Agent Alpha Discovery

Agent alpha discovery is the use of AI research agents to search for systematic trading hypotheses. The phrase often appears as AI agents for quant research, multi-agent alpha discovery, LLM quant research workflow, or agentic backtesting platform. In Corrai, those terms do not mean that an agent is allowed to approve a strategy. They mean agents help produce and investigate candidates inside a governed research loop.

The core rule is simple: agents can increase the number of ideas, but they cannot lower the evidence bar.

The agent roles

Corrai models alpha discovery as a collaboration between specialized agents and the researcher.

• Research agent: drafts hypotheses, compares related literature, and turns market observations into testable claims.
• Data agent: checks source coverage, schema changes, point-in-time availability, missing data, and staleness.
• Factor agent: proposes transformations, lookbacks, normalizations, ranking rules, and signal variants.
• Risk agent: inspects costs, turnover, leverage, drawdowns, regime concentration, and leakage risk.
• Judge agent: reads the evidence package and issues a conservative verdict. It does not optimize the strategy.

This separation prevents a common failure mode in AI research systems: the same agent invents the idea, tunes the backtest, explains the result, and approves the output. Corrai keeps generation and judgment structurally separate.

From observation to hypothesis

A useful alpha idea starts as a claim that can be falsified. Examples:

• Funding compression after volatility spikes mean-reverts over a short horizon.
• Cross-sectional momentum in liquid perpetual futures survives after fees and slippage.
• On-chain exchange flow changes lead realized volatility only in high-liquidity regimes.
• Earnings drift signals fail when publication lag and survivorship bias are modeled correctly.

The agent's job is to convert an observation into a research object: allowed universe, required data, label horizon, feature definition, execution assumption, and validation method. This makes the idea testable before the backtest result is known.

The trial ledger is the control surface

Agentic research can produce a large number of variants. That is useful for exploration and dangerous for inference. Every completed run changes the denominator of the search. If the system tests 500 variants but remembers only the best 5, the final Sharpe ratio is not interpretable.

Corrai records trial history so the Judge can apply selection-aware validation. The Deflated Sharpe Ratio needs an honest trial count. PBO needs a matrix of candidate performance. Agent alpha discovery only works if the search history remains attached to the winner.

What the agent can automate

Agents can help with repetitive research tasks:

• generate candidate factor families from a stated market mechanism
• create experiment plans for walk-forward validation
• compare failed trials and identify duplicate dead ends
• flag same-bar fill assumptions, missing fees, or look-ahead leakage
• summarize why a Judge verdict blocked promotion
• maintain research memory across sessions

They should not silently change validation rules after seeing results. They should not hide failed runs. They should not convert a weak result into a production claim.

Long-tail use cases

This workflow is relevant to searches such as AI alpha discovery software, multi-agent quant research platform, LLM trading strategy research, AI backtesting assistant, agentic systematic trading research, and AI factor research workflow. The practical user behind those terms is usually not asking for a black box. They need a faster way to explore ideas while preserving evidence.

For the visual workflow surface where these ideas become experiments, see Alpha Canvas Research Workflows.