Part I: Market Segmentation
Before accusing any individual stock of driving a crisis, we must first identify when the market itself entered abnormal regimes.
We therefore begin with the only element the market cannot conceal:
the timeline.
In the figure above, the timeline was segmented with respect to periods that exhibit similar weekly returns using dynamic programming (50 segments). Notably, the most severe segments coincide with well-known systemic crises, including the dot-com bubble burst, the 2008 financial crisis, and the COVID-19 shock.
These periods serve as our outbreak windows in the following analysis. We will focus on these segments to study how financial contagion emerges, propagates, and eventually dissipates.
Major crisis periods identified by market segmentation.
| Crisis episode | Start date | End date | Duration (weeks) | Cumulative market return | Interpretation |
|---|---|---|---|---|---|
| Dot-com bubble burst | 2000-09-08 | 2000-10-20 | 6 | −19.49% | Huge correction after the internet hype |
| Subprime financial crisis | 2008-10-03 | 2008-10-24 | 3 | −22.02% | Banking and credit collapse |
| COVID-19 market shock | 2020-02-21 | 2020-04-10 | 7 | −22.90% | Global shutdown |
All identified crisis periods exhibit cumulative losses exceeding 19%, confirming their systemic severity and justifying their use as outbreak windows in the subsequent contagion analysis.
Part II: Definition of contagion dynamics
The Main Immunology Pipeline
Before analysis, it is important to be clear on how the algorithm functions:
Three key dimensions must be introduced first. These dimensions are jointly used to evaluate the state and systemic relevance of each market entity.
- Network centrality: We measure how embedded each asset is within the market network. Centrality reflects how strongly a node is connected to others through correlation, mutual information, and causal exposure.
- Temporal leadership: Timing matters. Assets that exhibit abnormal stress early in the outbreak window carry more explanatory power than late movers. Early infection combined with strong connectivity raises suspicion regarding a possible triggering role (i.e. patient zero).
- Reproduction number (R₀): Inspired by epidemiological concepts, this metric quantifies the effective contagion pressure exerted by a distressed node on the rest of the network. Higher values indicate nodes with a greater capacity to transmit and amplify market stress through interconnected pathways.
Together, these three dimensions define the Pandemic Potential Index (PPI), a composite score that captures an entity’s ability to initiate, amplify or propagate financial stress within the market. This helps us find super-spreaders in the network.
For each case, we used 50 representative tickers for the sake of visualization. All the tickers are accessible through the Ticker Dictionary on the right of the page, providing insightful information about every single one of them.
Infection state classification
In parallel, each entity is assigned a health state based on its return dynamics over the outbreak window.
- An entity is labeled directly sick if its daily return crosses the −5% threshold. It stays sick until recovery condition is satisfied.
- Entities that maintain strong connections with sick nodes (through correlation or causal exposure) are classified as sick by contagion.
- Over time, entities may recover if their daily returns are positive for three consecutive days.
- Those that never meet the infection criteria remain healthy.
In each case file, we apply the same investigative procedure: we examine market behavior during the outbreak window, reconstruct the underlying interaction network, and identify key roles such as patient zero, super-spreaders, and high-risk entities.
Select a case file
You may now select one of the following crisis episodes to open its corresponding case file. Each case file presents a focused and comparable analysis of market behavior during that outbreak.
Case File: Dot-com Bubble
At the turn of the millennium, markets were drunk on promises. Anything with a website and a “.com” suffix was treated as the future of civilization. Valuations detached from fundamentals, and skepticism quietly exited the room. Then reality pushed back when the internet bubble burst.
The plot below shows the mean daily market return during the identified outbreak window.
During this period, the market records negative daily returns on nearly 77% of trading days, cumulating to a −19.49% loss over the entire window.
To grasp what's really happening here we shall analyze the internal structure of the market and track
how stress appears, spreads, and concentrates at the entity level.
You can press the RED BUTTON below to initiate the analysis.
Patient zero has been identified as entity BIIB.
Now that a patient zero has been identified, we might want to understand how the outbreak unfolded. It is interesting to understand the relative roles played by all assets during the crisis.
The 3D plot below places every stock in the market inside a common contagion space. Each point represents an entity, positioned according to three dimensions: reproduction index, centrality and temporal leadership (all normalized). These values are the mean values over the entire timeframe.
To move from this static snapshot to a dynamic view, we now follow how the outbreak unfolds over time. The timeline below tracks the daily evolution of asset states, showing when entities become infected, recover, or remain resilient. The slider bar can be used to scroll over the timeline.
Finally, we reconstruct the underlying interaction network. This network reveals the channels through which stress propagates, highlighting super-spreaders that amplify contagion through correlation.
Arrows represent potential contagion or influence between nodes (correlation, mutual information and causality). Red arrows indicate active contagion when both the source and target nodes are in a sick state. Orange arrows originate from super-spreader nodes, highlighting entities with a disproportionate influence on contagion dynamics. Gray arrows represent connections involving at least one recovered node, reflecting reduced or inactive transmission pathways. Light gray arrows correspond to baseline connections with no specific contagion state.
Each node is positioned according to four statistical moments of its daily return distribution: volatility, mean, skewness, and kurtosis, meaning that spatial proximity reflects similarity in return dynamics.
Cluster membership
The clusters shown in the 3D plot below emerge directly from how individual stocks behave during the outbreak window.
Each stock is represented as a point in a three-dimensional space defined by (i) the fraction of days where returns are lower than -5%, (ii) the average time needed to recover after suffering a -5% loss in a day, and (iii) the fraction of days spent in a healthy state. Stocks that occupy nearby positions in this space exhibit similar recovery dynamics. The clustering algorithm groups stocks solely based on this behavioral similarity.
The envelopes drawn around each group indicate the range of behaviors covered by a cluster. Clusters that lie far apart in the plot correspond to fundamentally different ways in which market stress propagates and dissipates during the crisis.
- Tickers belonging to cluster 0 struggle to recover and are sick for a longer time period. They are in a critical state and suffer great losses.
- Tickers belonging to cluster 1 on the other hand recover fast from a shock in the market.
- Tickers belonging to cluster 2 stay pretty much healthy most of the time, with a wide range of recovery times. They are the most immune group out of all.
- Tickers belonging to cluster 3 also struggle to recover but are overall less impacted by strong changes.
This classification helps us understand the role of each stock in the outbreak.
Part III: Sectorial Analysis
Maybe we have been looking at the problem from a wrong angle. What if the epidemic was actually caused by an entire sector of activity? This is what this section aims to explore.
We will reuse the algorithm, but this time, over the different sectors of the market. Below is a non-exhaustive bubble map of each sector and the 20 largest stocks they contain.
Similarly as before, we display all sectors on a reproduction index--temporal leadership--centrality map and the resulting PPI indices.
For the case of the dot-com bubble burst, the sectorial patient zero is Utilities.
The timeline below tracks how sector-level health evolves over time.
We then reconstruct the network of inter-sector interactions. This network reveals how stress is transmitted across sectors, identifying which sectors act as conduits that amplify or dampen systemic risk.
Case File: Subprime Crisis
In October 2008, the financial system stopped trusting itself. Years of hidden leverage, opaque balance sheets and mispriced risk surfaced almost simultaneously. What had been framed as a localized housing issue suddenly revealed deep fractures at the core of global finance. This outbreak is widely regarded as one of the most severe market upsets in modern history.
From the daily market returns below, large negative swings are compressed into a short time span, punctuated by sharp but short-lasting rebounds. Unlike the dot-com episode, here losses arrive in strong bursts.
Here the market records negative daily returns on 80% of trading days, with a cumulative loss of −22.02% over this time period.
Press RED BUTTON below and see what happens.
Patient zero has been identified as entity GS.
Now that a patient zero has been identified, we might want to understand how the outbreak unfolded. It is interesting to understand the relative roles played by all assets during the crisis.
The 3D plot below places every stock in the market inside a common contagion space. Each point represents an entity, positioned according to three dimensions: reproduction index, centrality and temporal leadership (all normalized). These values are the mean values over the entire timeframe.
To move from this static snapshot to a dynamic view, we now follow how the outbreak unfolds over time. The timeline below tracks the daily evolution of asset states, showing when entities become infected, recover, or remain resilient. The slider bar can be used to scroll over the timeline.
Finally, we reconstruct the underlying interaction network. This network reveals the channels through which stress propagates, highlighting super-spreaders that amplify contagion through correlation.
Arrows represent potential contagion or influence between nodes (correlation, mutual information and causality). Red arrows indicate active contagion when both the source and target nodes are in a sick state. Orange arrows originate from super-spreader nodes, highlighting entities with a disproportionate influence on contagion dynamics. Gray arrows represent connections involving at least one recovered node, reflecting reduced or inactive transmission pathways. Light gray arrows correspond to baseline connections with no specific contagion state.
Each node is positioned according to four statistical moments of its daily return distribution: volatility, mean, skewness, and kurtosis, meaning that spatial proximity reflects similarity in return dynamics.
Cluster membership
The clusters shown in the 3D plot below emerge directly from how individual stocks behave during the outbreak window.
Each stock is represented as a point in a three-dimensional space defined by (i) the fraction of days where returns are lower than -5%, (ii) the average time needed to recover after suffering a -5% loss in a day, and (iii) the fraction of days spent in a healthy state. Stocks that occupy nearby positions in this space exhibit similar recovery dynamics. The clustering algorithm groups stocks solely based on this behavioral similarity.
The envelopes drawn around each group indicate the range of behaviors covered by a cluster. Clusters that lie far apart in the plot correspond to fundamentally different ways in which market stress propagates and dissipates during the crisis.
- Tickers belonging to cluster 0 stay pretty much healthy most of the time, with a wide range of recovery times.
- Tickers belonging to cluster 1 are get sick almost all the time, but they manage to recover faster than the others.
- Tickers belonging to cluster 2 struggle to recover but are less impacted by strong changes. - Tickers belonging to cluster 3 also struggle to recover but get sick more times in this time period.
This classification helps us understand the role of each stock in the outbreak.
Part III: Sectorial Analysis
Maybe we have been looking at the problem from a wrong angle. What if the epidemic was actually caused by an entire sector of activity? This is what this section aims to do.
We will reuse the algorithm, but this time, over the different sectors of the market. Below is a non-exhaustive bubble map of each sector and the 20 largest stocks they contain.
We can use the same reasoning as for individual tickers but this time we apply it to our newly defined sectors.
The patient zero in this case is the Real Estate Sector, which totally makes sense in this context! This is a great sign.
The timeline below tracks how sector-level health evolves over time.
We then reconstruct the network of inter-sector interactions. This network reveals how stress is transmitted across sectors, identifying which sectors act as conduits that amplify or dampen systemic risk.
Case File: COVID-19
The COVID-19 crisis did not originate inside the market as economic activity was frozen by policy decisions, supply chains fractured overnight and uncertainty replaced all prior narratives. Public places were shutdown, international trade was disrupted and toilet paper shortages could be witnessed all around the globe.
The return series below captures this instability.
In this scenario, the market records negative daily returns on roughly 53% of trading days, with a cumulative loss of −22.90%, meaning the negative days carry a much stronger weight than the positive return days. The analysis using our pipeline will give proper insight into the COVID-19 Crisis.
You can press the RED BUTTON below to proceed with the analysis.
Patient zero has been identified as entity GILD.
Now that a patient zero has been identified, we might want to understand how the outbreak unfolded. It is interesting to understand the relative roles played by all assets during the crisis.
The 3D plot below places every stock in the market inside a common contagion space. Each point represents an entity, positioned according to three dimensions: reproduction index, centrality and temporal leadership (all normalized). These values are the mean values over the entire timeframe.
To move from this static snapshot to a dynamic view, we now follow how the outbreak unfolds over time. The timeline below tracks the daily evolution of asset states, showing when entities become infected, recover, or remain resilient. The slider bar can be used to scroll over the timeline.
Finally, we reconstruct the underlying interaction network. This network reveals the channels through which stress propagates, highlighting super-spreaders that amplify contagion through correlation.
Arrows represent potential contagion or influence between nodes (correlation, mutual information and causality). Red arrows indicate active contagion when both the source and target nodes are in a sick state. Orange arrows originate from super-spreader nodes, highlighting entities with a disproportionate influence on contagion dynamics. Gray arrows represent connections involving at least one recovered node, reflecting reduced or inactive transmission pathways. Light gray arrows correspond to baseline connections with no specific contagion state.
Each node is positioned according to four statistical moments of its daily return distribution: volatility, mean, skewness, and kurtosis, meaning that spatial proximity reflects similarity in return dynamics.
Cluster membership
The clusters shown in the 3D plot below emerge directly from how individual stocks behave during the outbreak window.
Each stock is represented as a point in a three-dimensional space defined by (i) the **fraction of days where returns are lower than -5%**, (ii) the **average time needed to recover** after suffering a -5% loss in a day, and (iii) the **fraction of days spent in a healthy state**. Stocks that occupy nearby positions in this space exhibit similar recovery dynamics. The clustering algorithm groups stocks solely based on this behavioral similarity.
The envelopes drawn around each group indicate the range of behaviors covered by a cluster. Clusters that lie far apart in the plot correspond to fundamentally different ways in which market stress propagates and dissipates during the crisis.
- Tickers belonging to cluster 0 recover very quickly from being sick, and rarely get sick.
- Tickers belonging to cluster 1 have mixed attributes, but they are still very much impacted and don't recover fast.
- Tickers belonging to cluster 2 are the stocks that are most impacted by the outbreak.
- Tickers belonging to cluster 3 show fast recovery and better health conditions compared to the other clusters.
This classification helps us understand the role of each stock in the outbreak.
Part III: Sectorial Analysis
Up to this point, the analysis has focused on individual assets. However it is bold to assume that only one entity is responsible for a market shutdown as these epidemics often propagate through entire sectors.
In this section, we shift the level of observation from assets to sectors. Each sector is treated as an aggregated entity whose health reflects the combined behavior of its constituent stocks during the outbreak window. Below is a non-exhaustive bubble map of each sector and the 20 largest stocks they contain.
The 3D plot below positions sectors in a contagion space analogous to the one used for individual assets. It captures how early sectoral stress emerges, how central each sector is within the market structure, and how strongly stress propagates from one sector to another.
This change in scale allows us to distinguish between crises driven by a small number of influential sectors and those that spread broadly across the economic landscape.
The patient zero in this case is Consumer Staples.
The timeline below tracks how sector-level health evolves over time.
We then reconstruct the network of inter-sector interactions. This network reveals how stress is transmitted across sectors, identifying which sectors act as conduits that amplify or dampen systemic risk.
Part IV: Crisis Signatures
Up to this point, each crisis has been examined in isolation through its own case file. We now take a step back and ask a broader question:
Do different crises stress the market through the same underlying mechanisms or through fundamentally different ones?
To answer this, we compress each crisis into a latent market signature using Principal Component Analysis (PCA) applied to the immune phenotype features introduced earlier (variance, health state, returns, etc.).
Rather than focusing on individual assets, PCA extracts the dominant collective mode that explains how the market behaves as a system during each outbreak.
Each crisis is therefore summarized by:
- its dominant PCA mode (PC1), capturing the main direction of market stress,
- its secondary PCA mode (PC2), capturing the main alternative stress pattern that PC1 does not explain,
- the fraction of variance explained by each mode, measuring how synchronized the market becomes,
- and the feature loadings defining the underlying crisis mechanism.
Each bubble represents a crisis period. The vertical position and size of a bubble reflect the fraction of variance explained by the selected PCA component: larger and higher bubbles indicate more synchronized, low-dimensional market behavior. Hovering over a bubble will reveal the standardized feature loadings that define the corresponding crisis mechanism. Positive loadings indicate features that intensify the crisis mode, while negative loadings indicate stabilizing or opposing effects.
What this comparison reveals
The crisis signatures highlight clear structural differences between major market shocks.
The dot-com bubble emerges as the most synchronized episode. Its dominant PCA mode explains a larger fraction of the variance than in the other crises (roughly 57% of the variance is explained by this dominant mode), indicating that market behavior collapsed onto a single prevailing stress mechanism. This mode is driven by a sharp rise in sick assets and longer recovery times, while the fraction of healthy assets contributes in the opposite direction. In this sense, the dot-com crisis reflects a broad erosion of market health concentrated along one dominant dimension.
The subprime crisis displays a less concentrated structure. Its lower explained variance suggests a more heterogeneous propagation of stress, consistent with a crisis originating within specific financial institutions and spreading through tightly connected parts of the system rather than affecting all assets uniformly.
The COVID-19 shock, while severe in magnitude, is characterized by a dominant mode that explains even less variance. This indicates that market stress unfolded through multiple concurrent channels, reflecting an abrupt shock impacting assets and sectors in uneven ways.
Overall, these signatures show that crises with similar aggregate losses can differ substantially in their internal dynamics. PCA reveals that each episode is governed by a distinct latent stress structure, shaping how quickly, how broadly, and how uniformly financial contagion spreads through the market.
Part V: Systemic Risk Radar
Another way to compare these three crises is to look at how different they can be in terms of liquidity exchanges, volatility, causality between the nodes, how large the mean contagion index is, and how much the market goes down. The plot below gives a clear picture of what is happening in these three time periods.
We can easily assess that, overall, the Subprime Financial Crisis had the strongest market impact, as nearly all indicators dominate those of the other two crises (particularly contagion, negative weekly returns, and liquidity disruption). By contrast, the COVID-19 episode exhibits higher causality and volatility. This plot therefore provides a compact way to compare crisis severity. This means that higher values across multiple dimensions signal a more systemic market breakdown.
Under this immunology-inspired lens, the Subprime Financial Crisis resembles a highly contagious outbreak, with stress spreading broadly across assets, severe liquidity impairment, and deep persistent losses. The COVID-19 shock instead appears as an acute episode, characterized by violent volatility and strong causal coupling but more limited propagation and faster stabilization. Finally, the Dot-com bubble, emerges as a slower and more contained process with weaker transmission and a milder systemic footprint.
Conclusion
Based on the results found in the previous sections, we can confidently say that using the immunology approach on a single entity is less relevant than looking at a broader perspective (e.g. sectorial), but it still reveals microscopic dynamics where the sectorial approach is more macroscopic. It is useful to understand the propagation mechanisms, the speed at which a disease propagates between nodes and how factors such as centrality, temporal leadership and contagion come in play. We can tell which stocks are more or less affected than others and thereby group them together.
The patient zero analysis with the sectorial approach can yield coherent results but is highly dependent on the window that we decide to choose, as it is a temporal parameter. However, super-spreaders are more insightful as they reveal which nodes are key in crisis periods and tell us who really takes this financial disease to a higher level and cause a widespread outbreak.
Comparing different crises, we found that each of them had an immunological signature and that they all spread through different contagion pathway mechanisms using PCA. Where some outbreaks were more synchronized and coordinated, other showed more chaos. Using the risk radar, factors such as liquidity exchange and volatility reveal other parameters that can be used to compare between cases.
Overall, this project has limitations as to where and how it is applied. Nevertheless, it is still valuable as a conceptual and analytical bridge between network science and financial risk analysis. The framework is inherently sensitive to modeling choices such as thresholds, time windows and data resolution, which constrains its use as a predictive tool. However, these same constraints make it well suited for comparative and diagnostic purposes. By highlighting structural vulnerabilities and identifying nodes that disproportionately amplify stress, the approach opens perspectives for future extensions.