration, copulas, and the Ornstein-Uhlenbeck framework.

👉 Tested on assets like S&P 500 stocks, ETFs, and forex, multi-view clustering, which retains more feature information, consistently outperformed single-view clustering. The multi-view approach led to higher Sharpe ratios and reduced drawdowns compared to traditional methods.

👉 By integrating diverse data at multiple frequencies, this clustering approach shows strong potential for improving statistical arbitrage returns, particularly in high-liquidity assetsk to interpretable factors, making it practical for real-world investment applications.

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! 🌟🤖💼