# Model methodology and limitations

## Question

"If a $20T public/private market-value crash occurs by March 2027, which observable failures would need to happen in the preceding five months, which companies would likely show stress earliest, and how could financial distress propagate?"

This is an **inverse conditional** question — one can calculate a *path consistent with a target* without knowing the unconditional probability of arriving there.

## Economic edge ontology

For each `data/flows.json` record, `source → target` denotes the direction of the nominal cash/spending/guarantee commitment (not necessarily cash already settled).

- **Equity:** investor → issuer. Issuer distress transmits asset-markdown pressure *back to investor*. An "up to" investment must not be booked as funded equity.
- **Cloud:** customer → supplier. Customer credit or capacity cuts transmit unearned-revenue stress *to supplier*. A multiyear cloud total is not equivalent to loss of cash today.
- **Hardware:** buyer → supplier. Buyer capex cuts transmit *to seller*; only disclosed amounts can be weighted.
- **Debt:** lender → borrower. Borrower default or collateral impairment transmits *to lender*. A facility limit is not always outstanding principal.
- **Guarantee/backstop:** provider → covered entity. Conditional payment stress may transmit *to provider*, capped by facility and contract triggers; no modeled call is a reported liability.
- **Warrants:** dilution/optionality, not a funded equity purchase. Undisclosed nominal relationships do not enter dollar-weighted propagation.

Each edge has `stage`, `horizon`, `note`, `source_url`, `included`. Two counterparties with multiple contracts appear as independent ledger rows; the view aggregates them for readability but the engine handles each row separately. `included:false` typically indicates overlap, missing values, or optional tranches; toggle separately. **Do not sum differently typed gross contracts to estimate net debt, profit or market value.**

## Directed propagation

For counterparties A and B, a relevant relationship with headline value V (USD billions) is scaled as:

`notional_intensity = ln(1 + V) / ln(301)`.

At each of three propagation steps, stress increments follow reverse-risk direction for each edge:

`Δ risk_recipient = coefficient(edge type) × notional_intensity × scenario_channel_weight × path_progress × max(0, risk_source − base_risk_recipient)`.

The index is clipped to [0,1]. A cap limits incremental changes per step. The coefficients, base risk indices and stage-specific path are **judgmental controls** used to make mechanics interpretable, *not estimates fit to defaults*. High nominal values have diminishing impact to prevent a $300B cloud contract from being treated as a direct $300B credit exposure. The visualization is a **causal hypothesis explorer**; it must not be used to estimate a company's default probability.

Circle area uses `base_radius × sqrt(1 − cumulative_markdown)` to reflect shrinking assumed capitalization. Companies with missing/unverified cap data use relative size proxies. Color is a *modeled composite stress index*: stable <34, watch 34-54, warning 55-76, crisis 77-100. Baseline indices intentionally reflect heightened credit exposure of leveraged neocloud operators compared with cash-generating hyperscalers, but values remain uncalibrated assumptions. The dashboard CoreWeave CDS override uses an intentionally simple index mapping `clip(0.12 + 0.00061 × spread_bps, 0.12, 0.98)` to permit interactive canary sensitivity; **the mapping is not a conversion from spread into default probability**.

## Scenarios and timeline

Six snapshots: Oct 25, Nov 30, Dec 31, Jan 31, Feb 28 and Mar 31. For the March thesis, vulnerability intensifies during Nov-Feb while most hypothesized equity repricing is back-loaded to March itself. Other scenarios use different paths. No intermediate or future state is presented as a historical observation.

The stress cases intentionally distinguish five drivers:

1. **Funding/credit:** widening borrowing costs, failed refinance, guarantees, sponsor liquidity, counterparty credit.
2. **Demand/ROI:** reduced AI inference economics, utilization and cash-conversion shortfalls, weaker order books.
3. **Collateral/depreciation:** accelerated obsolescence and GPU resale values; asset-backed loan haircuts.
4. **Duration/capex:** power/interconnect bottlenecks, delayed commissioning, customer prepayment dependence.
5. **Equity repricing:** earnings expectations and valuation-multiple compression, even without defaults.

## Avoiding $20T double counting

The `uniqueBucketLoss` function assigns disjoint equity-market buckets by issuer:

- Mag 7 inclusive of Nvidia (25% haircut).
- **Only Nvidia's incremental semi shock above the Mag 7 loss** (25 percentage points in the severe case), not a second 50% of Nvidia.
- Non-Mag7 semiconductors (50%).
- Other infrastructure excluding neoclouds and semiconductor names (30%).
- Private AI labs, *excluding SpaceX* (55% assumption).
- SpaceX, which has been publicly traded since June 2026 (55% in severe case).
- Neoclouds outside the other buckets (70% in severe case).

With starting bucket bases of $23.727T, $5.09T Nvidia, $6.23T non-Mag7 semis, $2.0T other infrastructure, $2.9T private labs, $1.95T SpaceX and $0.2T separate neoclouds, the severe scenario yields **approximately $13.7T**, leaving **about $6.3T of additional loss required** to reach the $20T thesis. Many bucket bases are deliberately editable working estimates rather than validated market-wide aggregates. A full reconstruction would need a point-in-time complete company list, all cross-ownership links, public float/share counts, issuer classification and explicit private valuation marks.

This is *market-value change*, **not** investors' realized cash loss, capital expenditure, or destroyed physical assets. Private markdowns and public stocks cannot be valued with equally frequent and transparent price discovery.

## Probability / calibration readiness: three distinct events

The model currently displays **Not calibrated**, rather than using an arbitrary 3% starting prior or a 5.5% subjective posterior. It distinguishes:

1. **AI equities decline at least 20%** by the horizon: requires a defined universe, dated index and option/downside distributions, and validation of tail model assumptions.
2. **Systemic financing distress**: requires independent issuer-level default/refinancing data, credit spreads, capital availability, maturity schedules, cross-issuer correlations and actual capex cancellation signals.
3. **Exactly $20T public/private market-value loss by March 2027**: requires non-overlapping issuer valuations, private valuation uncertainty, spillovers beyond AI, timing calibration and an explicit distribution for joint tail events.

**Do not confuse** a market price or model vulnerability index with a default probability or system crash probability. CoreWeave's CDS-based tool is an approximate *risk-neutral issuer default calculation* from \(\lambda = s/(1-R)\) and \(1-e^{-\lambda t}\), not a real-world default probability nor a probability of a March 2027 global crash. Its constant spread, recovery and liquidity-premium assumptions are strong simplifications. The CoreWeave credit *index* mapping is separately uncalibrated and used only to illustrate color/sensitivity.

The prior educational subjective Bayesian calculation and the optional single IPO-cluster multiplier were **removed from the live dashboard**. They had unverified likelihood ratios and could give a false impression of quantitative precision. IPO evidence remains dated and qualitatively classified, avoiding treating several correlated listing deferrals as independent systemic defaults.

## Falsifiability and early-warning procedure

In the Oct–Nov period, assess whether financing distress spreads from CoreWeave to two or more unrelated neoclouds and whether Anthropic/OpenAI funding proceeds on the indicated terms. In December, track actual lender refusal and collateral advance rates. In January, examine cloud utilization and demand, customer prepayments, hyperscaler capex revision, server orders, and conversion of RPO into realized revenue. In February, demand confirmation via earnings guidance, write-downs or genuine restructurings. If those links do **not** materialize, the particular March 2027 crash path should lose credence even if AI equities remain volatile.

`data/canary_watchlist.csv` contains proposed review cadence, hard provisional watch thresholds and evidence sources. Those thresholds are *screening rules* for investigation; they are not industry-standard covenants or calibrated systemic event predictors.

## Roadmap toward quantitative validation

Future versions should ingest dated financial histories and implement panel/hazard survival models for issuer solvency, empirical funding stress factor loadings, Monte Carlo marginal shocks with correlated draws, cash runway and debt-maturity computations, scenario conditioning (reverse Monte Carlo or importance sampling), and full sensitivity/error budgets. Backtesting and anti-lookahead controls are essential. A credible tool must separate *causal pathways* from merely correlated valuation changes.

## October 20 Bloomberg IPO-stress evidence

The neocloud market-access feed separates **market price discovery** (Elice's IPO first-day return and CoreWeave's drawdown) from **successful financing** (Elice raised ~$150M), **unpriced/paused offerings** (Firmus, AirTrunk, SB Energy) and **indicative valuation resistance** (Nscale). It also retains counterevidence: Elice's improved H1 operating profitability. A postponed IPO does **not** equal a debt default; offerings are influenced by rates and market windows.

These observations are **not edges** in the circular cash-flow network and are **not additional losses** in the $13.73T terminal capitalization ledger. The dashboard no longer shows the original subjective IPO-cluster multiplier. Correlated listing deferrals remain grouped as one qualitative market-access signal, not several independent likelihood multipliers, and the dated feed has no automatic live refresh.

## Expanded full-width reverse-stress stages

The six-month path is an analyst-authored **conditional sequence**, shown immediately below the network and canary ranking, not a model-inferred projected future. Each stage records its causal transmission chain, requirements, provisional numeric screening thresholds, observed evidence available by October 25, counterevidence/falsifiers, and which financial mechanism would transmit stress to the next stage. Future stages remain explicitly unobserved and cannot become 'confirmed' merely because the animation slider advances.