Consider what it means for a sector’s news coverage to be volatile. The obvious interpretation is that sentiment scores are moving around a great deal. But there is a subtler and ultimately more valuable form of volatility: the kind where the fundamental character of the conversation changes, even before the numbers reflect it. This is what the Semantic Volatility Index, or SVI, is designed to detect.
The SVI is a composite indicator developed exclusively for Canary 2.0. It aggregates five distinct sub-component signals, each measuring a different dimension of narrative instability, into a single number on a scale of zero to one hundred. The higher the score, the more the underlying language and framing of AI sector coverage is in flux.
Why Sentiment Alone Is Not Enough
Imagine that AI sector news has been consistently bearish for two weeks: earnings disappointments, cautious guidance, some analyst downgrades. The aggregate sentiment score is negative. Now imagine that today’s news carries the same negative score, but the stories driving it are entirely different. Not earnings concerns, but export controls, antitrust investigations, and supply chain disruption. The number is the same. The meaning is not.
This is what practitioners refer to as a narrative regime shift: a change in the type of risk present in the market, not just the level of it. Regime shifts of this kind have historically preceded significant repricing events, precisely because the vocabulary of earnings disappointment is familiar and largely priced in, whereas the vocabulary of geopolitical disruption opens exposure to tail events that markets have not fully discounted.
The SVI is designed to detect the early stages of these shifts. It does so by looking not at where sentiment has arrived but at how the language of the conversation is evolving, and by measuring that evolution across five complementary dimensions.
The Five Sub-Components
Each sub-component of the SVI captures a different aspect of narrative stability. Together they cover a range from lagging confirmation signals to genuine leading indicators that can fire several days ahead of any visible movement in sentiment scores.
The rolling standard deviation of daily sentiment scores across the most recent seven trading days. A high reading confirms that sentiment has been fluctuating significantly. Useful context, but by definition it registers only after the fact.
The spread of individual article sentiment scores within a single day. A high reading means today’s articles are pulling in different directions — some strongly positive, others strongly negative — even if the average looks moderate. This reveals internal tension that aggregate scores conceal. Because sentiment dispersion is a lower bound on semantic divergence, a high SC2 reading is almost certainly understating the true degree of narrative disagreement present that day.
The day-over-day change in weighted sentiment compared to the most recent prior trading day. Large absolute movements, particularly those concentrated in specific categories, signal that something has changed in the tone of coverage since yesterday. The category breakdown accompanying this sub-component is often more informative than the overall figure, since opposing movements across categories can cancel each other out in the aggregate.
A statistical measure of how different the vocabulary in this week’s articles is from last week’s, computed using Jensen-Shannon Divergence (JSD), a rigorous information-theory metric for comparing two language distributions. New words entering the discourse signal narrative change before it registers in sentiment scores. SC4 operates on the body text of articles, giving it access to the full richness of each piece rather than the compressed language of headlines alone.
Measures how the balance of the eight narrative frames has shifted between the current seven-day period and the prior seven-day period, again using Jensen-Shannon Divergence. This is the most semantically rich of the five sub-components. It captures not just changes in vocabulary but changes in the fundamental type of story being told about the sector. A rotation from Growth Momentum and Technical Breakthrough toward Regulatory Risk and Market Correction is precisely the kind of structural change the SVI is designed to surface.
Jensen-Shannon Divergence Explained
Both SC4 and SC5 use Jensen-Shannon Divergence as their core mathematical measure. JSD quantifies how different two probability distributions are. In SC4, it compares the distribution of words used in this week’s AI sector articles against those used last week. In SC5, it compares the distribution of narrative frames across the same two periods.
A JSD value of zero means the two distributions are identical: the language and framing have not changed. As the value rises, the divergence between periods grows, indicating that the vocabulary or framing of the conversation has shifted. JSD has two properties that make it well suited to this task: it is bounded between zero and one, making it directly interpretable, and it is symmetric, meaning the comparison produces the same result regardless of which period is treated as the reference.
The Composite Score
Each sub-component produces a raw value on its own natural scale. These are normalised to a common range and then combined into the composite SVI score using a weighted average, with the leading indicators carrying greater weight than the lagging and concurrent ones. The weights reflect a deliberate prioritisation: signals that fire ahead of events are more valuable than those that confirm what has already happened.
The P-Wave Analogy
In seismology, an earthquake generates two types of waves. P-waves, which are compressional waves, travel fast and arrive first. S-waves, the shear waves that cause destructive ground movement, arrive second, sometimes minutes later. Detecting the P-wave provides a window of warning before the damage arrives.
Vocabulary drift and frame distribution shift are the P-waves of narrative change. They travel ahead of the sentiment shear wave. The SVI is the instrument that detects them.
When the SVI is elevated, the conversation about AI companies has already begun to change. Whether or not the aggregate sentiment score has moved yet, the underlying linguistic signals are telling a different story. In the data gathered since Canary 2.0 launched, those signals have consistently preceded visible sentiment movements by several days.
Independent Research Validation
The SVI’s design draws on a growing body of academic research into the relationship between textual signals and market behaviour.
Multiple independent studies have confirmed that categorising financial news by the type of story being told, rather than simply scoring its sentiment, is a better predictor of market volatility than sentiment analysis alone. Research examining the vocabulary of financial news across different macroeconomic regimes has found that Jensen-Shannon Divergence between successive periods correlates strongly with periods of elevated prediction error in financial models. In other words, when the language is shifting, standard sentiment models become less reliable. This is precisely the condition the SVI is designed to flag.
Research into how markets respond to communication has also demonstrated that markets calibrate to stability. When a sustained period of consistent messaging gives way to a significant change, the market response tends to be larger than it would have been without the preceding period of calm. A sustained run of CONTINUING labels in the frame distribution establishes a narrative equilibrium; if that equilibrium breaks, the signal may carry more weight than a single isolated reading would suggest.
The SVI’s sub-component architecture also reflects the established finding that news predicts the direction of market volatility more reliably than it predicts the direction of asset prices. Canary is accordingly designed as a volatility and regime detector rather than a price forecasting tool.
What the SVI Does Not Do
The SVI measures narrative volatility. It does not predict share price movements, and it makes no claim to forecast the direction in which sentiment will move next. A high SVI score indicates that the story is changing; it does not say whether the change will prove positive or negative for the sector.
This distinction is important. The value of the SVI lies in the early warning it provides: the opportunity to examine the underlying signals closely, review the specific vocabulary and framing shifts driving the reading, and make an informed judgement about what they might mean. It is an instrument for directing attention, not a substitute for analysis.
Regime Shift versus Normal Volatility
Not all elevated SVI readings indicate a regime shift. Day-to-day variation is normal, and the SVI will reflect it. What distinguishes a genuine regime shift is persistence: multiple sub-components elevated simultaneously, over multiple consecutive days, with CONTINUING frame momentum in the distribution chart.
A single day’s high reading is a prompt for attention. Several days of consistent elevation is a prompt for action.
The SVI in Practice
A sustained High or Extreme SVI reading, particularly one accompanied by CONTINUING labels on risk frames in the distribution chart, has in practice preceded visible directional moves in category sentiment by several days. The most useful posture when the SVI is elevated is not to act on the number directly, but to look closely at what is driving it: which sub-components are elevated, which categories are driving within-day dispersion, and which frames are showing momentum. The SVI is the signal that says something is happening; the rest of the dashboard tells you what.
The Semantic Volatility Index is a proprietary component of Canary 2.0, developed and maintained by KnowEntry. The sub-component design, weighting methodology, and normalisation thresholds are calibrated continuously as the dataset matures. The system has been in live production since early 2026 and is updated daily alongside the main Canary sentiment output.
One of the practical disciplines built into the SVI from the outset is empirical calibration. The thresholds and weights that determine how each sub-component contributes to the composite score are not set once and fixed permanently. They are reviewed regularly against accumulated data, and adjusted when the evidence supports a change. A system calibrated on its first week of data would not serve its purpose well in month six; the SVI is designed to improve as its understanding of the AI sector’s normal behaviour deepens.