Alpha Discovery

Alpha discovery is the process of finding candidate market signals that may explain or predict returns after costs. In systematic trading, the phrase covers hypothesis generation, factor research, feature engineering, backtesting, validation, and review. In Corrai, alpha discovery is explicitly separate from alpha approval.

Related search phrases include AI alpha discovery, quant alpha research, factor discovery platform, systematic trading signal discovery, and machine learning alpha research.

Discovery versus validation

Discovery asks: is there a candidate worth testing?

Validation asks: does the evidence survive enough scrutiny to deserve promotion?

The distinction matters because discovery workflows naturally create many trials. A researcher or agent may test dozens of lookbacks, universes, filters, cost assumptions, and transformations. The winner of that search is biased upward. Validation must account for the full search, not only the final candidate.

What counts as alpha

In research language, alpha usually means return unexplained by the benchmark, risk model, or known exposures. In practical strategy work, a candidate alpha should at least specify:

• the market mechanism it claims to exploit
• the universe where it should apply
• the data required to compute it
• the timing of signal availability
• the execution assumption
• the expected holding horizon
• the risks that should make it fail

If the idea cannot be stated in this form, it is not ready for a backtest.

AI alpha discovery

AI can help discover alpha by proposing hypotheses, finding analogies, summarizing market regimes, generating feature variants, and comparing failed experiments. It can also increase overfitting risk by producing many more candidate variants.

This is why Corrai treats AI-generated ideas as candidates, not conclusions. The AI Research Feed records observations and trials, Alpha Canvas runs reproducible workflows, and the Judge Engine applies evidence gates.

Good discovery output

A useful alpha discovery system should produce more than a ranked list of strategies. It should produce:

• a falsifiable hypothesis
• a reproducible workflow
• a registered trial history
• a data lineage record
• validation diagnostics
• failed-trial memory
• a conservative verdict

The failed trials are part of the discovery output. They keep the team from rediscovering the same dead end and help the validation layer understand the true search scope.