1. Deployment went free
The slowest step in retail quantitative trading was writing and wiring the code. It is no longer a step. None of it is illegal, most is well built. The question is what the button ships.
| Platform | What it ships | Reach and dates |
|---|---|---|
| Robinhood | A customer’s own large language model agent, connected through an official machine interface, trading a ring-fenced account | May 2026; the largest retail-facing brokerage to try it |
| Composer, now SoFi | A plain-English prompt turned into a backtested strategy in under a minute | Over $200 million in daily trading volume and more than 3,000 community-built strategies, October 2025; acquired June 2026 and placed in front of 14.7 million members, deployable within seconds |
| QuantConnect | An agentic developer tool that writes the strategy code itself | A community it counts at over half a million |
| Horizon | A plain-English idea to a live brokerage deployment; its own footer says it is not a broker, a dealer or an adviser | Launched July 2026; about ninety seconds end to end |
| Combined | Deployment as a button | Tens of millions of accounts |
2. What free deployment ships
This desk runs the same loop with one difference: a gate that rises with every look, stresses doubled costs and halved returns, and admits no strategy that skipped the queue. Pairs trading is the worked example, the class these platforms generate on request. Its information is real and too small by an order of magnitude: an edge below its trading cost. Buyers see the raw backtest, never the deflated one.
| Measure | Value |
|---|---|
| Hypotheses filed by the engine | 402 |
| Adjudicated by the gates | 251 |
| Survivors kept | 2 |
| Died in engineering before any statistic ran | 146 |
| Pending | 5 |
| Pairs-trading variants tested and closed, 11 August 2026 | 5, each with the pipeline’s standard one-day execution delay |
| Information coefficient of the pairs signal | Real, confirmed by a formal test, and economically negligible |
| Widely viewed replication of the classic distance rules, reported annual return | 8%, with honest inclusions |
| That replication’s best portfolio once execution waits a single day | Over 7% falls to under 2% |
| Pair-level convergence in this book, gross | Roughly 0.3% a year |
| Cost of trading that book | 0.7% a year |
| Time to find such an edge, and to deploy it | Ninety seconds, ninety-one |
3. Who funds whom
Prediction markets keep a public ledger, so the skill test can be run. The ninety-second pipeline arms the losing majority first: speed without a selection discipline is noise, just faster. That playbook is automatable too, so agents arm both sides. But one side needs a discipline the platforms do not ship, the other only volume, which they manufacture. The wave does not democratise alpha. It industrialises the noise that pays for it.
4. The herd becomes a mechanism
Agents on the same few foundation models, prompted alike and reading the same signals, write correlated books. The herd stops being an accident and becomes an attack surface. The branching ratio would measure the condition. Nobody has estimated it on this wave.
| Claim | What the record says |
|---|---|
| Published strategies decay | Returns fall 26% out of sample and 58% after publication, with post-publication crowding most of the difference (McLean and Pontiff, Journal of Finance, 2016) |
| The uninformed fund the informed | Over the full Polymarket ledger, 3% of accounts hold persistent, statistically real skill; the other 97% trade with no measurable edge, and their aggregate losses are precisely what funds the 3%. Luck does not persist: most accounts that look brilliant in one period flip to losses in the next. The skilled playbook is reacting first to public information, enforcing the law of one price, and betting against behavioural bias (Gomez-Cram, Guo, Jensen and Kung, London Business School and Yale, 2026; the theory is Grossman and Stiglitz) |
| Common models are procyclical | Positions driven by common models and data create procyclical, fire-sale dynamics (Bank of England, 2025) |
| Homogeneous agents can be attacked | “Systemic traps”: fabricated signals designed to synchronise homogeneous agents into congestion and cascade, so a crafted headline that triggers a synchronised sell-off is a published threat category rather than a thought experiment (Franklin, Tomasev, Jacobs, Leibo and Osindero, Google DeepMind, 2026). Simulations already find that large language model agents show less strategy variance than human traders |
| Activity is increasingly self-caused | The branching ratio, the fraction of events caused by other events rather than by news, is estimable with self-exciting point process models and rose towards criticality over the algorithmic era (Filimonov and Sornette) |
| The rulebook is elsewhere | FINRA’s 2026 Regulatory Oversight Report covers firms’ internal generative-artificial-intelligence use, governance and hallucination risk, and says nothing about retail customers wiring autonomous agents to brokerage accounts, five months before Robinhood shipped exactly that |
5. The gap in the record
Regulators are watching a different film. United States alerts on artificial intelligence trading are almost entirely fraud-framed, and crowding, correlated deployment and decay from legitimate mass tools appear nowhere. The academic side asserts the causal chain without measuring it, because deployment counts are private. The open contribution: match platform-declared strategy counts and deployment dates against the later out-of-sample decay of the classes they deploy, as McLean and Pontiff matched publication dates. The data sits inside three or four companies.
6. What this does not establish
No causal measurement is presented here; none exists yet, and the receipts above are surrounding evidence, not proof. The museum’s numbers are one engine, one universe, daily bars. The herding results are simulations, not field data. The branching-ratio rise is the literature’s estimate on an earlier era. This museum is itself an artificial intelligence research pipeline. The difference was never the machinery but the bar it must clear, the multiplicity it must confess and the mortality table it must publish. Those take longer than ninety seconds.
Provenance, bs-prov/1.0
- Method
- No new statistics were computed for this paper. Museum figures are recorded store verdicts: 402 filed, 2 survivors; the five-look pairs family of 11 August 2026, information coefficient and cost figures as adjudicated and critic-verified.
- Sources
- Composer volume and strategy counts (company release, October 2025); SoFi acquisition and relaunch (SoFi investor relations, June 2026); Robinhood agentic trading (TechCrunch, May 2026); QuantConnect community count (company site, company-claimed); the Horizon launch claim (trade press, July 2026; deliberately not linked); McLean and Pontiff, Journal of Finance 2016; Gomez-Cram, Guo, Jensen and Kung, working paper, June 2026; Bank of England, Financial Stability in Focus, April 2025; FINRA 2026 Regulatory Oversight Report, December 2025; Franklin, Tomasev, Jacobs, Leibo and Osindero, AI Agent Traps, 2026; Filimonov and Sornette on endogeneity, with the Bacry, Mastromatteo and Muzy survey.
- Related
- Working Paper No. 13 (the deflated bar), No. 22 (who has to trade), No. 25 (what daily bars cannot see), No. 28 (the constraint audit), and the pairs family verdicts of 11 August 2026.