Reading the Next Bubble Before It Breaks

Building a risk index for transformative-technology markets — and putting it to its first live test. A behavioural risk index for the AI market.

For financial analysts, risk officers, and Institute partners

VM

Vadim Matyushkin Behavioral Scientist & Sociologist · ICDS

EditionResearch Brief · November 25, 2026

Scope

Date: July 2026.  Format: open research brief for external readership.

This is the story of an index. It begins with an old pattern — the way transformative technologies produce fortunes and ruins in the same decade — and asks whether that pattern leaves a trace in today’s data early enough to act on. It follows a risk index from its first idea to its first live test on the AI market, and reports, without flattering the instrument, what the test revealed.

Key Findings

An old pattern wearing new clothes

Every generation is told that this time is different. Every generation is partly right, and that is exactly what makes it dangerous.

In the 1840s, British investors poured their savings into railway companies. The technology was genuine and the economics were about to remake the world — and roughly nine in ten of the specific companies went bankrupt. The lines were still laid. The trains still ran. But the shareholders who had backed the wrong company lost everything, while those who owned the physical track beneath the boom survived to own the century.

In 2000, the same story wore the language of the internet. Cisco, Intel, and Microsoft, the giants of the moment, fell sixty to eighty per cent and lived. The thousand small companies with a dot in their name mostly vanished. The web was real; most of its first businesses were not.

This is not a coincidence repeated twice. It is the signature of a particular kind of market — one where a genuinely transformative technology arrives faster than the economics that would justify its prices. Call it a transformative-technology bubble. It is not a fraud and not a mania in the ordinary sense; it is a race in which the destination is real and most of the runners still fall.

The AI market of 2026 fits the family portrait. The circular financing is documented — Nvidia invests in OpenAI, which holds a stake in AMD, while Microsoft is both a large OpenAI shareholder and a large customer of a cloud firm Nvidia part-owns. Industry revenue sits below fifty billion dollars against roughly a trillion in investment. Five companies hold thirty per cent of the S&P 500. And yet, unlike 2000, the leaders are deeply profitable and their products are already woven into the real economy. The question is therefore not whether the whole thing collapses, but which parts reprice, how hard, and — the part that matters for anyone holding a position — whether it can be seen coming.

Who survives, and who does not

Before you can forecast a fall, you have to know who is standing where the ground is soft.

The railways and the dot-coms taught the same lesson twice: a transformative-technology market is never one market. It is a stack of layers with different economics and different odds of survival. Reading the AI market the way the survivors of past bubbles would have read theirs, four layers separate out — and each carries its own risk of repricing.

The recurring shape of a transformative-technology bubble: the technology delivers, the infrastructure survives, and most of the individual companies built on top of it do not. The winners are usually those who owned the layer everyone else had to rent.

The moral of the stack is the moral of the whole brief. Forecasting an ‘AI crash’ as one event is the wrong question, because the four layers will not fall together or at the same time. The right question is which layer, and when. Answering it needs an instrument — and the rest of this story is about building one.

Building the instrument

The balance sheet tells you what already happened. By the time it speaks, the position is already lost.

Traditional risk monitoring watches two dials: the share price and the quarterly report. Both are rear-view mirrors — they describe the crash after the metal has bent. The wager behind this index is that in a market ruled by social media and trading algorithms, the earliest signal is not financial at all. It is behavioural: the way a market talks about an event, in the days around it, before the price has finished moving.

Three signals, in plain terms

The index listens for three things, and each answers a different question about a shock.

From signals to an index

Building the instrument meant three deliberate moves. First, gather the data: for ten AI-sector events between December 2024 and June 2025, four independent streams each — price for the technology ETF and for Nvidia, the volatility index, media volume, and media tone; forty verified media queries in all. Second, line each event’s signals up against what the price actually did afterwards. Third, let the statistics say which signal earns its place — rather than deciding in advance which one ought to matter.

The result was humbling in a useful way, because it did not confirm the first guess.

Statistics from the ten-event calibration corpus; strengths reported as computed.

The fear spike was the honest core: on ten events it alone reached significance for the immediate drop. Tone came second, as a leading indicator of how long a wound takes to heal — a direction rather than a settled number at this sample size. Narrative loudness, the signal one might expect to matter most in the age of viral markets, turned out to be the weakest price predictor, useful only to classify how loud a story runs, not how far a price falls. Ten events do not settle a science, and the index was published knowing it.

The rule it produced

From those three signals came a simple rule an analyst could use inside a single session. A sharp fear spike within forty-eight hours marks a cascade — expect a deeper fall and a slow recovery over fifteen to twenty trading days. A fall in fear marks relief — expect a bounce the next day. Anything in between is a narrative event, its recovery shaped by tone and by how loud the story runs. That was the instrument. Then the market gave it something to read that it had never seen.

The first live test: June 2026

A model that has only ever seen the past is a memory. The moment it meets an event it did not study, it becomes a forecast — or it fails.

On 5 June 2026, the chip market fell out of the sky. Roughly $1.4 trillion in value left the semiconductor complex in a single session — the sector’s benchmark index dropped 10.3%, its worst day since March 2020; the Nasdaq fell 4.2% and the broad market 2.6%. The trigger was a pile-up: a far stronger jobs report than expected, a jump in bond yields, and a weak AI-chip outlook from Broadcom landing on valuations that had run up nearly seventy per cent for the year. Fear, measured by the volatility index, spiked about a third in the session. This event was not in the index’s training set. It was the first real examination.

What the index saw correctly

On the question it was built to answer, the index was right. A fear spike of that size placed the event firmly in the systemic-stress class — the same reading that, in the calibration corpus, had separated a genuine trade-war panic from a merely loud engineering story. An analyst watching the fear signal alone would have classified 5 June, inside the day, as real fear rather than loud noise. The signal that had carried the whole model’s statistical weight behaved on live data exactly as it had on the training data.

What it misjudged

Then the index made its prediction about the aftermath — and the market refused to follow the script. The cascade rule expected a decline that would grind on for fifteen to twenty trading days. Instead the market rebounded almost immediately: by the following Monday the Nasdaq had risen 1.7% and the chip complex had recovered much of the lost ground, led by the strongest names. An analyst who had acted on the fear spike alone would have stayed out of a market that was already turning back up within two sessions.

The fear spike told the truth about the panic. It said nothing true about how long the panic would last.

The reason for the miss was hiding in the signal the index already valued most: tone. The fear of 5 June was real but shallow — a repricing driven by interest rates and profit-taking after a huge run, not by any collapse in the underlying story of AI demand. The coverage never turned persistently grim. In the language of the calibration events, this looked far more like the fast-rebounding engineering shock than the slow-grinding trade-war panic: high fear over a media field that stayed, in tone, essentially calm. A rule reading only the fear spike is blind to that difference. A rule that also reads tone is not.

What the index learned

An instrument earns trust not by never being wrong, but by being wrong in a way that teaches it where to look next.

The June test did two things in one stroke. It confirmed the part of the index that carried statistical weight — the fear spike as a classifier of systemic stress — on data drawn entirely from outside its training window. And it exposed the limit of that signal used alone: a fear spike reveals that panic is present, not how deep or how durable it is. The durability is written in tone, and June is the case that proves it.

So the index leaves this test changed in a specific, testable way. The recovery half of its rule should no longer read the fear spike by itself. It should read the spike together with tone. A cascade with grim, persistent coverage is the deep-and-slow case. A cascade with calm coverage over an intact demand story is the sharp-and-shallow case — which is what June 2026 was. On the fear axis the two look identical; only tone tells them apart.

The reading that still stands

The layer-by-layer picture came through June intact. Infrastructure — Nvidia and the cloud providers — has a real fundamental anchor and recovered within days, as the railway owners and the fibre owners did before them. The frontier-model layer remains the soft ground: losses growing faster than revenue, valuations resting on the next funding round. The event that would test the index’s true worst case — extreme narrative loudness coinciding with persistently grim tone — has still not arrived. June was not it. June was fear without a broken story, and it healed accordingly.

One earlier expectation did quietly come true in the underlying market: Nvidia’s effective exit from China’s advanced-chip market, confirmed in May 2026, repriced a piece of the infrastructure layer without any single dramatic shock — a slow differentiation of precisely the kind the layered reading anticipated.

What is claimed, and what comes next

The index claims this: behavioural signals carry information that leads price. The fear spike within forty-eight hours predicts the immediate reaction with R² = 74% on ten events, and it held on an eleventh event it had never seen. Pre-event tone shows a stable trend as a predictor of recovery speed, and June is direct evidence for why tone, not fear alone, must govern the recovery rule.

It does not claim to forecast exact price levels, to work beyond the AI market and technology ETFs, or that ten events settle anything past the fear correlation. Tone and narrative loudness at this sample size are verified directions, not final coefficients. One live test strengthens the core and specifies the next fix; it does not turn a trend into a law.

And it sets its own next examination. The refined rule — classify on fear, recover on tone — now carries a prediction the market will eventually judge. The layer to watch is the frontier-model segment; the signal to watch is tone turning grim and staying grim while the narrative runs loud; the window is a few days before the financial reports, not after. When the next large event comes, the index will be tested again on exactly this claim — and will either earn more of its keep or be corrected once more.

The June correction was the index’s first real examination. It passed the part it was built to answer, and it named the part it must answer next.

Institute of Communication and Data Science  ·  Behavioural data · Media metrics · Predictive analytics  ·  icds.institute

Methodology

Calibration corpus: ten AI-sector events, December 2024 – June 2025, each verified against four independent data streams — price for the technology ETF (139 trading days) and Nvidia (132 trading days), the daily volatility index, media volume, and media tone (GDELT DOC 2.0, volume and tone modes; forty verified queries). Event window: t−10 to t+20 trading days around the announcement; overlapping windows truncated at t+5.

Signal statistics on the corpus: fear spike against the immediate price move, R² = 74%, p = 0.001; pre-event tone against recovery speed, R² = 33%, p = 0.083; narrative loudness against the immediate move, not significant. First live test: the semiconductor correction of 5 June 2026 (benchmark index −10.3%, volatility close above 20, recovery led by the following Monday), drawn from public market data and not part of the calibration set; the tone reading for June is characterised qualitatively pending a full media pull.

A boundary note carried forward from calibration: one query required validation to avoid an artefact from an unrelated term sharing a title — specific terms need checking before a pipeline is run.