How Product Analysts Build Churn Prediction Neural Nets
For product data analysts · Based on Ko Romeo Deep Learning Neural Network Builder
// TL;DR
This guide shows product data analysts how to translate a business question — like 'will this customer churn?' — into a working neural network. You'll learn to identify whether your output is binary, continuous, or categorical, build and standardise an input layer from your feature set, choose the right activation functions, and train the model with gradient descent and back propagation. Use it when you need a model that discovers feature combinations your dashboards miss, and when you must explain to stakeholders why the network flags certain customers.
Why use a neural network instead of hard-coded rules?
As a product analyst, you could hand-write churn rules: if logins dropped and support tickets rose, flag as at-risk. But that's the hard-coded paradigm — you specify every rule. A neural network belongs to the other paradigm: you build the architecture, supply labelled examples of customers who churned and stayed, and let the network discover the rules by adjusting weights. It will surface feature combinations you'd never think to write down.
Each hidden neuron manufactures a brand-new derived attribute. One might combine account age, balance, and product count to detect a specific at-risk profile. This automatic feature engineering is deep learning's core advantage over the SQL segments and threshold rules you're used to.
How do you frame a churn problem for a neural network?
Start with the prediction goal. Churn is binary — will the customer exit or stay? That means a single output neuron using a Sigmoid function (interpreted as probability of churn) or a Threshold function (strict 0 or 1). Decide this first, because it defines your output activation.
Then list every feature you have for one customer: tenure, balance, age, number of products, activity level. Each becomes one input layer neuron. Standardise them all — a balance in the thousands and a tenure in single digits are on wildly different scales, and un-standardised inputs will wreck training. Standardise to mean 0 and variance 1, or normalise to a 0–1 range.
Which architecture and activation functions work for classification?
Add one or more hidden layers with the Rectifier function — the most popular choice for hidden layers. Pair that with a Sigmoid output for a clean probability of churn. This Rectifier-plus-Sigmoid combination handles the vast majority of binary classification problems.
Don't try to pre-specify what each hidden neuron detects. The whole value is that the network finds patterns autonomously across your entire customer base.
How does the model learn from your data?
Initialise weights and define the cost function C = ½(ŷ - y)², where y is whether the customer actually churned and ŷ is the network's prediction. The network runs an epoch — one full pass through every customer row — computes total cost, then back-propagates to update weights. All customers share the same single network; there's no separate model per row.
Because you likely have tens of thousands of customers, use Stochastic Gradient Descent. Batch gradient descent updates weights only after processing every row and becomes painfully slow at scale. SGD updates after each individual customer, trains faster, and helps avoid getting stuck in local minima. Never attempt brute-force optimisation — the curse of dimensionality makes it impossible.
How do you explain the model to stakeholders?
After training, inspect the hidden neurons. Near-zero weights mean a feature doesn't matter to that neuron; non-zero weights reveal the at-risk combination it learned. This lets you tell a product manager, "the model isn't just looking at declining logins — it found that low balance combined with short tenure and few products is the strongest churn signal." That's an insight your existing dashboards likely missed.
Next step: Export a labelled dataset of past churned and retained customers, confirm you have enough volume and processing power (the two non-negotiable ingredients), standardise your features, and prototype a single-hidden-layer Rectifier-plus-Sigmoid network. Validate its churn predictions against a holdout period before shipping it into your retention workflow.
// FREQUENTLY ASKED QUESTIONS
How much data do I need to train a churn neural network?
There's no universal minimum, but deep learning has two non-negotiable ingredients: large data volume and sufficient processing power. Audit both before starting. With too little data, your network will underperform — exactly as deep learning did in the 1980s before the technology was ready. Thousands of labelled churned-and-retained customers is a reasonable starting point for a binary classifier.
Should churn output use a Sigmoid or Threshold function?
Use Sigmoid if you want a probability of churn between 0 and 1 — ideal for ranking customers by risk and setting your own intervention threshold. Use the Threshold function only if you need a strict binary 0 or 1 with no nuance. Most product analysts prefer Sigmoid because a probability score is far more actionable for retention campaigns.
Can a neural network tell me why a customer is likely to churn?
Partly. By inspecting hidden neuron weights after training, you can see which feature combinations the network learned to associate with churn — near-zero weights signal irrelevance, non-zero weights signal importance. This won't give a per-customer explanation like a decision tree, but it reveals the driving feature combinations, often ones your dashboards never surfaced.