The Un-Blinding: Numerai Names Its Stocks as Finance's Agents Get Real Jobs

The dataset gets names

Numerai spent a decade as finance's great anonymized experiment: thousands of data scientists building models on obfuscated features, never knowing which row was Apple and which was a zinc miner. On October 16, 2026, the blindfold came off. Numerai is revealing the stock identifiers for every row in its dataset, inside a new Unity dataset of 1,040 features, and Numerai and Signals will merge into a single tournament with one Meta Models on January 1, 2027. m:…4941

The un-blinding is not about nostalgia for tickers. It is about agents. The Unity dataset allows agents to perform tasks like churn measurement, sector modeling, and time-series forecasting — the kind of work that requires knowing what the thing is, not just its statistical shadow. The performance claim is striking: Unity doubles the NCORR of the Signals benchmark model over the validation period. And the economics of agent-driven research have collapsed to near zero: Numerai recorded 77,000 MCP research sessions since July 2026, and a researcher built three new signals using Claude Opus 5.5 for a cost of $33. Thirty-three dollars. The marginal cost of a quant idea is now a rounding error. m:…4941

The agents get jobs

While Numerai was un-blinding its data, Hong Kong's banks were giving their agents performance reviews. In the HKMA's second GenA.I. Sandbox cohort, over 60% of use cases involved AI agents — and the results read like a procurement justification. A market-risk agent shortened the cycle of case-specific risk analysis by 95%. A bank reduced customer-onboarding time by 40% using advanced fraud detection enabled by AI. The cohort achieved sophisticated outputs with around 30% less computing power than the first. These are not demos; they are cost lines and cycle times. m:…9973

The supervision is moving as fast as the deployment. The HKMA's expectations are shifting from human-in-the-loop to human-on-the-loop oversight for AI applications — the human moves from the steering wheel to the supervisor's chair — and the authority issued new supervisory guidance on AI adoption aligned with Financial Stability Board standards. The message to bank executives is unambiguous: the agents are in the building, doing real risk work, and the governance model is being written around them in real time. m:…9973

The tail gets measured

And then, the counterweight. The Bank of Canada published a working paper on AI and financial tail risks that applies Expected Shortfall and skewness decomposition to the question every risk officer is quietly asking: what happens when everyone's agents trade the same way at the same time? The paper distinguishes between AI-sector, macroeconomic, systemic, and financial scenarios — an attempt to isolate the AI-specific contribution to extreme outcomes from the ordinary background noise of markets. m:…5869

The paper analyzes AI exposure across five markets, and its existence is the signal. A central bank is now formally modeling the possibility that the same technology cutting risk-analysis cycles by 95% could concentrate risk rather than disperse it. Put the three acts together and the week's shape is clear: the data gets names so the agents can do real work, the agents get real jobs with real cycle-time gains, and the supervisors start pricing the correlated tail. The $33 signal is the exciting part of this story. The Expected Shortfall is the part that will matter. m:…5869