At Context Analytics, we are constantly innovating within our S-Factor Sentiment Feed—testing new metrics, uncovering relationships, and engineering features that push the boundaries of quantitative trading with alternative data. Our mission is to transform unstructured text into structured, actionable signals that give traders and portfolio managers a measurable edge.
Background: The Foundation of the S-Factor Feed
Our S-Factor Feed includes multiple sentiment-derived metrics from Twitter (X) that quantify market chatter around each security over time. Among the key components are:
- Raw-S – The aggregated raw sentiment from all tweets mentioning a given security within a 24-hour window.
- S-Volume – The total number of tweets captured for that security during the period.
- S-Dispersion – The percentage of unique accounts contributing to that S-Volume, indicating how broad the conversation is across market participants.
These foundational variables capture different dimensions of social sentiment — magnitude, activity, and diversity. However, until now, they have not been directly combined into a single normalized metric that captures the average sentiment intensity per participant.
Defining the Sentiment per Account Metric
To address this, we engineered a new feature derived from our existing S-Factor components — the Sentiment per Account metric. It is defined as:
This formulation normalizes aggregated sentiment by both the number of contributing accounts and their distribution. Conceptually, it reflects how positive or negative the average account is tweeting about a given security, rather than being dominated by high-volume or highly vocal users.
While S-Score and Raw-S capture overall tone and magnitude, the Sentiment per Account metric provides a refined view that can be added alongside—especially useful in identifying situations where a single influential account could skew aggregate sentiment. By adjusting for dispersion and tweet volume, this metric better isolates the breadth-adjusted sentiment strength of a crowd.
Research Design: Testing Predictive Efficacy
To evaluate this new signal, we conducted a historical backtest spanning from January 2019 through the present, using the following design:
- Universe: U.S. equities with price > $5.
- Filter: Securities with at least three unique accounts tweeting within the 24-hour window to ensure sufficient crowd participation.
- Portfolio Construction: Securities were ranked daily by their Sentiment per Account value and sorted into five equal-weighted quintiles (Q1–Q5).
- Rebalancing Frequency: Daily.
- Benchmark: SPY (S&P 500 ETF).
Results: Clear Monotonic Spread and Strong Outperformance
The results show a clear monotonic relationship between Sentiment per Account and subsequent returns. Across the sample period, the average daily coverage included approximately 160 securities per quintile, ensuring robust breadth across time.
- Q1 (lowest Sentiment per Account) securities consistently underperformed all other groups.
- Q5 (highest Sentiment per Account) significantly outperformed the lower quintiles and the SPY benchmark.
- The Q5 portfolio delivered roughly +60% cumulative outperformance versus SPY and an additional +4% annualized alpha, while maintaining similar Sharpe and Sortino ratios.
- The Q5–Q1 long-short spread demonstrated improved risk-adjusted returns compared to SPY, reinforcing that sentiment intensity per unique participant carries independent predictive power.
It’s also important to note that all quintiles were equally weighted, while SPY is heavily weighted toward a few securities—making the observed spread even more compelling, as it is driven by broad participation rather than concentration in large-cap names.
Interpretation: Why It Matters
The findings suggest that Sentiment per Account can serve as a valuable addition to trading models. While total sentiment (Raw-S or S-Score) measures how much the market is talking, the new metric captures how strongly each participant feels.
This distinction is critical in social-driven alpha research: markets often respond not only to aggregate chatter but also to the intensity and uniformity of conviction among participants.
In essence, overall sentiment is important, but who is talking—and how consistently positive or negative they are—can also be predictive of next-day returns.
Looking Ahead: The Power of Feature Engineering with S-Factor
This new feature is just one example of what’s possible within the Context Analytics framework. Our S-Factor Feed provides the granularity and flexibility for quantitative teams to engineer proprietary sentiment-driven signals. From dispersion-adjusted metrics like Sentiment per Account to cross-sectional combinations with volatility, liquidity, or fundamental data—there’s vast potential for innovation.
We’ll continue publishing our findings as we explore new relationships between sentiment, market behavior, and performance.
For more information on the S-Factor Feed, sentiment signal research, or integration options, visit www.contextanalytics-ai.com or contact us to discuss data trials and model applications.
TL;DR:
Context Analytics introduced a new “Sentiment per Account” metric within its S-Factor Feed, normalizing sentiment by the number and diversity of unique accounts. Backtests (2019–present) show that securities with higher Sentiment per Account outperform by ~60% vs SPY, confirming its predictive power. This feature refines crowd sentiment signals by emphasizing how strongly each participant feels, not just how much they talk.