💰 quantitative finance

Predicting Stock Price Direction on Earnings Announcement Days using Multi-modal Deep Learning

This study demonstrates that a Transformer-based multi-modal deep learning model, which integrates firm fundamentals, technical indicators, and FinBERT-derived news sentiment, outperforms both logistic regression and LSTM baselines in predicting stock price direction on earnings announcement days by achieving superior sensitivity to volatile movements.

Original authors: Manuel Noseda, Nathan Soldati, Marco Paina

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Manuel Noseda, Nathan Soldati, Marco Paina

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to guess whether a stock's price will go Up, Down, or stay Neutral on the day a company announces its earnings report. This is a notoriously tricky game because the market is noisy, and prices can jump wildly in an instant.

The researchers from ETH Zurich treated this like a detective story. They built a "super-sleuth" system using Artificial Intelligence to solve the mystery. Here is how they did it, explained in everyday terms:

1. The Three Clues (The Data)

To make their prediction, the AI didn't just look at one thing. It gathered three different types of clues, like a detective combining fingerprints, witness statements, and motive:

The Company's Health Report (Fundamentals): They looked at 15 hard numbers, like how much profit the company makes or how much debt it owes. This is the "official" data.
The Recent Mood of the Market (Price Dynamics): They looked at the stock price over the last 30 days and used simple math to see if the price was speeding up or slowing down.
The News Whisper (Sentiment): They used a special AI reader called FinBERT to read thousands of financial news articles. This AI doesn't just read words; it understands the tone. Is the news happy, sad, or indifferent?

2. The Three Detectives (The Models)

The team trained three different "detectives" to solve the case and compared who was best:

The Old-School Detective (Logistic Regression): This is a simple, linear thinker. It looks at the clues and draws a straight line to a conclusion. It's fast but often misses the nuance.
The Careful Observer (LSTM): This is a deep learning model that reads the story day-by-day, like reading a book one page at a time. It is very cautious. If it's not 100% sure, it plays it safe.
The Big-Picture Visionary (Transformer): This is a more advanced model that can look at the whole 30-day story at once. It uses "attention" to figure out which specific days in the past were the most important, rather than just reading them in order.

3. The Tricky Part: The "Neutral" Trap

The biggest challenge was that most of the time, stocks don't do anything dramatic on earnings day; they stay Neutral.

Imagine a classroom where 67% of the students are sitting quietly, and only a few are jumping up and down.
If a detective just guesses "Everyone is sitting," they will be right most of the time, but they will miss all the exciting jumps.
The researchers had to teach their AI to care about the rare, wild jumps (Up or Down) just as much as the boring quiet days. They used a special scoring system that punished the AI heavily if it guessed the wrong direction (e.g., saying "Up" when it actually went "Down").

4. The Results: Who Won?

The paper found that there is no single "perfect" detective; it depends on what you value more: Safety or Sensitivity.

The Careful Observer (LSTM) was the "Safe Bet":
- It was very conservative. It mostly predicted "Neutral" to avoid making expensive mistakes.
- Result: It made the fewest costly errors. If you are risk-averse (scared of losing money), this is the model you want. It's like a driver who only turns the wheel when they are absolutely sure.
The Big-Picture Visionary (Transformer) was the "Sensitive Hunter":
- It was much better at spotting the wild price jumps. It caught more of the "Up" and "Down" movements.
- Result: It had the highest overall success score (Macro F1) for finding these movements. However, because it tried to catch everything, it made more mistakes about which direction the stock would go. It's like a driver who spots every pothole but might swerve too often.
The Old-School Detective (Logistic Regression) struggled:
- It was too simple to handle the complex mix of news, prices, and history. It made frequent, expensive directional errors.

5. The "Secret Sauce" (News Sentiment)

The researchers tested what happened if they removed the "News Whisper" (sentiment) from the clues.

The Verdict: Every detective got worse without the news.
Why? The news provided extra information that the numbers and price charts couldn't see on their own. It acted like a stabilizer, helping the AI make more reliable guesses.

The Bottom Line

The paper concludes that predicting stock movements on earnings day is a balancing act.

If you want to minimize risk and avoid big mistakes, the LSTM (the cautious model) is better.
If you want to catch volatile movements and are willing to accept more risk, the Transformer (the sensitive model) is better.

Crucially, the study proves that you cannot rely on just the numbers; you must include the "mood" of the news to get the best results.

Technical Summary: Predicting Stock Price Direction on Earnings Announcement Days using Multi-modal Deep Learning

Problem Formulation
This study addresses the challenge of predicting the directional movement of stock prices specifically on Earnings Announcement (EA) days. The authors frame this as an Event-Driven Price Movement Prediction task. Given a multi-modal temporal sequence representing the 30 days leading up to an EA, the objective is to classify the post-announcement price direction into one of three classes: UP ( $R \geq 3\%$ ), DOWN ( $R \leq -3\%$ ), or NEUTRAL (otherwise). The problem is characterized by significant market noise, high-impact price discontinuities, and a severe class imbalance where the NEUTRAL class accounts for approximately 67% of observations.

Methodology
The research employs a multi-modal deep learning approach, integrating three distinct data modalities into a unified feature space of dimension $d=21$ :

Firm Fundamentals ( $d_f=15$ ): Quarterly and annual metrics (e.g., Net Margin, ROE, Debt-to-Equity) serving as proxies for financial health.
Market Dynamics ( $d_m=3$ ): Adjusted closing prices and short-term momentum indicators (3-day and 6-day simple moving averages).
News Sentiment ( $d_n=3$ ): Sentiment scores derived from financial news articles processed via FinBERT, a language model fine-tuned for the financial domain. For days with multiple articles, sentiment vectors are averaged.

Data Processing and Modeling
The dataset covers the top 500 U.S. companies by market capitalization over a ten-month window. Preprocessing involves temporal imputation (forward-filling for non-trading days), cross-sectional mean imputation for sparse fundamentals, and z-score standardization. To address the class imbalance, the authors utilize a weighted cross-entropy loss function that penalizes directional errors (UP vs. DOWN) more heavily than neutral misclassifications.

Three model architectures are evaluated:

Logistic Regression (Baseline): A non-temporal multinomial classifier trained on flattened 30-day windows.
LSTM: A 2-layer stacked recurrent network (hidden dimension 64) designed to capture sequential dependencies.
Transformer: A self-attention-based encoder architecture (2 layers, 4 heads) utilizing sinusoidal positional encodings to capture global temporal patterns across the 30-day window without strict sequential inductive bias.

The models are trained and evaluated using metrics aligned with financial decision-making: Precision for actionable signals, Macro-averaged F1-score (to balance performance across imbalanced classes), and a custom cost metric that assigns higher penalties to UP↔DOWN misclassifications.

Key Results
The experimental results reveal a distinct trade-off between sensitivity and risk control:

LSTM Performance: The LSTM adopts a conservative strategy, frequently predicting the NEUTRAL class. This results in the lowest custom financial cost and high precision for directional signals but yields a lower Macro-F1 score due to missed volatile events.
Transformer Performance: The Transformer model demonstrates superior sensitivity, identifying more UP and DOWN events than the LSTM. Consequently, it achieves the highest Macro-F1 score (0.390) but incurs a higher custom cost due to increased directional errors.
Baseline Performance: The logistic regression baseline exhibits an unstable decision profile with a bias toward predicting UP movements, leading to the highest financial cost and frequent costly directional swaps.
Impact of Sentiment: An ablation study confirms that removing news sentiment features degrades performance across all models. The effect is most pronounced in the LSTM, suggesting that sentiment data plays a crucial stabilizing role in deep sequential models.

Significance and Claims
The paper claims to contribute to the relatively under-researched area of EA-specific stock prediction by demonstrating the value of combining "hard" information (fundamentals, prices) with "soft" information (news sentiment) in a multi-modal framework. The authors assert that while linear models are insufficient for capturing the complex temporal and cross-modal interactions in EA data, deep learning architectures offer distinct advantages depending on the investment strategy. Specifically, the study highlights that the choice of model depends on the target application: risk-averse strategies may favor the stability of the LSTM, while volatility-seeking strategies may benefit from the Transformer's higher sensitivity. The work underscores that incorporating news sentiment provides complementary predictive signals that improve stability and performance beyond what fundamentals and price dynamics alone can offer.