Learn / Research process
Answer page / research process
Topic cluster / Signal calibration and thresholdsA signal can look statistically strong and still be untradeable when the edge is too small, too slow, too crowded, or too expensive to express under real execution, sizing, and capacity constraints.
What to remember
A strong signal is not automatically a good trade. It becomes untradeable when the expected edge cannot survive the actual cost of expressing it, or when the portfolio cannot hold the position at the size, speed, or frequency the backtest quietly assumed.
That is why a model can be directionally impressive and still fail in production. Prediction quality and tradeability are related, but they are not the same research question.
The common failure mode is that the model is evaluated in a cleaner world than the one the execution stack lives in. A tiny edge may disappear once spread, slippage, funding, borrow, participation limits, or rebalance interaction with other sleeves are included.
Backtests often reward the model for being right without making the strategy pay for how hard it was to be right. That problem is especially severe when a fast continuous signal is turned into many small trades that each look harmless in isolation.
This is where Alphora's separation between research logic, portfolio context, and forward evidence becomes useful. If the signal only works at unrealistic rebalance speed or only at a size the shared portfolio cannot support, the problem should show up in cost-aware validation, the portfolio simulator, and eventually paper tracking instead of being hidden behind one attractive score.