WFGY/TensionUniverse/Experiments/Q105_MVP/README.md

TU Q105 MVP: toy systemic crash warnings

Status: MVP notebook A is implemented and fully offline. This page will be updated again when notebook B is added.

This page reports toy experiments for TU Q105.
The aim is to show how small networked systems can expose tension between local early-warning indicators and global crash risk.

Navigation

• ← Back to Experiments index
• ← Back to Event Horizon (WFGY 3.0)

---

0. What this page is about

TU Q105 looks at prediction of systemic crashes.

Instead of real financial markets or infrastructures we start with toy networks:

• nodes with local balance sheets or loads
• edges representing exposures or dependencies
• simple failure rules for cascades

We then define:

• early-warning indicators
• crash outcomes
• and tension observables when indicators fail to warn

The current MVP implements Experiment A as a single offline notebook.
All runs are fully simulated and deterministic once the random seed is fixed.
No API key is needed.

---

1. Experiment A: network contagion and local indicators

1.1 Research question

In a simple contagion model on a network, can we define a scalar observable T_warning that

• is small when early-warning indicators successfully flag upcoming crashes
• grows when indicators remain calm but large cascades occur

The goal is not to build a realistic risk system.
The goal is to make the tension between local indicators and global crashes explicit and measurable.

1.2 World and early-warning schemes

The notebook builds a tiny networked world and three naive early-warning schemes.

Configured world

• Nodes: 20 in total (5 core nodes, 15 periphery nodes)
• Network structure: core–periphery with additional ring edges in the periphery
• Crash definition: a cascade is a crash when at least 40% of nodes fail
• Scenarios per scheme: 500 independent shock-and-propagation runs
• Observed crash rate across all runs: about 0.8% of scenarios crash

Each node carries a simple load or capacity variable.
Loads are shocked and then propagated through the network using a threshold failure rule.
Once a node fails its load is pushed to neighbours, possibly triggering further failures.

Early-warning schemes

Before shocks are applied we compute a few aggregate indicators and apply three schemes:

• global_mean: uses the global mean of load divided by capacity as a single indicator
• tail_sensitive: focuses on how many nodes are close to their threshold
• core_focused: gives extra attention to core nodes in the network

Each scheme turns its indicators into a simple warning rule, for example
"issue a warning when mean load is above a threshold".

1.3 T_warning: how tension is measured

For each scheme and each scenario the notebook records:

• whether a large cascade (crash) occurred
• whether the scheme issued a warning before the crash window
• the overall crash rate in the world

From these outcomes we compute:

• FN_rate: fraction of crashes with no prior warning (false negatives)
• FP_rate: fraction of warnings that occur in non-crash scenarios (false positives)
• quiet_crash_rate: crash rate conditional on no warning

The scalar tension observable T_warning is then defined so that:

• high false negative rate and high quiet crash rate give a high T_warning
• excessive false positives also increase T_warning
• lower values mean better calibrated warnings for this particular world

In the current configuration all three naive schemes behave badly.
They almost never issue warnings, so every crash is a quiet crash and T_warning is high.

1.4 What the current run shows

The notebook prints a summary table for the three schemes:

• global_mean
• tail_sensitive
• core_focused

For the current world we obtain:

• Crash rate across all scenarios: about 0.008
• For every scheme:
  • FN_rate close to 1.00 (all crashes happened with no warning)
  • FP_rate close to 0.00 (warnings are almost never issued)
  • quiet_crash_rate close to 0.008
  • T_warning around 3.008 for all three schemes

In words:

The world does occasionally crash.
The simple early-warning schemes almost never speak up.
All systemic crashes happen in quiet conditions.
Tension is high because the indicators are not aligned with actual crash risk.

The notebook also saves two plots in this folder.

Crash probability versus mean load ratio

Each point shows the empirical crash probability in a bin of the mean load to capacity ratio.
Most bins at lower ratios have almost zero crash probability.
The rightmost bin with higher ratios has a noticeably higher crash probability.

This shows that the toy world itself does carry a usable signal:
when the system is globally more loaded, crashes become more likely.

T_warning per scheme

All three schemes end up with nearly identical and relatively high T_warning values.
They fail to exploit the available signal and leave crashes almost completely unwarned.

This is the intended tension for TU Q105:
even in a small 20 node world it is possible to measure how badly a given
early-warning design matches the actual structure of crash risk.

1.5 How to reproduce

To rerun or inspect Experiment A:

1. Open the notebook with the Colab badge above or from this repository: Q105_A.ipynb.
2. Read the header comments that describe the network, failure rules and indicators.
3. Run all cells. The notebook will:
   • build the network
   • simulate 500 scenarios for each early-warning scheme
   • compute T_warning and print a summary table
   • regenerate the two plots (Q105A_indicator_vs_crash.png and Q105A_T_warning_per_scheme.png)
4. Optionally modify the world parameters or warning thresholds to explore how
   T_warning changes when indicators are improved or degraded.

No API key is needed. Everything runs offline inside the notebook.

---

2. Experiment B: model based versus data based risk assessment

This section describes a planned follow up experiment. The notebook is not yet implemented.

2.1 Research question

Can we define a tension observable T_model_data that captures when a simple
model based risk score disagrees with an empirical data based score?

2.2 Setup (planned)

Using the same simulation engine, the future notebook will:

• generate datasets of scenarios with
  • basic features
  • whether a crash occurred
• train a small supervised learner to predict crash probability from features
• compare
  • model based risk scores from the structural indicators
  • data based risk scores from the learner

T_model_data will be defined as a function of:

• divergence between the two risk scores
• misranking of scenarios by each score relative to true crash outcomes

2.3 Expected pattern

We expect:

• low T_model_data when structural and data based scores agree
• higher T_model_data when a structural story and empirical patterns diverge

2.4 How to reproduce (future)

Once Q105_B.ipynb exists:

• open and inspect the feature definitions
• train the learner and compute both risk scores
• compare T_model_data across parameter settings

---

3. How this MVP fits into Tension Universe

TU Q105 treats systemic crash prediction as a tension between:

• local indicators and global outcomes
• structural models and data based models

The current MVP provides:

• a network contagion toy model with T_warning for early-warning schemes
• visual evidence that the toy world carries a usable signal while naive schemes fail to use it

Both notebooks in this series are designed as inspection friendly demos,
not as real risk management systems.

For overall context:

• Experiments index
• Event Horizon (WFGY 3.0)

---

Charters and formal context

This page follows:

• TU Effective Layer Charter
• TU Encoding and Fairness Charter
• TU Tension Scale Charter