Predictive Modelling for Donor and Customer Behaviour

Introduction

Many predictive models claim to forecast customer or donor behaviour, but struggle to influence real decisions.

Scores are produced, yet no one knows how to act on them.
Models perform well in validation but degrade quietly over time.
Predictions explain what might happen, but not why or what to do next.

The problem is rarely algorithm choice.
It’s that predictive modelling is treated as an isolated exercise rather than a behavioural decision system.

Why prediction is important

Predictive models increasingly influence:

• targeting and prioritisation

• retention strategies

• resource allocation

• long term engagement planning

When models are poorly designed:

• stakeholders lose trust

• bias and leakage go unnoticed

• models become brittle as behaviour shifts

• analytics teams spend more time defending outputs than improving them

Well designed behavioural models do the opposite.
They create shared understanding, support action, and adapt as patterns change.

Modelling behaviour, not outcomes

At a programme level, effective predictive modelling starts with a mindset shift.

Instead of asking:

“Can we predict this outcome accurately?”

The better question is:

“What behavioural signals precede this outcome, and how stable are they?”

For donor or customer behaviour, these signals typically fall into three groups:

• Temporal behaviour
  Recency, frequency, trend over time

• Engagement intensity
  Depth of interaction, responsiveness, consistency

• Contextual patterns
  Channel mix, topic exposure, lifecycle stage

modelling flowchart

cycle of data prediction

The goal is robust prediction that survives behavioural change.

Example: designing behaviour first features (Python)

Below is a simplified example of feature construction focused on behavioural stability rather than raw attributes.

import pandas as pd

df = pd.read_csv("customer_activity_features.csv")

features = pd.DataFrame()

features["recency_days"] = df["days_since_last_activity"]
features["activity_trend"] = (
    df["activity_last_30d"] - df["activity_prev_30d"]
)
features["engagement_consistency"] = (
    df["active_months_last_6"] / 6
)
features["avg_value"] = (
    df["total_value"] / df["transaction_count"]
).fillna(0)

target = df["future_engagement_flag"]

None of these features identify who the individual is.
They describe how behaviour is evolving.

This makes models:

• more explainable

• more privacy aware

• more resilient to data change

Modelling with intent, not just performance

from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression

X_train, X_test, y_train, y_test = train_test_split(
    features, target, test_size=0.2, random_state=42
)

model = LogisticRegression(max_iter=1000)
model.fit(X_train, y_train)

Simple models are often preferred at this stage because:

• coefficients are interpretable

• feature impact is transparent

• behaviour can be reasoned about

Advanced modelling is a choice, not a default.

A reusable framework for behavioural predictive modelling

A robust framework for predicting donor or customer behaviour:

1. Define the decision the model supports, not just the outcome

2. Engineer behaviour first, privacy aware features

3. Avoid raw identity or sensitive attributes

4. Validate stability across time windows

5. Interpret drivers, not just scores

6. Monitor behavioural drift post deployment

This framework prioritises actionability and trust over marginal accuracy gains.

Although implementations vary across organisations, these principles apply broadly to most data analytics environments.

Generalised advice for analysts building predictive models

• Treat predictions as inputs to decisions, not answers

• Prefer features that describe change, not state

• Validate models against future behaviour, not past patterns alone

• Expect models to decay and design for revision

• Document what the model should not be used for

Strong predictive models shape conversations, not just rankings.

 Reflection: impact, learning, and application

Predictive modelling for donor or customer behaviour is most powerful when it is grounded in behavioural understanding rather than technical complexity.
Models that explain why behaviour changes are easier to trust, easier to act on, and easier to maintain.

The key learning is that prediction is a design discipline, not just a machine learning task.
By focusing on stable behavioural signals and decision context, analysts can build models that remain valuable even as data, tools, and strategies evolve.

For other analysts, this approach is immediately applicable.
Start by redefining the problem around decisions, engineer behaviour first features, and choose models that can be explained as well as evaluated. Over time, this creates predictive systems that support strategy rather than just reporting.