Predictive analytics for marketers, built on your own data
Predictive analytics for marketers ranks customers and leads by what they are likely to do next: buy again, churn, upgrade, answer a campaign or grow in value. DT Nova builds those scores from your CRM, orders, website and app data, tests them against what really happened, and feeds them into the email, ads and sales tools your team already uses. Data stays in the EU as standard, the models are explainable, and the code is yours.
- You see the result before it goes live
- Fixed price for the agreed scope, in writing before work starts
- Source code, domain and accounts in your company's name
The predictions marketers use most
| Prediction | Question it answers | What changes |
|---|---|---|
| Churn risk | Who is about to stop buying or cancel? | Retention offers and contact go to the right customers |
| Purchase propensity | Who is likely to buy in the next weeks? | Campaigns, ad audiences and frequency |
| Customer lifetime value | Which new customers will be worth most? | How much to spend to win and keep each segment |
| Lead scoring | Which leads are worth sales time now? | Order of follow-up and handover rules |
| Next best offer | Which product or plan fits this customer? | Recommendations in email, app and site |
Each of these is a ranking, not a certainty. The value comes from treating the top and the bottom of the list differently: saving the retention budget for customers who would otherwise leave, sending fewer messages to people who will buy anyway, and giving sales the leads most likely to close. Retail and e-commerce forecasting of demand and stock is a related but different job, covered in our guide to predictive analytics for retail.
Start with what your tools already predict
Before building anything, check what you already have. Google Analytics 4 has predictive metrics (read 11 November 2026): purchase probability, churn probability, predicted revenue and in-app purchase probability. Google's help page sets the conditions: at least 1,000 returning users who met the predictive condition and 1,000 who did not, over a seven-day period within the last 28 days, with model quality sustained over time; purchase predictions also need purchase events with value and currency. Many CRM and email platforms offer similar scores.
These are worth using when they qualify. Their limits are just as clear: smaller sites do not reach the thresholds, the model is a black box you cannot tune, and it only sees what that one tool sees. A customer who emails support three times, returns two orders and stops opening newsletters looks the same to web analytics as a happy one. Our own models earn their place when the signals that matter sit across several systems.
Data first, model second
Most of the work in predictive marketing is joining the data correctly. We connect the CRM, the order or subscription system, email and ad platforms, web and app analytics and, where relevant, support tickets, and match them to one customer record. Then each training example is built only from what was known at the time: a churn model trained on data that already includes the cancellation date predicts the past perfectly and the future badly. This leakage is the most common reason predictive projects disappoint, and we test for it explicitly.
We also check early whether the data will carry a model at all. If the outcome is rare, the history is short or the key events are not recorded, we say so and propose what to collect first, often a few events in the app or site and a cleaner link between orders and customers. Our data analytics consulting covers that groundwork, and the types of analytics guide explains where prediction sits beside reporting and diagnosis.
Models a marketer can explain
We start with well-understood methods: regression and gradient-boosted trees for propensity and churn, and established customer value models for lifetime value. They are accurate enough for marketing decisions, fast to retrain and, most importantly, explainable. Each score comes with the main reasons behind it, such as fewer orders in the last quarter, a price plan change or no app use for three weeks, so the team can write the right message and trust the list. More complex models are used only when they measurably beat the simple ones on held-out data.
Proving it works
A model that ranks well on past data still has to change results. We test in two steps. First on history the model never saw, to check that the customers it ranks highest really did churn or buy more often. Then live, with a random control group that is treated as before: the difference in retention, conversion or revenue between the scored group and the control group is the measured effect. Without a control group, a campaign aimed at likely buyers always looks good, because those people would have bought anyway.
Putting scores where the team works
A score in a spreadsheet changes nothing. We write scores back to the CRM and the email or marketing automation platform as fields and segments, send audiences to ad platforms through their official interfaces, and show the trend in a simple dashboard. Retraining runs on a schedule, and a drift check warns when the model's accuracy falls. If your marketing automation itself needs building or tidying first, see marketing automation for small business; for the CRM, see CRM development.
Profiling and the GDPR
Scoring customers for marketing is profiling, and the GDPR sets the rules. Article 21(2) (read 11 November 2026) gives people the right to object at any time to processing of their data for direct marketing, which includes profiling related to it, and Article 21(3) says that after an objection the data shall no longer be processed for those purposes. Our pipelines carry that flag through to every model and every exported audience. Article 22 restricts decisions based solely on automated processing that produce legal effects or similarly significantly affect a person. Choosing who receives a newsletter rarely reaches that level; personalised prices, credit limits or refusing service can, and those are designed with human review and their own legal assessment. In the EU, a price personalised on the basis of automated decision-making must also be disclosed to the consumer, as our AI dynamic pricing page explains.
Two more rules shape the data. Website and app tracking for marketing needs consent where the ePrivacy rules require it, so models must work with the share of users who consented. And special category data, such as health or beliefs inferred from behaviour, is kept out of marketing models entirely.
How a project runs
- One decision. We pick the marketing decision the score should change, such as retention offers or lead order, and the number that will show it worked.
- Data check. Sources, history, outcome counts and consent, with a written go or no-go.
- First model. Built, tested on held-out history and explained to the team.
- Live test. Scores in your tools with a control group for a set period.
- Run. Retraining, drift checks and a monthly look at the measured effect.
You own it
The data pipeline, the models, the training code and the documentation are yours, in your repository and your cloud account, hosted in an EU region as standard. Scores are written into your own systems, so nothing depends on keeping us as the supplier.
Questions we get
What can predictive analytics do for marketers?
Rank customers and leads by what they are likely to do next: buy, churn, upgrade, respond to an offer or become valuable over time. Those rankings decide who gets which message, which leads sales sees first, where retention budget goes and how large a segment is worth targeting. The value comes from acting differently on the top and bottom of the list.
Is the predictive audience feature in Google Analytics enough?
Often it is a good start. GA4 offers purchase probability, churn probability and predicted revenue, but only when the property has enough examples: Google's help page says at least 1,000 returning users who met the condition and 1,000 who did not, over a seven-day period within the last 28 days, and model quality must hold over time. Smaller sites do not qualify, and GA4 only sees website and app behaviour, not your CRM, orders or support history.
How much data do we need for our own model?
Enough past examples of the outcome you want to predict, with the information you had before it happened. For churn or repeat purchase that usually means at least a year of customer history and several hundred cases of the outcome. We check this first and tell you if the data will not support a reliable model, and what to collect so it will.
Is predictive marketing allowed under the GDPR?
Yes, with care. Scoring customers for marketing is profiling. GDPR Article 21(2) gives people the right to object at any time to processing for direct marketing, including profiling related to it, and after an objection the data may no longer be used for that purpose. Article 22 restricts decisions based solely on automated processing with legal or similarly significant effects. Marketing scores rarely reach that level, but price or credit decisions can, and we treat those separately.
How do we know the model works?
We hold back a group. Before launch the model is tested on past data it never saw; after launch a random control group is treated as before, so the difference in sales, retention or conversion between the scored group and the control group is measured, not assumed.
Get a written quote
Tell us in a few lines what you need. We reply within one working day with a written quote: scope, timeline and a fixed price or an hour estimate. Asking is free and commits you to nothing.