A hypothesis from a model
A candidate is built on a Labs model: the liquidation map with exposure history, OI, order flow, options regimes or Monte Carlo.
Method
Finding a condition that once worked is easy. It becomes an alpha only if it holds on data it has never seen.
Fixed before the test
X, Y and the horizon are fixed before testing. The condition threshold is never tuned on the data its result is measured on.
Out of sample only
The hit rate is measured walk-forward on data after the development period. Overlapping horizons are collapsed so one move is not counted many times.
Against the base rate
P(Y|X) is compared with P(Y) on the same symbols and period. It needs an edge of 5 pp or more and a 95% lower bound above the base rate.
Corrected for search
The more candidates are tried, the stricter the significance bar: the number of attempts is accounted for, as in the deflated Sharpe ratio.
Review
Every two weeks an alpha is recomputed on new data. A weakening one moves to watch; one that loses its edge is retired.