Models that predict, forecast, optimise and segment, with the method chosen to fit the problem and the drivers attached to every answer.
We build statistical and machine learning models that predict what happens next, such as which customers are about to leave, which product they are likely to take, which enquiries convert, how much of a defaulted balance you recover and what a portfolio does over its life. We choose the method to fit the problem, from a regression to a boosted tree or a neural network, and attach the drivers behind every prediction.
We forecast the quantities your planning depends on, such as volumes, revenue, losses and demand, across the horizon you plan over, for decisions such as inventory, staffing and cash. We can also build the monitor that reports, whenever a forecast number moves, what drove it and by how much.
We build the models that choose the best setting to maximise the desired outcome, such as pricing where we estimate willingness to pay and price elasticity.
We find the groups hiding in your data and turn them into segments you can act on, using clustering to locate where behaviour genuinely differs rather than where convention says it should. Those segments then carry through to the decisions that follow, such as classifying customers into risk tiers so each one gets the treatment it warrants, targeting campaigns at the people most likely to respond, assigning the contact strategy that suits each group, and building separate models where a single model across the whole book would hide the differences that matter.
We build an independent version of your model from the raw data and compare the results.
The bottleneck is rarely the fitting. At Gini we found it sits in the documentation, the testing and the review cycle, so that is what we automated. We build scorecards and probability of default, loss given default and exposure at default models in weeks rather than months, with the model document generated from the code as the model is built rather than written six months afterwards.
Because the acceleration comes from removing repeated manual work rather than from skipping steps. Data preparation, variable screening, diagnostics, test packs and documentation are generated by tooling Gini maintains and reuses. That leaves the practitioner the judgement calls: segmentation, the treatment of outliers, the choice of method. The evidence trail a review sees is longer than a hand-built model produces, because the tooling records every step it took.
At Gini we choose the method to fit the problem, from a regression through to a boosted tree or a neural network, and we attach the drivers behind every prediction. Where a decision affects a customer and has to be explainable, that constrains the method, and we say so before building rather than after.
That is the usual arrangement and the better one. Your analysts know the portfolio and the systems, and they are the people who will maintain the model, so they build alongside Gini’s consultants and hold the reasoning by the time we leave. Knowledge transfer is easier to achieve during the build than to bolt on at handover.
