Introduction
Standard financial models, such as Geometric Brownian Motion or basic Lévy processes, typically characterize stock price "jumps"—sudden, significant movements—as stochastic, purely random events. While these models can account for the statistical "fat tails" observed in market returns, they often overlook the external, real-world catalysts that drive these anomalies. Social media sentiment, public panic, or hype are frequently the precursors to such volatility. This paper addresses this gap by introducing a modified Lévy jump-diffusion model that incorporates "market sentiment memory" to predict price jumps in an online learning framework.
Methodology
The proposed framework integrates three core components to transform sentiment signals into predictive insights:
- Modified Lévy Jump-Diffusion Model: Unlike traditional models where jump intensity is constant or random, this model assumes that jump magnitudes are governed by a "sentiment memory" process.
- Market Sentiment Memory: Utilizing the VaderSentiment tool, the researchers extract idiosyncratic (company-specific) and macroeconomic (market-wide) sentiment from daily Twitter data. Sentiment is modeled as an exponentially decaying process, reflecting the reality that public opinion accumulates and lingers over time.
- Unscented Kalman Filter (UKF): To handle the non-linear relationship between social media sentiment and stock returns, the authors employ a UKF. This allows for "online learning," where the model continuously updates its parameters as new data arrives.
The architecture converts a non-Markovian process (where the future depends on the entire history) into a higher-dimensional Markovian state-space representation suitable for the UKF optimization algorithm.
Data and Experimental Setup
The study utilized two primary data streams: social media sentiment from Twitter and historical financial data from Yahoo Finance. The analysis focused on three major technology stocks: Facebook (FB), Microsoft (MSFT), and Twitter (TWTR), with training and testing periods generally spanning from 2010/2013 to 2017.
Key Results and Discussion
The model's performance was evaluated based on its ability to predict "jumps," defined as price movements exceeding 1.96 standard deviations from the mean. The results demonstrated significant predictive power:
- Facebook (FB): 64.79% precision
- Twitter (TWTR): 60.85% precision
- Microsoft (MSFT): 52.96% precision
A qualitative analysis revealed that "outbreaks" in sentiment memory—sharp spikes in the sentiment memory parameter—served as strong leading indicators of abnormal price jumps. For instance, in the Facebook dataset, sentiment spikes accurately preceded jumps on specific trading days (e.g., Day 30 and Day 47). This suggests that markets are somewhat "inefficient" in how they digest public sentiment, creating a window of predictability that the UKF can exploit.
Conclusion and Future Work
This research highlights that sentiment is not a series of isolated events but a decaying process with memory. While the model achieved high precision in jump identification, some limitations remain. The UKF's return predictions often exhibited higher magnitudes than actual returns due to the volatility of raw sentiment data. Additionally, the model's performance during low-sentiment periods tends to default to previous returns. Future research could explore refining the "memory" component by multiplying historical jumps by decaying factors to improve magnitude accuracy and directionality.
Summary updated