From prediction to decision

Prediction tells you who is at risk. It does not tell you that acting helps.

This is the gap most analytics never crosses. A model identifies a group; someone acts on the group; the numbers move; nobody can say whether the action caused it. The customers you contacted were the ones most likely to respond anyway, so the campaign looks effective whether or not it was.

Nexyron treats the action itself as something to be measured.

Estimating effect, not association

The question is not "how do treated and untreated subjects differ" but "what would have happened to these subjects if we had done nothing". Nexyron estimates that difference, and reports the uplift attributable to acting rather than the correlation of belonging to a group.

That changes the answer you get. Some subjects will respond whatever you do. Some will not respond whatever you do. Some respond only because you acted, and a few react badly to being contacted at all. Only the third group is worth spending on, and it is rarely the group a risk score puts at the top.

Refusing to compare the incomparable

Estimating an effect requires that the subjects you acted on and the ones you did not be similar enough to compare. Often they are not.

Nexyron checks this before it will be confident. Where the two populations do not overlap enough to support a comparison, it says so rather than producing a number that looks authoritative and is not. This is the single most common way that decision analytics misleads people, and declining is the correct behaviour.

Comparing courses of action

Given a budget and a constraint, several policies are usually available: contact the top thousand, contact everyone above a threshold, contact nobody in a segment. These can be compared against the evidence already collected, before committing to any of them.

Closing the loop

Interventions and their outcomes are recorded as part of the model. Which subjects were acted on, when, how, and what followed.

That record is what makes the next question answerable. "Did that work" and "what should we do differently this quarter" stop being matters of opinion and become questions with evidence behind them. It is also the part that compounds: each cycle of acting and measuring makes the next estimate better grounded.