May 6, 2026
Three questions to answer before moving forward with machine learning
Are you interested in using machine learning models? In this article, we review three topics that define the success of these models.
More and more companies are seeing the value of implementing machine learning models to automate complex analyses and improve decision-making.
The conversation quickly turns to tools, vendors, and budgets.
But there are important questions to answer first. The success of a predictive model depends less on the technology itself and more on three conditions that must be met before it can be launched.
1. Is the problem defined with sufficient precision to model it?
Having a clear overall objective is not enough. "I want to predict the risk of fraud or stockouts" is not sufficient to define useful predictive models.
Two questions that help to reach the correct level of definition:
- Is the problem the same for all customers, products, and stores, or are there segments with different dynamics?
For example, a bank that manages consumer, mortgage, and business loan portfolios will need separate models. Or a retail company that sells different products in different cities and neighborhoods will have to create models for each location and product line with similar dynamics.
Trying to capture everything in a single model generally produces poor predictions in all cases. Undesirable biases.
- What specific decision will the model support, over what time horizon, and at what point?
A credit scoring model that supports the initial approval of a loan is different from one that monitors the risk of an existing portfolio. The former works with applicant data at the time of application. The latter uses accumulated behavior over time. They are different models, with different uses and success criteria.
The more specific the question, the more useful the answer the model can provide.
2. Do you have the data to answer the question?
Things get more complicated here.
It's not about having a lot of data, but about having the right data in a usable condition. If it's scattered across multiple systems, inconsistent, or requires manual consolidation, that's the problem that needs to be solved before even thinking about the model.
The flow below only works if the first stage is properly implemented.
3. Can you integrate the prediction where the decision is made?
This aspect is often underestimated.
A model that performs well but isn't connected to the system where it operates changes nothing. The value isn't in the prediction the model generates. It's in the decision that number allows you to make, at the moment it needs to be made.
In the case of the credit scoring example presented above, the model has real value when the probability of default reaches the analyst or the approval system at the moment the application is evaluated. Not in a report that someone reviews days later, when the exposure has already materialized.
If these three elements are well-structured, the conditions are in place for a predictive model to generate real impact.
If there are doubts about any of these components, that's the starting point. It's the work that needs to be done first for the most advanced solutions to work well.
