Engage
Crisis & Risk
Technology
Index
2026
ICDS RESEARCH TEAM
AI Risk Protection Index (ARPI)
A Leading Early-warning System for Risks in The AI-company Market
Share
Segment Analysis
How does this research apply to your industry and market context?
Find out more about a specific country or a specific market for goods or services.
Request a Comment
This article introduces the AI Risk Protection Index (ARPI), an early-warning system for detecting emerging risk in AI-sector equities through behavioral media signals rather than price-dependent indicators such as VIX. Drawing on GDELT media tonality and volume data alongside Yahoo Finance price series (139+ trading days) and FRED VIX data, the index combines five active metrics — Tonality Divergence, Information Asymmetry, Latent Price Divergence, Ambiguity Score, and Frontier Disclosure Index — into a single composite score classified across four risk quadrants. A retrospective calibration against past events (December 2024 – June 2025), including an L-Pattern signature identified in two cases (NVDA Q4, GPT-4.5), suggests behavioral signals may precede price movement by 3–15 trading days, with the tonality-jump coefficient replicating in sign a pattern previously observed in the ICDS FMCG corpus (N=66). The index remains in a verification stage: it has been calibrated retrospectively rather than tested prospectively, and the current sample (N=10) is explicitly identified as sufficient for directional VIX correlation but insufficient for stable statistical weighting of individual components. Real-time monitoring with dated signal logging is proposed as the next validation step.
Abstract
Research Presentation
Download PDF

AI Risk Protection Index (ARPI)

A leading early-warning system for risks in the AI-company market
Financial bubbles carry diagnostic markers refined over two centuries of economic history. Valuations detached from reality — stocks trade at multiples unsupported by corresponding profit. Debt-financed growth — companies scale using borrowed money rather than cash flow. Narrative displaces numbers — “this time it’s different.” Circular financing — money circulates within a closed loop of participants, creating an illusion of growth. Market concentration — a small number of companies pull the entire index upward.
Applied to the AI sector, each of these markers is present — but with varying intensity across its different segments.

What the numbers say

The circular financing scheme is documented: OpenAI holds a stake in AMD, Nvidia invests in OpenAI, OpenAI counts Microsoft as a major shareholder, and Microsoft is a major customer of CoreWeave, in which Nvidia holds a stake. At the same time, Microsoft accounted for roughly 20% of Nvidia’s revenue. Total AI-industry revenue is estimated at under $50 billion against a trillion dollars in investment. OpenAI, with roughly $12 billion in revenue, posts an operating loss of $8 billion — and expects losses to double in 2026. Market concentration has reached a half-century high: at the end of 2025, five companies accounted for 30% of the S&P 500. AI-related stocks delivered 75% of the S&P 500’s gains since ChatGPT’s launch.
On the other hand, JPMorgan concluded that AI does not meet the classic criteria of a financial bubble. Unlike the dot-com era, the current boom is led by companies with years of proven profitability — Meta, Amazon, Microsoft. Revenue and net income at the largest players are growing quarter over quarter, and demand for AI services remains high. A 2026 NBER study found that although 90% of companies report no measurable impact of AI on productivity, executives forecast productivity growth of 1.4%.

How the situation differs from the dot-com era — and why it matters

In 2000, the largest companies (Cisco, Intel, Microsoft) also fell 60–80%. But they survived. The small ones disappeared entirely. The structure today is different. Layers 1 and 2 of the AI market are embedded in the real economy far more deeply than telecom was in 2000. Google is no longer competitive without AI. This lowers the risk of total collapse, but does not eliminate a painful repricing.
The most precise historical analogue is not the dot-com crash but the story of 19th-century railroads: the technology is real, the economics are revolutionary, but 90% of individual companies went bankrupt. Those who controlled the physical infrastructure survived.
The key conclusion from this analysis: predicting an “AI crash” as a single event is the wrong framing. The right framing is to track in which layer, and when, a repricing will occur. And that requires a model.

1. Timeline: how the market arrived at where it is now

And could anyone have seen it earlier?
Let’s look back across the three years the AI market has existed — a path from euphoria to the first institutional warnings. Let’s treat this path as a detective story: at what point were the clues already on the table, waiting for someone to put them together?

Key observation: the gap between “the signal appeared in the data” and “consensus formed” ranged from 6 to 18 months. For AI wrappers, the first data on unviable unit economics appeared in niche media in late 2023. Mainstream consensus formed in autumn 2025. Potential lead window: ~18–24 months.

2. Could we have seen this in December 2024?

Retrospective check
Let’s attempt a retrospective check: were there signals in the data that could have been captured — and even acted on? The answer is yes. Below is a comparison of what the system would have shown at the time, and when the market arrived at the same conclusions.

Can a predictive model be built on this data?

Such an attempt has been made — the model is in the verification stage. We are calibrating it on events that have already occurred, December 2024 – June 2025. The next step is to launch real-time monitoring with dated signal logging. The model can be validated once the events occur. If the forecast is confirmed, the system will be conclusively verified.

A critically important distinction: retrospective analysis shows the signals were there. And we could have seen the emerging trend earlier — if our predictive system had already been running at the time.

3. Predictive model built on observation vs. market consensus: where we agree — and where we diverge

Map of positions: consensus, polarization, three contrarian theses
The value of a predictive model is tested not where it agrees with the market, but where it diverges. Divergence itself is the source of informational advantage.

Three ARPI contrarian positions:

  • OpenAI: the market is polarized; ARPI records the Ambiguity Score as an independent risk signal — regardless of who is right in the bull/bear debate. Uncertainty itself creates risk.
  • xAI: the market sees a successful funding round. ARPI sees Semantic Bleed masking a real negative signal, and a widening FDI. A contrarian position.
  • Anthropic: the only contrarian position in the opposite direction — not HIGH, but MODERATE. The market overlooks it; ARPI flags it as the likely Layer 3 winner.

4. Six dimensions — the ARPI architecture

Five active metrics plus one deflator. All data from GDELT and open sources.
Each metric is independent of market prices — all are calculated from media tonality, media volume, and narrative structure. This is a fundamental difference from VIX-dependent models: the ARPI signal appears before VIX begins to move.
ARPI = normalize( TDI×2.0 + IAI×1.8 + LPD×2.2 + AS×1.5 + FDI×1.8 ) × SBI Scale:
0–30 = LOW
31–55 = MODERATE
56–75 = HIGH
76–100 = CRITICAL

5. Operational matrix: BRI × Tonality → recovery profile

The quadrant is determined before the event — based on TDI and AS, calculated continuously
The matrix classifies any AI event into one of four quadrants and predicts the price-movement profile.
BRI HIGH (> 2.0) — loud narrative BRI LOW (< 0.5) — quiet narrative
Tonality
NEGATIVE
D — CRITICAL
Narrative intensity + systemic fear.
Decline −15...−25%. Recovery 20–40
days or L-pattern.
C — QUIET BLOW
Quiet narrative with fear → W- or L-pattern.
Liberation Day: BRI 0.87, tone −2.33 →
−13%, 16 days.
Tonality
POSITIVE
A — NARRATIVE NOISE
Loud narrative without fear → V-shaped
rebound.
DeepSeek: BRI 4.26, tone +0.23 →
recovery in 1 day.
B — EVENT NOISE
Quiet and positive → absorbed by the
trend.
NVDA Q1 FY2026: BRI 0.05, neutral tone
→ +5% at t+20.

The paradox of clarity: ambiguously bad (60/40) is more dangerous for a stock than unambiguously bad (90/10). Algorithmic traders find opportunity precisely in the zone of uncertainty. Robinhood: volatility 6.3% at moderate disbalance vs. 2% when one-sided (ICDS Article 8).

6. L-Pattern Detector: quiet disappointment as the worst-case scenario

IAI + LPD: “major news — quiet media field” — a harbinger of protracted decline
L-pattern — a decline without recovery by t+20 — the most dangerous profile in the corpus. Recorded twice: NVDA Q4 and GPT-4.5. Both cases share the same configuration.

LPD = 1 when:
(1) |XLK t+1| < 0.5% — neutral immediate reaction;
(2) Δtone(post−pre) < −0.5 — tonality worsened after the event.
Accuracy on N=10: 2/2, without false positives.

Weight in ARPI: ×2.2 — the highest.
This is precisely what the turning point in Layer 3 will look like — not like DeepSeek (loud), but like NVDA Q4 (quiet).

7. Frontier Decoupling Index: a proxy for non-public risk

First public tool for assessing Layer 3 risk (OpenAI, xAI, Anthropic)
Goldman, Bridgewater, and JPMorgan assess Frontier AI through Nvidia capex, patent filings, and Azure contract volume. FDI uses a different dimension — the public narrative, which aggregates the informed opinion of thousands of independent authors before it materializes in financial results.

FDI = 10-day rolling average tonality [OpenAI + xAI + Anthropic] minus 10-day rolling average tonality [Nvidia + Google + Microsoft].
A widening negative spread (FDI < −1.2) is a trigger for HIGH in Layer 3.
Current signal (May 2026): FDI = −0.9 for OpenAI, trend pointing downward.

8. Current ARPI and signal profile for Layer 3

Indicative estimates (May 2026) and the signal sequence leading up to a critical event
✓ normal · ⚠ attention · ✗ alarm · ↓ rising trend

What this would look like in the data — signal profile for Layer 3:

We do not know the exact date. We know what it will look like in the data — 5 to 15 trading days before the market feels it.

What ARPI claims:

  • Behavioral media signals contain leading information 3–15 trading days ahead of price movement.
  • The L-Pattern Detector identifies “quiet disappointment” with 2/2 accuracy on N=10.
  • Tonality Jump (TDI) reproduces the sign of the coefficient from FMCG (N=53) — a candidate for a universal ICDS law.
  • FDI is the first public proxy for assessing the risk of non-public AI companies.
  • ARPI is in the verification stage. The next step is real-time monitoring with dated signal logging.

What ARPI does not claim:

  • ARPI did not predict these events — it was calibrated on them retrospectively.
  • N=10 is sufficient for VIX correlation, but not for the statistical weights of TDI and IAI outside a directional trend.
  • VIX as the “only significant predictor” is partly tautological. This is precisely why an independent GDELT signal matters more.
Methodology
Price data: Yahoo Finance (XLK, NVDA, 139+ trading days). VIX: FRED VIXCLS. Media volume and tonality: GDELT DOC 2.0 API (timelinevolraw, timelinetone). Cross-corpus validation: ICDS FMCG (N=66, 26 countries, Articles 1–8), ICDS Fintech (N=12, 31 peaks). BRI: (E%/100)×(rel_vol/1000)×k. Formula: ARPI = normalize(TDI×2.0 + IAI×1.8 + LPD×2.2 + AS×1.5 + FDI×1.8) × SBI. Cross-Corpus Validator (CCV): agreement in the sign of the TDI coefficient across the FMCG and AI corpora — the condition for recognizing a pattern as universal across sectors.
ICDS — Institute of Communication and Data Science
Behavioural Data · Media Metrics · Predictive Risk Analytics · icds.institute
May 2026
Related Research

You may also find relevant

ENGAGE WITH ICDS
Interested in commissioning research or accessing institutional data?