Data Science
Predictive models that are explainable, in production, integrated in the system where decisions are made
What we build
Models that cover the problems that recur most often with our clients. Not an open catalogue. These are the models we have built, and for each one we know what data is needed, how to develop it, where it tends to fail, and what result is the one that actually matters.
Churn prediction
Anticipate which customers are about to leave and why, to build a segmented retention strategy See case →Customer lifetime value
Estimate how much each customer is worth without needing a transaction history to prove it See case →Purchase propensity
Identify which customers are most likely to buy again, and what product See case →What data to start with
Your available data defines the effectiveness of the model. The more mature the data, the more robust the model. Knowing your case is the first deliverable of the diagnostic.
When the history is too short or the event to predict is not well recorded, the right move is to build that logic first.
No model delivers a black box
Every prediction comes with the signals that explain it, so you can decide what to do with it.
A customer at risk due to price and another due to a drop in activity need different actions. Without the reason, both go into the same campaign.
There is also a formal reason: being able to justify why a customer was added to a list.
What a model predicts — and what it cannot →What "in production" means
A model that lives in a notebook and produces a report is not a model in production. It is a one-off analysis, static, and ages from the next day.
Runs on its own, at the cadence the business decides.
Daily, weekly, monthly, or triggered by a specific event. No one needs to remember to run it.
Read more →Delivers results where your team already works.
As an additional field to segment or filter on, or as input for other decision systems. No new tools, no one moving a file.
Read more →Prevents losing value over time.
Models degrade because the world they learned stops resembling the one they face. Anticipating this allows triggering retraining so it never loses value.
Read more →Makes it easy to understand and adjust.
With the written procedure for how it runs, what to review, and what to do when something goes out of range.
Read more →Where the model ends and your judgement begins
Models estimate the probability that an event will occur. The decision of what to do with that remains yours, because it depends on each customer's margin, the cost of each action, team capacity, and commercial strategy. That is your territory, not the model's.
- Estimates probabilities
- Calculates expected values
- Ranks priorities
- Who to call
- What offer to make
- Which campaign to launch
That boundary can be moved, and it is a separate project: a decision layer that combines the results of several models with your business rules and proposes the action. See the full case →
After delivery
A model in production does not stay still. The behaviour it learned changes, and over time that requires adjustments: retraining, reviewing the signals it uses, sometimes a partial rebuild.
Not sure where to start or what makes sense to build first?
Strategic Advisory →Frequently asked questions
How much data is needed?
Less than usually assumed, and not measured in rows. To predict whether something will happen, what matters is how many times it has already happened: one hundred thousand customers with two hundred cancellations give less than twenty thousand with two thousand. To estimate a quantity, how many complete cycles exist. To segment, that there is variety of behaviour. Your case comes out of the diagnostic.
Our data is messy and disorganised. Does it still work?
Almost always yes. Empty fields, duplicates, and criteria that changed mid-way are normal, and cleaning them up is part of the work. What blocks is not mess but absence: if something was never recorded, it cannot be recovered.
Do we need a data warehouse or a data team?
Neither. What is needed is a place where data from different sources can be stored and joined. If it does not exist, it is set up as part of the project, and today storage is very cheap: the work is in organising the data, not in storing it.
How is this different from what our BI already does?
BI is descriptive: it organises and shows what already happened. A predictive model is prospective: it estimates what has not happened yet, case by case. They are different layers and are used together.