Prediction is useful when it changes the order in which you spend time and money. It becomes dangerous when teams treat the forecast as proof.
In growth work, I use prediction as a prioritization layer. Historical data, audience behavior and model output can tell us which experiments deserve to run first. The experiment still has to earn the result.
Forecast expected value, not certainty
Every test has an upside, probability of success, cost and time-to-learn. A rough expected-value calculation is often enough to improve the queue:
Expected Value = probability of meaningful lift × economic upside − test cost − delay cost.
The numbers do not need false precision. The discipline forces the team to compare opportunities economically instead of politically.
Include learning value
A test can be valuable even when it does not win if it answers an important question. Testing a new audience, price framing or acquisition channel may reduce uncertainty that affects millions of pesos of future spend.
This is why I separate commercial value from learning value. A small test can have low immediate revenue upside and still deserve priority because it resolves a major strategic unknown.
Use priors carefully
Historical conversion rates, creative performance and audience data can create priors. But markets change. Seasonality, competitive pressure, product changes and measurement changes can make old data less representative.
A predictive model should make the assumption visible: what period, what population and what mechanism support the forecast?
Design the smallest credible test
The objective is not the smallest possible spend. It is the lowest-cost experiment capable of changing the decision. Too little sample creates noise and delays learning because the team must repeat the test.
Predefine kill, continue and scale rules
- Kill: downside crosses an agreed threshold.
- Continue: signal is directionally positive but insufficient.
- Scale: the result clears the minimum economic hurdle and remains stable under more spend.
The portfolio view
A healthy growth program has a mix of proven optimizations, adjacent bets and a limited number of high-uncertainty experiments. If every test is safe, the team is probably optimizing locally. If every test is radical, the company is paying too much tuition.