Does sentiment help in asset pricing? A novel approach using large language models.

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:

  • Does sentiment help in asset pricing? A novel approach using large language models and market-based labels

  • A Bayesian-based classification framework for financial time series trend prediction

• 🥐 Asset Pricing Insights: In this edition, I introduce a paper examining how momentum trading strategies handle market turning points. The research analyzes the effectiveness of combining slow and fast momentum signals, revealing key insights about trend reversals and portfolio optimization across global equity markets.

Today’s Sponsor

Handle your phone calls 24/7 with AI

Boost your business with Synthflow AI Phone Calls—available 24/7 to handle thousands of calls concurrently. Scale effortlessly as your needs grow and speed up lead response times. The assistants can book appointments into your calendar, transfer the calls to a human rep, and extract valuable info from each call. Integrates with your tech stack, get started from $29/month.

Let AI answer the phone

AI-Finance Insights

“Does sentiment help in asset pricing? A novel approach using large language models and market-based labels”

👉 Researchers have developed a novel approach to sentiment analysis in financial markets using a state-of-the-art large language model (SMARTyBERT) combined with market data-driven labeling. This innovative method addresses the critical challenge of data labeling in NLP by deriving labels from next-day excess returns relative to the Fama-French five-factor model, eliminating reliance on subjective human annotations.

👉 The study utilizes a comprehensive dataset spanning 2012-2019, incorporating diverse text sources for all CRSP universe companies: earnings call transcripts, Bloomberg headlines, and StockTwits tweets. This multi-source approach captures sentiment from various investor perspectives, enhancing the model's understanding of market sentiment.

👉 Testing showed impressive results, with a long/short equal-weighted portfolio strategy yielding a 35.56% annualized return and a Sharpe ratio of 2.21 before transaction costs (33.21% and 2.06 after costs). The strategy significantly outperformed FinBERT, a traditional human-annotated model, which showed negative returns, demonstrating the superiority of the data-driven labeling approach.

👉 The research found that low sentiment consistently predicts negative next-day excess returns, while high sentiment shows no similar effect. The high-low sentiment (HLS) factor proved significant at the 1% level in multifactor asset pricing, even when controlling for Fama-French five factors plus momentum.

Key Paper Metrics 🔻

• ➡️ Annualized Return: 35.56% (pre-costs), 33.21% (post-costs)

• ➡️Sharpe Ratio: 2.21 (pre-costs), 2.06 (post-costs)

• ➡️Effect: Strongest in small-cap stocks

• ➡️Data Period: 2012-2019

• ➡️Coverage: Full CRSP universe

“A Bayesian-based classification framework for financial time series trend prediction”

👉 Researchers have developed an innovative tri-state labeling algorithm combined with machine learning and deep learning models to predict financial market trends. This novel approach classifies price movements into three states (up, down, and no-action), with the no-action state helping filter out noisy market periods and reduce false signals.

👉 The study implements multiple classification models including Support Vector Machines (SVM), XGBoost, Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU). These models are optimized using Bayesian hyperparameter optimization to maximize performance. The framework uses combinatorial purged K-fold cross-validation to prevent data leakage and look-ahead bias.

👉 Testing on selected S&P 500 stocks showed impressive results, particularly with the XGBoost model achieving an annualized Sharpe ratio of 2.823, significantly outperforming recent sophisticated approaches like CNN-based and transformer models. The framework maintained consistently low maximum drawdown values, typically under 10%.

👉 The research demonstrated strong performance across stocks with varying levels of systematic risk (beta), from low-risk stocks like Clorox (CLX, β=0.17) to high-risk stocks like Macy's (M, β=2.09), showing the framework's robustness across different market conditions.

Key Metrics:

➡️ Best Annualized Sharpe Ratio: 2.823 (XGBoost)

➡️ Maximum Drawdown: Generally < 10%

➡️ Model Performance (Average):

• XGBoost: 2.823 SR

• SVM: 2.007 SR

• LSTM: 1.672 SR

• GRU: 1.675 SR

➡️ Test Period: Jan 2020 - Nov 2021

➡️ Coverage: Selected S&P 500 stocks

Sponsor 👇

Join the Revolution in Private Market Investing

At UpMarket, we make it possible for accredited investors to tap into high-growth private companies. With companies like xAI and Anthropic redefining what’s possible in AI, and SpaceX pushing the boundaries of space exploration, UpMarket’s platform has become a gateway for future-focused accredited investors. Join our network of 500+ investors who have placed over $175M in private opportunities since 2019.

Join now

Asset Pricing Insights

“Momentum Turning Points”

📢 Slow and Fast Time-Series Momentum: Characterizing Stock Market Cycles.📈 Keep reading!🔻

👉 Researchers have developed a novel approach to characterize stock market cycles using slow and fast time-series momentum strategies. This innovative method addresses the challenge of balancing noise reduction and quick reaction to market turning points, providing insights into market behavior and potential investment strategies. .

👉 The study utilizes a comprehensive dataset spanning a 50-year period from 1969 to 2018, incorporating U.S. stock market excess returns. The researchers define four market cycles based on the combination of slow (12-month) and fast (1-month) momentum signals: Bull, Bear, Correction, and Rebound states.

👉 Analysis showed impressive results, with distinct return characteristics for each market cycle. Bull states exhibited high average returns (9.5% annualized) with low volatility, while Bear states showed negative returns (-7.7% annualized) with the highest volatility. Correction and Rebound states displayed intermediate characteristics, potentially indicating market turning points.

👉 The research found that combining slow and fast momentum signals can effectively identify states with negative expected returns (Bear states), which is difficult to explain using traditional risk-based asset pricing models. The study also revealed close connections between market cycles and macroeconomic indicators, with Bear states occurring frequently in early recession periods.

Key Metrics:

• ➡️ Bull State: 9.5% annualized return, 11.3% volatility

• ➡️ Bear State: -7.7% annualized return, 20.8% volatility

• ➡️ Correction State: 6.5% annualized return, 17.8% volatility

• ➡️ Rebound State: 9.6% annualized return, 17.3% volatility

• ➡️ Data Period: 1969-2018

• ➡️ Coverage: U.S. stock market

👉 The researchers also analyzed intermediate-speed momentum strategies, which blend slow and fast signals. These strategies demonstrated higher Sharpe ratios than pure slow or fast strategies. A novel decomposition of alpha revealed that market timing drives about two-thirds of the alpha, while volatility timing accounts for the remaining one-third.

👉 This study provides valuable insights into market behavior and offers potential improvements to momentum-based investment strategies. The findings challenge traditional risk-based explanations for market returns and suggest that behavioral factors may play a significant role in market dynamics.

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! 🌟🤖💼