AI, algorithms and prices — when a real shock enters the machine
The theory: the systems that form prices read signals, not causes — and when a physical shock hits, margin rules can turn the move into a cascade. The accompanying study states, case by case, which part is already documented and which part is still theorized.
The verdict, first
The theory is partially correct in its mechanism and not demonstrated in its full form. The central mechanism — systems that read signals rather than causes, and a margin loop that amplifies — is documented case by case. The complete systemic cascade, with modern AI as the decisive actor, has no case at all: it is a conditional scenario, warned about by regulators and not yet played by reality. The nodes and their status are declared in section 08.
How it was done: every claim in the study was checked against its primary source (SEC, CFTC, Bank of England, IMF, FAO, ICCO, independent reviews). Anything that could not be verified against a primary source did not go in. Press estimates are cited as press.
01 · The theory — the author's theory
The automated systems that today form and respond to prices in markets — and the artificial intelligence now being added to them — embed the past: a statistical idea of what things "should" cost. They do not read the physical cause. When a real shock hits (a plague, extreme weather, a strike), the price first moves because of scarcity; then margin and liquidity rules turn that move into a self-amplifying loop — price, margin calls, forced sales or purchases, more price — and spread it to similar products, even those with no physical relation to the shock.
Refinement 1 — orchestration (October 23, 2026)
If the cascade happens, AI will not do it alone: it will be orchestrated by humans, or both. In every documented case there is a human hand upstream (the rule that was set, the position that was taken, the leverage that was chosen). AI would add speed, scale and homogeneity — not a different mechanism.
Refinement 2 — the pattern has already been played (October 23, 2026)
The pattern is not only a possible accident: it has already been deliberately played by humans (GameStop, January 2021, verified against the SEC's official report). It is repeatable in the future by AI systems or by a single human, and extensible to mass participation. In the author's formulation, computing power is the limiting factor of scale. A precision from the laboratory: in the case already played, the brake that was observed was institutional — the clearinghouse's collateral rule — and the practical limit was coordination; computing power limits detection and speed.
Scope — what the theory does NOT claim
It does not claim that AI sets the retail price of physical goods in a store; it claims that the financial reference price is formed in systems that read signals, not causes. It is a theory about markets and finance.
What would refute it
Major physical shocks in already highly automated markets absorbed without price/loss disproportion or margin loops; rising model diversity in successive surveys; event studies showing that, at equal inventory and leverage, more AI does not mean more amplification.
02 · The mechanism, in three layers
Reference prices are not set by a single class of system. Attributing to "AI" what the other two layers did is the most common error in this debate:
1 · Automatic rules and infrastructure — margins, collateral, limits, circuit breakers, liquidation. They do not "think": they execute. This is the layer that has amplified the most in real cases.
2 · Classical algorithms and HFT — they execute orders and provide liquidity by reading price, volume and volatility. They withdraw when volatility exceeds their parameters.
3 · AI / machine learning — still a minority in core market decisions (the Bank of England confirms this in 2026), growing in surveillance, execution and risk management.
The endogenous-risk literature (Daníelsson and Shin) describes the underlying pattern: the danger does not come only from outside; the system generates it itself when everyone reacts the same way to the same signal at the same time.
03 · The documented cases
Flash Crash — May 6, 2010 documented fact
The Dow Jones fell nearly 1,000 points intraday (≈9%) and recovered much of it within minutes; some stocks traded at one cent. The joint SEC/CFTC report concluded that a large automated E-mini futures sell program, executed against volume with no reference to price or time, in an already stressed market, triggered the fall; algorithmic liquidity providers withdrew or traded among themselves at absurd prices. Lesson: nobody "decided" the price; the system read flow, not causes.
Knight Capital — August 1, 2012 documented fact
A software deployment reactivated dormant code (the Power Peg function) and the firm flooded the market with erroneous orders in 154 stocks. Losses above US$460 million in one day; an SEC fine of US$12 million (order 34-70694); the firm survived only through an external rescue. Lesson: the dominant layer was code + deployment, not intelligence.
WTI oil at a negative price — April 20, 2020 documented fact
The May futures contract settled at −US$37.63 per barrel (EIA/CFTC): storage at Cushing was effectively full and the expiring contract obliged holders to take physical barrels. Those who could not take delivery paid to exit. Lesson: here the physical world disciplined the financial one — the contract discovered, the hard way, that it represented a real thing.
Nickel at the LME — March 2022 documented fact
After the invasion of Ukraine, the nickel price more than doubled within hours and exceeded US$100,000/t intraday on March 8. The LME suspended the market and cancelled that day's trades. The independent Oliver Wyman review documented the loop: collapsing liquidity, record margin calls, short covering and higher prices, in a spiral. The large short position the press attributed to Tsingshan (100,000–200,000 tonnes, according to Reuters/Bloomberg) does not appear in the official report, which anonymizes positions: it is cited here as a press estimate, not a verified fact.
Gilts and LDI funds — September/October 2022 documented fact
After the British "mini-budget", gilt yields rose violently and pension LDI funds had to sell gilts to meet collateral calls, pushing yields higher still: the margin→sales→price loop in pure form. Net gilt sales by LDI funds between September 23 and October 14 are estimated at £36 billion (Bank of England, Pinter, Staff Working Paper 1019, 2023). The Bank intervened, buying gilts to prevent a chain collapse of the funds.
Food — 2022 documented fact
The FAO food price index hit its all-time high in March 2022: 159.3 points, +12.6% in one month (cereals +17.1%; wheat +19.7%; vegetable oils +23.2%). The war cut exports from two breadbaskets and export bans (India banned wheat exports in May 2022) amplified the shock. Contagion among similar products ran through real substitution (one oil for another) and trade policy (one country after another closing exports), not through algorithms. With the Black Sea grain corridor and substitution, prices eventually returned to levels close to those before the invasion.
Cocoa — 2023/24 documented fact
The price rose from ~US$2,600/t in early 2023 (ICCO 2023 average: US$3,261/t) to an intraday high of US$11,878 in New York on April 19, 2024 (×4.6) and the all-time record of US$12,931 on December 18, 2024 (×5.0), after a combined production fall in Côte d'Ivoire and Ghana of ~−23% in 2023/24 (ICCO) — the two countries produce about 60% of the world's cocoa — and the largest global deficit in over 60 years. The scarcity was real; the breakdown of futures-market liquidity was the added financial layer. A useful epilogue: in 2026 cocoa rose again by ~80% between February–March (~US$2,850) and July–August (~US$6,480), without repeating the 2024 breakdown — a data point consistent with the conditional form of the theory: a strong shock without the fragilities of 2024, and no cascade.
GameStop — January 2021: the pattern, played documented fact
The orchestration case. From ~US$19.95 (2020 close) to US$483 intraday on January 28, 2021 (+~2,700% in the month), according to the SEC's official report (October 2021). Short interest stood at 109.26% of shares outstanding and 122.97% of the free float (SEC; the ~140% figure that circulated comes from S3 Partners). An official nuance this study does not hide: short covering contributed in windows but was "a small fraction of overall buy volume" — the driver was sentiment buying — and the SEC found no evidence of a gamma squeeze. The fully documented margin loop ran through the clearinghouse: the NSCC made intraday calls of US$6.9 billion on January 27 (including a special charge of US$4.8 billion, waived on the 28th), and brokers restricted buying in response to those calls — not under pressure from funds. Contagion to "similars" with no physical or financial relation was the largest of the episode: KOSS +480%, AMC +301%, NAKD +252%, EXPR +214% at the January 27 closes. Melvin Capital lost ~53% in January (press) and was wound down in May 2022. Reading: the trigger was coordinated human orchestration; the amplification, automatic rules.
04 · The shocks that did NOT cascade — documented fact
A theory that only collects confirmations is useless. Six large physical shocks were absorbed without a financial cascade:
Hurricanes Katrina (2005) and Harvey (2017) — gasoline spiked and fell back within days: high inventories and the coordinated strategic-reserve signal (IEA/SPR).
US drought 2012 — corn rose ~63% to a record and the adjustment was orderly: rationing by price and substitution of origins.
African swine fever in China (2018–21) — a huge fraction of the world's pig herd was lost without a financial spiral: a physical market, with no leveraged derivatives chain.
Grains during COVID (March–April 2020) — initial panic absorbed by ample inventories.
Red Sea (2023–24) — freight tripled and new shipping capacity absorbed it.
Panama Canal drought (2023–24) — administrative rationing and substitute routes, with no price spiral.
The brake pattern: available inventories or capacity + real substitution + the absence of a leveraged derivatives chain. It is the exact mirror of the conditions that did amplify in section 03.
05 · The contrary reading, in its strongest form
Price often works: in 2012 it rationed demand and moved purchases to other origins without a spiral — that is the system doing its job, not failing. The physical world also disciplines the financial one: in WTI 2020 it was the real world (there was nowhere to put the barrel) that crushed the contract. And the negative cases outnumber the cascades: the cascade is the exception, conditioned on prior fragilities, not the system's automatic response to every shock. The theory holds as a conditional risk, not as destiny.
06 · What the regulators say — regulatory warning
IMF — Global Financial Stability Report, October 2024, chapter 3: among the largest risks perceived by participants are herding and concentration; the chapter recommends reviewing margin requirements and circuit-breaker calibration, strengthening monitoring including non-bank entities, mapping interdependencies and coordinating oversight of critical AI providers.
Bank of England / FCA — 2024 AI survey: 75% of firms already use AI and 10% use foundation models; concentration among the top three providers reaches 44% in models, 73% in cloud and 33% in data; 46% of firms declare only a "partial" understanding of the models they use.
Bank of England, Financial Policy Committee (April 2026) — notes, at the same time, little evidence of substantial use of advanced AI in core market decisions: the layer-3 risk is today more warning than fact.
Sarah Breeden (Bank of England) — "Agents of change", ECB Forum in Sintra, June 2026: "If AI agents respond similarly to the same prompts or triggers, they could amplify volatility in stress." The Bank is experimenting with the BIS Innovation Hub and the Bundesbank (Project Logos) on "simulation methods to understand which aspects of agent design could drive herding behaviour", and raises safeguards "analogous to circuit breakers or kill switches that would limit or stop trading market-wide if faulty AI models cause market meltdown".
FSB and BIS — adoption is growing faster than supervisory frameworks; the FSB's follow-up (2025) monitors adoption and its vulnerabilities.
07 · The cascade scenario — scenario, no complete case
Step 0 · Prior fragility: low inventories, a concentrated origin, high leverage, homogeneous models, concentrated providers.
Step 1 · A real physical shock: plague, extreme weather, a strike, a closed route. Step 2 · the price jumps on scarcity. Step 3 · margins rise; collateral calls arrive. Step 4 · forced sales or purchases to meet them. Step 5 · liquidity evaporates: algorithmic providers withdraw. Step 6 · contagion to similar products through substitution, policy or sentiment. Step 7 · chain reactions: export bans, controls, interventions. Step 8 · the financial price, now detached from the physical one, passes into the real economy.
No documented case runs through all eight steps; the cases in section 03 run through stretches of it, and GameStop ran the orchestration stretch.
08 · Declared nodes and their status
In this house, a theory's percentage is never typed in: it is derived from declared nodes or declared theorized.
| Node | Status | Basis |
|---|---|---|
| N1 · Systems read signals (price, volume, volatility, margins), not physical causes | documented | Flash Crash and Knight; regulatory doctrine |
| N2 · Margin → liquidity → price loop | documented | Nickel, gilts/LDI, cocoa, the NSCC in GameStop |
| N3 · Contagion to similar products | documented | By substitution and policy (2022) and by coordinated sentiment (GameStop). NOT documented by "algorithmic confusion" |
| N4 · A human-orchestrated trigger | documented | Once: GameStop 2021 (SEC) |
| N5 · The brakes: inventories/capacity, substitution, no leveraged derivatives chain | documented | Six negative cases (section 04) |
| N6 · Modern AI as a decisive amplifier of a crisis | theorized | No case; warned by regulators (section 06) |
| N7 · A complete systemic cascade from a single physical shock | theorized | Conditional on Step 0 (section 07) |
| N8 · Repetition of the pattern by AI systems or a single human; computing power as the limit of scale | theorized | The author's formulation, October 23, 2026 |
09 · Watchable indicators — the living part
Version 1 is never rewritten; status changes are noted with a date in later notes.
1 · The annual BoE/FCA survey: % of core decisions using advanced AI; top-3 concentration; declared "partial" understanding. 2 · FSB follow-ups: correlation and herding indicators when they exist. 3 · A fragility screen by commodity: stocks-to-use ratios (ICCO, USDA) in concentrated origins (one origin ≥50–60% of supply) and initial margin levels at ICE/CME/LME — high fragility = Step 0 assembled. 4 · An episode log: every commodity spike greater than ×2 in under 6 months is logged with its outcome (margin loop or absorption?). The log decides whether N6 and N7 change status.
10 · Primary sources
- SEC/CFTC — Findings Regarding the Market Events of May 6, 2010.
- SEC — administrative order 34-70694 (Knight Capital), 2013.
- EIA / CFTC — the WTI episode, April 2020 (settlement at −US$37.63).
- Oliver Wyman — independent review of the LME nickel market, 2023.
- Pinter — Bank of England Staff Working Paper 1019, 2023 (LDI sales: £36bn).
- FAO — food price index, March 2022 (159.3).
- ICCO — Quarterly Bulletin of Cocoa Statistics, 2024 (averages, records and 2023/24 production).
- SEC — Staff Report on Equity and Options Market Structure Conditions in Early 2021 (GameStop), October 2021.
- IMF — Global Financial Stability Report, October 2024, chapter 3.
- Bank of England / FCA — AI in financial services survey, 2024.
- Bank of England — Financial Policy Committee record, April 2026.
- Breeden, S. — "Agents of change", Bank of England, June 2026.
- FSB — reports on AI adoption and financial stability, 2024–2025.
Theory and authorship: Juan Carlos Euyoque Aguilera (JCEA). Study: JCEA Labs. Record v1 sealed on October 28, 2026 with the Origin Method — root fingerprint 60ca4723c77aa8d77193e44725acfb06c45350f7f0c6f93071b7a4a2925de9f3. Layer 3 (OpenTimestamps): stamp registered with 3 calendars, Bitcoin anchoring pending at the time of sealing — declared out loud; public verification will show it once complete. License CC BY 4.0.
Mathvsmyth