💰 quantitative finance

Overreaction as an indicator for momentum in algorithmic trading: A Case of AAPL stocks

This paper demonstrates that machine learning models integrating high-frequency Twitter sentiment and volatility data can effectively predict and monetize short-term overreactions in Apple (AAPL) stock, revealing that negative emotions and volatility drive intraday momentum patterns that outperform traditional rules at ultra-short horizons.

Original authors: Szymon Lis, Robert Ślepaczuk, Paweł Sakowski

Published 2026-02-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Szymon Lis, Robert Ślepaczuk, Paweł Sakowski

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 the stock market as a giant, chaotic playground where millions of people (and robots) are constantly shouting, whispering, and reacting to news. Sometimes, the crowd gets so excited or so scared that they push prices up or down way too far, way too fast. This is called an overreaction.

For decades, traders have tried to catch these moments to make money. But it's like trying to catch a speeding bullet with your bare hands.

This paper is about a team of researchers from the University of Warsaw who decided to build a super-smart robot detective to catch these overreactions in Apple (AAPL) stock. They wanted to see if they could predict when the crowd is about to panic or get overly excited, using two main tools: high-speed market data and Twitter emotions.

Here is the breakdown of their adventure, explained simply:

1. The Setup: The "Emotion Radar"

The researchers didn't just look at stock charts. They built a radar that listens to the "emotional heartbeat" of the internet.

The Data: They watched Apple's stock price tick-by-tick (every 1, 5, 10, or 15 minutes).
The Mood Ring: Simultaneously, they scanned thousands of tweets about Apple. They didn't just look for "good" or "bad" words; they used advanced AI (like a super-smart translator) to detect specific emotions: Fear, Sadness, Joy, Anger, Surprise, and even Disgust.

The Analogy: Imagine you are a trader standing in a crowded stadium. You can see the scoreboard (the stock price), but you can also hear the crowd screaming. Sometimes the crowd screams because they are terrified (Fear), sometimes because they are shocked (Surprise). The researchers wanted to know: Does the crowd screaming "Fear" mean the price is about to crash, or does it mean the price has already crashed and is about to bounce back?

2. The Problem: Too Much Noise

The market is noisy. At the 1-minute level, it's like trying to hear a conversation in a hurricane. The price jumps up and down so fast due to tiny technical glitches (microstructure noise) that it's hard to tell if it's a real emotional reaction or just a glitch.

The Finding: At 1-minute speeds, the "robots" (Machine Learning models) struggled. The noise was too loud. The only thing that mattered was how wild the price was swinging (volatility), not really what people were saying on Twitter.

3. The Sweet Spot: The 10-Minute Window

As they slowed down their view to 5, 10, and 15 minutes, the picture got clearer.

The 10-Minute Magic: This was the "Goldilocks" zone. It wasn't too fast (noisy) and not too slow (old news).
The Result: Here, the researchers found that classic human behavior was the strongest predictor. When people got scared or surprised, they overreacted, and the price kept moving in that direction for a while.
The Twist: Interestingly, at this specific 10-minute mark, a simple rule ("If people are scared, sell") actually worked almost as well as the super-complex AI robots. The crowd's behavior was so predictable that you didn't need a PhD in math to spot it.

4. The AI vs. The Human Intuition

The researchers tested four different types of "AI detectives":

XGBoost & Random Forest: Think of these as super-organized librarians. They look at thousands of rules and patterns to find the answer.
Deep Neural Networks (DNN): Think of this as a genius art student who sees patterns humans can't.
BiLSTM: Think of this as a storyteller that remembers the sequence of events (what happened 5 minutes ago matters for what happens now).

The Verdict:

At ultra-fast speeds (1 min), the AI was better than simple rules, but the profit was tiny because the market was too chaotic.
At medium speeds (5-10 mins), the AI and the simple rules both made money. The AI was slightly better at spotting the exact moment to jump in, but the simple "overreaction" rules were surprisingly strong.
Key Insight: The AI didn't just guess; it learned that Fear and Sadness on Twitter were the biggest triggers for price drops, while Joy and Surprise often signaled a price bounce.

5. The "Why" (The Secret Sauce)

The researchers used a special tool called SHAP (which is like an X-ray for the AI) to see why the robots made their decisions.

Volatility: The most important clue was how fast the price was moving.
Fear: When the "Fear" meter on Twitter spiked, the AI knew a crash was likely or that a drop had already happened.
Surprise: When people were shocked, it often meant a new opportunity was opening up.

The Big Takeaway

This paper tells us that emotions drive the market, and we can actually predict them.

For the 1-minute trader: It's a noisy mess; focus on the speed of the price, not the tweets.
For the 10-minute trader: This is where the magic happens. The crowd gets emotional, overreacts, and then corrects itself. You can make money by catching this wave.
The Future: While simple rules work well, combining AI with Emotion Data gives you a sharper edge. It's like having a weather forecast that doesn't just tell you it's raining, but tells you why it's raining and how long the storm will last.

In a nutshell: The market is a giant emotional rollercoaster. By listening to the screams of the crowd on Twitter and using smart computers to filter out the noise, we can figure out when the ride is about to go up or down, turning human panic and excitement into a profitable strategy.

1. Problem Statement

The paper addresses the challenge of predicting and monetizing short-term market overreactions (extreme price deviations relative to fundamentals) in high-frequency trading. While momentum is a well-documented anomaly, traditional linear models and rule-based strategies often fail to capture the complex, non-linear interactions between investor sentiment, volatility, and market microstructure at intraday frequencies.

The authors investigate whether emotion-driven overreactions contain a predictable structure that can be systematically exploited using:

High-frequency emotional data extracted from Twitter (X).
Non-linear machine learning (ML) models capable of capturing regime dependencies and threshold effects.
Volatility-normalized return signals to distinguish genuine overreactions from noise.

The study focuses specifically on Apple Inc. (AAPL) to test if these mechanisms hold for a liquid, high-volume asset.

2. Methodology

A. Data Construction

Asset & Period: Apple Inc. (AAPL) from January 1, 2019, to June 30, 2022.
Market Data: Intraday OHLCV data sampled at four frequencies: 1, 5, 10, and 15 minutes.
Sentiment Data: Twitter messages containing "$AAPL" (English only).
- Feature Extraction: Emotion probabilities were extracted using ModernFinBERT (a transformer-based model), generating vectors for emotions: anger, disgust, fear, joy, neutral, sadness, surprise.
Preprocessing:
- Timestamps aligned to US/Eastern trading hours (04:00–20:00).
- Anti-Look-Ahead: Features ( $X_t$ ) are constructed using data available up to interval $t$ ; the target is the overreaction state of interval $t+1$ .
- Scaling: Z-score normalization applied using only training statistics.

B. Target Definition: Volatility-Scaled Overreaction

The authors define overreaction ( $OR_{t+1}$ ) as a three-class classification problem based on volatility-normalized returns:
$OR_{t+1} = \begin{cases} +1 & \text{if } r_{t+1} > \theta\sigma_t + 2TC \\ -1 & \text{if } r_{t+1} < -(\theta\sigma_t + 2TC) \\ 0 & \text{otherwise} \end{cases}$

$r_{t+1}$ : Log-return.
$\sigma_t$ : Rolling realized volatility (20-interval window).
$\theta$ : Threshold parameter (tested 1.5 to 5.0) controlling event rarity.
$TC$: Transaction costs (set to 10 bps one-way; 20 bps round-trip).

C. Machine Learning Models

Four non-linear classifiers were trained to predict $OR_{t+1}$ :

XGBoost: Gradient-boosted trees with multi-class softprob objective.
Random Forest: Ensemble of decision trees.
Deep Neural Network (DNN): Feed-forward network (64-32 hidden layers, ReLU, dropout).
Bidirectional LSTM (BiLSTM): Recurrent network to capture temporal dependencies in sentiment sequences.

Training Protocol: Chronological split (60% Train, 20% Val, 20% Test) with expanding-window time-series cross-validation.
Class Imbalance: Inverse-frequency class weights applied to handle the dominance of the "neutral" (0) class.
Execution: Trades executed at the open of $t+1$ if $P(OR_{t+1}=k | X_t) > c$ (threshold $c$ optimized on training data).

D. Evaluation Metrics

Performance: Annualized Sharpe Ratio, Sortino Ratio, Maximum Drawdown, and Annualized Returns.
Statistical Significance: Jobson–Korkie test with Memmel correction to compare Sharpe ratios between ML models and benchmarks.
Interpretability: SHAP (Shapley Additive exPlanations) values to decompose feature contributions and validate behavioral hypotheses.

3. Key Contributions

Novel Integration: First study to combine transformer-based emotion embeddings (ModernFinBERT) with volatility-normalized overreaction definitions in a high-frequency ML framework.
Frequency-Specific Insights: Systematic analysis across 1, 5, 10, and 15-minute horizons, revealing that the efficacy of ML vs. behavioral rules is highly dependent on the sampling frequency.
Behavioral Validation: Uses SHAP to empirically demonstrate that fear, sadness, and volatility are the primary drivers of predicted overreactions, providing a mechanistic link between NLP features and price dynamics.
Economic Viability: Demonstrates that ML strategies can generate statistically significant risk-adjusted returns even after accounting for realistic transaction costs and slippage.

4. Key Results

A. Predictive Performance by Frequency

1-Minute (Ultra-High Frequency):
- Noise Dominance: Microstructure noise overwhelms sentiment signals.
- Winner: Random Forest (Sharpe 0.90) significantly outperformed the overreaction benchmark (Sharpe -1.01), but trade volume was extremely low (only 2 trades), indicating rare opportunities.
- Insight: Sentiment is too slow-moving to be predictive at this granularity; volatility is the sole dominant driver.
5-Minute:
- ML Advantage: BiLSTM (Sharpe 0.69) significantly outperformed the benchmark (Sharpe -0.94).
- Insight: Recurrent architectures successfully captured temporal dependencies in emotional dynamics that simple rules missed.
10-Minute (The "Sweet Spot"):
- Benchmark Dominance: The Classical Overreaction Rule achieved the highest performance (Sharpe 1.43, 34% annual return), outperforming all ML models (DNN Sharpe 0.65).
- Insight: At this horizon, classical behavioral momentum (mean reversion/continuation) is the strongest signal. ML models failed to add value over simple heuristics here.
15-Minute:
- Convergence: XGBoost (Sharpe 0.87) and the Overreaction benchmark (Sharpe 0.66) performed comparably.
- Insight: As the horizon lengthens, the signal-to-noise ratio improves, but the distinct advantage of complex ML diminishes as behavioral patterns become more linear and predictable.

B. Statistical Significance (Jobson–Korkie Test)

1 & 5 Minutes: ML models significantly outperformed benchmarks ( $p < 0.05$ ).
10 Minutes: The Overreaction benchmark significantly outperformed ML models (statistically insignificant difference in favor of ML, but the benchmark had higher absolute Sharpe).
15 Minutes: No significant difference between ML and benchmarks.

C. SHAP Interpretability

Volatility: The single most important feature across all frequencies. High volatility consistently pushed models toward sell signals (interpreting volatility spikes as liquidity gaps or order-book imbalances).
Fear & Sadness: Strong negative drivers. High levels of these emotions increased the probability of overreaction (sell) predictions, aligning with loss aversion theory.
Joy & Surprise: Positive drivers for buy signals, particularly at 10- and 15-minute horizons.
Neutral Sentiment: Surprisingly influential; high neutrality often correlated with buy signals, suggesting that a lack of emotional noise indicates market stability.

5. Significance and Implications

For Algorithmic Trading: The study proves that non-linear ML models are essential for ultra-short horizons (1-5 mins) where noise and complex interactions dominate. However, for intermediate horizons (10 mins), simpler behavioral rules may be sufficient and more robust.
For Behavioral Finance: It provides empirical evidence that emotions (specifically fear) are not just noise but contain predictive structure regarding price dislocations. The "negativity bias" is quantifiable and exploitable.
For Market Efficiency: The findings challenge the strong form of market efficiency at intraday scales, showing that sentiment-driven overreactions create predictable, exploitable anomalies that persist for specific time windows.
Limitations: The study is limited to a single asset (AAPL) and Twitter data. It does not explicitly model order-book dynamics or latency, which are critical for real-world HFT execution.

Conclusion

The paper concludes that emotion-driven overreactions are predictable, but the optimal method depends on the time horizon. Machine Learning excels at ultra-short frequencies by filtering noise and modeling non-linear sentiment-volatility interactions, while classical behavioral momentum rules dominate at the 10-minute mark. The integration of NLP-derived emotion features with volatility-normalized signals offers a robust framework for next-generation algorithmic trading strategies.