LLMs for Time Series: Breaking New Ground in Stock Market Prediction

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:

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

  • LLMs for Time Series: Breaking New Ground in Stock Market Prediction

  • Deep Learning for Options Trading: An End-to-End Approach

• 💊 AI Essentials: The section on top AI & Quant Finance learning resources: Today, I’m sharing an insightful video titled "Recent Advancements of Financial Machine Learning with an Emphasis on Large Language Models." This talk provides a deep dive into the latest machine learning breakthroughs in finance, including large language models (LLMs), transformers, and their applications in decision-making and time series forecasting. If you're interested in cutting-edge financial technologies, this is a must-watch!

• 🥐 Asset Pricing Insights: In this edition, I introduce a paper that presents an advanced transformer-based model for forecasting stock movements. By leveraging long-range dependencies and integrating numerical inputs such as market sentiment, this model significantly outperforms traditional quantitative trading strategies, particularly in volatile and complex market environments.

Today’s Sponsor

We put your money to work

Betterment’s financial experts and automated investing technology are working behind the scenes to make your money hustle while you do whatever you want.

Learn more

AI-Finance Insights

“LLMs for Time Series: Breaking New Ground in Stock Market Prediction”

👉 A groundbreaking study challenges the conventional wisdom that Large Language Models (LLMs) can't effectively predict financial market returns. Using the Chronos model with 11 million parameters, researchers tested both zero-shot and fine-tuned predictions on major American stocks, focusing on residual returns after removing common factors.

👉 The study utilized three different approaches to extract residual returns: IPCA, PCA, and Fama-French factors. The most successful implementation achieved a remarkable Sharpe ratio of 3.17 for PCA over a 15-year period (2002-2016), translating to a t-statistic of 12.27. However, when accounting for trading costs (3 basis points per trade), the net Sharpe ratios turned negative.

👉 A key finding was that pre-trained Chronos models perform better with an exponential moving average (EMA) parameter of α=0.3, helping overcome biases from the model's trend-oriented training data. This configuration effectively captures market inefficiencies, though it doesn't outperform traditional Short-Term Reversal (STR) strategies.

👉 The research compared multiple approaches including CNN Transformers, autoARIMA, and STR strategies. Fine-tuned versions of Chronos showed promise, with optimal results achieved using 15 training steps. Notably, increasing to 40 steps led to decreased performance, suggesting a potential loss of pre-trained intelligence.

“Deep Learning for Options Trading: An End-to-End Approach”

👉 Researchers have developed a new deep learning approach to options trading that learns trading signals directly from market data, without requiring specific pricing models or market assumptions. This research offers an alternative to traditional methods that depend on complex market dynamics specifications.

👉 Testing on a decade of S&P 100 equity options data, their models showed improved performance compared to conventional strategies. The LSTM model achieved a Sharpe ratio of 1.329, notably higher than benchmark approaches. The study focused on delta-neutral straddle options with portfolio-level volatility targeting at 15% annually.

👉 The implementation includes turnover regularization to address transaction costs, with the LSTM model maintaining performance up to 50 basis points in costs. The research found that mean-reversion strategies generally performed better than momentum strategies, and simple linear models proved effective with proper regularization.

👉 During market stress periods, including the COVID-19 selloff, the models demonstrated consistent performance. The framework's design allows for potential application to other derivatives and instruments where sufficient market data exists.

👉 The study contributes to the growing field of quantitative trading by showing how machine learning can be applied to options trading without relying on traditional pricing models. The results suggest potential for improving options trading strategies through data-driven approaches.

Sponsor 👇

Man who predicted Crypto Market: AI will be 10X bigger

James Altucher, the investor who famously called Bitcoin's rise, is now predicting something even bigger: Artificial Intelligence.

Altucher, known for his bold predictions, believes AI could dwarf previous tech booms—becoming the first $100 trillion industry. He suggests that early investors have the potential to turn a modest $10,000 investment into $1 million over the next few years.

But this kind of opportunity won’t last long. Altucher says the “wealth window” will close soon.

He’s created a short video detailing exactly how investors can position themselves now for maximum profits. If you missed out on Bitcoin, this could be your second chance.

Plus, he’s giving away one of his top AI stock picks for free. It’s a rare chance to get ahead of what could be the biggest wealth-building opportunity of our time.

See how you can take advantage of AI here

AI-Essentials

In this edition, we're featuring "Recent Advancements of Financial Machine Learning with an Emphasis on Large Language Models." Speakers Nino Antulov-Fantulin and Petter Kolm explored how cutting-edge ML tools like large language models (LLMs) and transformers are transforming decision-making and forecasting in finance. Moderated by Dr. Hossein Kazemi, the talk focused on LLMs like GPT-3, their training, and applications in the financial sector. 👇

Asset Pricing Insights

“From attention to profit: quantitative trading strategy based on transformer”

👉 The paper introduces an advanced transformer-based model for predicting stock trends. By leveraging long-range dependencies and combining numerical inputs like market sentiment, this model outperforms traditional quantitative trading strategies, especially in volatile and complex markets.

👉 The authors tested the model on over 4,600 stocks from the Chinese market between 2010 and 2019, showing that the transformer-based strategy not only generates higher returns but also offers more stable results with lower turnover rates compared to over 100 traditional factor-based methods.

👉 Notably, the transformer model's ability to integrate stock price and turnover data, along with its enhanced architecture, provides more accurate predictions, yielding annual excess returns as high as 19.43%.

This paper uncovers a novel perspective on momentum trading by revealing how individual stock momentum is intertwined with factor momentum. The research presents compelling evidence that most factors exhibit positive autocorrelation, significantly influencing stock returns.

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! 🌟🤖💼