How to Utilise Your Customer Data to Make Predictive Models

 

Introduction

If you’re looking for the perfect tool to help you understand how your customers are going to behave before they start doing it, then you’re in need of predictive analytics. It may seem too good to be true, can you really stay one step ahead of your customers allowing you to take important actions at the right point in time? The reality is that you can, and you probably already have the available data to get you started.

The world of business is becoming ever more data-driven as the availability of data explodes, but it's knowing what to do with that data that will get you ahead. In the case of analytics and predictive models, it’s using your past and present data to help you build better picture of what is coming and using this to your advantage to make connections with your customers by knowing what they require next and identifying the new trends as they arise.

How it works

Predictive analytics are a suite of processes that connect the actions of the past to a model of the future. It applies analysis, statistics and machine learning techniques to data to generate predictions and forecasts that can drive your business decisions.

Predictive analytics begins and ends with data, and is driven by clear goals and strategy. The analytical process will identify patterns and trends in your existing data and use them to make predictions. For example, given data about leads who did and did not convert in the past, a predictive model will identify traits that are common among both groups, so that in the future you can predict whether a given lead will convert or not.

Before you begin your analysis, you should know what it is that you want to predict. The goal you choose will influence your modelling techniques and your data requirements. For example, there are two types of predictive model: classification models and regression models. The former predicts a label (it answers the question of, "what is it?") and the latter predicts a quantity (it answers the question of, "how much?").

Once you have a clear goal in mind, there are five key stages to developing and implementing your predictive analytics models:

1 Data collection

Your data can come from a variety of sources and combinations of sources, including customer and sales databases, web analytics, and surveys. You might already have this data available to you, but if not, try to let your goals drive the data collection process. Are you trying to identify market segments who are most likely to buy? If so, you should be gathering data that will identify these segments from your existing customers. Going forward, collect as much data as you can, because you don't know what will be useful until you start testing your data.

2 Data cleansing and transformation

Data in its raw form is messy, and cleaning it is vital to the success of your predictive model. This will involve removing anomalous points and identifying and addressing areas of missing data. Next your data needs to be transformed into a useable format for the analysis and modelling stage, which will involve combining the data from your various sources into a single, accessible dataset. The workload of the data transformation stage will depend on the volume of data, your goal, and the process of modelling you choose. Your data will need to be split into a training dataset and a testing dataset in order to build your predictive model.

3 Data analysis

Your predictive model will inform you as to what the future will look like, but analysing your data will tell you about the past and present. Inspect your dataset to find patterns and get a picture of where your marketing strategies have been successful.

4 Building a predictive model

Here's the technical bit! At this stage you might need to bring in a little outside help, or scroll past this section to see a list of shiny tools that will do it for you.

There are many different types of predictive model, which can be broadly divided between parametric models and non-parametric models.

Parametric models have a predetermined form (for example, a line of best fit), based on a fixed number of parameters. These models are much simpler than non-parametric models, and require less data to develop, but are constrained by the initial assumption of the form and therefore have a lower accuracy than their non-parametric counterparts. A parametric model is a good choice for a small dataset of limited complexity, because they are simple and fast both to train and interpret.

Non-parametric models do not include any prior assumptions about the relationship between the variables in your data. They require larger datasets and a greater time investment, but the result is a more powerful and flexible predictive model with higher accuracy.

Choose your model based on the size of your dataset, your prediction goals and your available resources, and develop it using a machine learning algorithm, statistics or a curve-fitting tool. Again, which of these you choose will depend on your available resources.

Using a machine learning algorithm to develop your model will require some training, in which you feed complete historical data into the algorithm and tweak the parameters of the model to better fit the data, and some testing, in which you compare your model's predictions to your testing dataset.

Examples of machine learning algorithms

How to utilise your customer data to make predictive models 2

5 Evaluation

The great thing about predictive analytics models is that they evolve and improve over time as more data becomes available. Once you have deployed your predictive model, monitor its predictions against the real-world outcomes to ensure that they are accurate and that they meet your objectives. Determine the ROI of your predictive analytics project by measuring the revenue gains from increased sales, cost reductions from streamlined marketing processes against the overall cost of developing the model. This will help you to direct future analytics.

As is often the case, the answer to the question you set out to answer with predictive analytics might spark more questions. Prepare to be inspired to develop more models on the back of what you learn.

Tools for predictive analytics

There are a multitude of tools available to help you build your predictive model, whether you want to build it yourself from scratch or hide the statistics behind a friendly GUI.

If you want the job to be done for you with minimal customisation on your part, choose one of these pre-built AI solutions to take care of the technical stuff:

  • Tableau is a bells-and-whistles business intelligence solution that features prediction capabilities that are easy to strap to your data.
  • Microsoft CRM Dynamics 365 offers an AI-driven add-on, Sales Insights, which features predictive models for lead and opportunity scoring.
  • Salesforce offers a similar AI driven product, Einstein, which features customisable predictive models.
  • Nudge.ai uses predictive modelling to help you predict the risk of churn for your customers.

If you have some statistical/coding skills:

  • R is an open-source programming language developed for statistics.
  • Python is a more general-purpose programming language that has a wide range of predictive analytics utilities.
  • Microsoft Azure Machine Learning Studio is a browser-based platform that offers user-friendly drag-and-drop machine learning tools. With a bit of wrangling, the results of your Microsoft Azure ML experiments can be published to Power BI for organisation-wide viewing.
  • IBM SPSS Statistics features a set of statistical tools that integrate with a large library of extensions and customisable solutions.
  • MATLAB offer a suite of tools for statistics, curve-fitting and machine learning.

How to utilise your customer data to make predictive models 3

Applications of predictive analysis

Your predictive analytics should start with a goal, and there are many to choose from.

Lead scoring

How sure are you of the potential value of every lead that you come across? By analysing all of your past conversions and non-conversions, a predictive model can give you insights as to which leads are - and are not - likely to convert in the future. This will help you to streamline your marketing and sales budgets by focusing on the most likely candidates.

Segmentation

It's no secret that segmentation makes for more effective marketing processes, but predictive modelling can segment your potential customers by much more than their demographics. A predictive model trained by behavioural data - such as time spent browsing different pages of your website - can make predictions and recommendations based on their psychology, leading to a much clearer view of how you can appeal to them.

Content distribution

Imagine if every time you wrote a compelling whitepaper, it was read by an audience of people who are all likely to take action after reading it. This is the power of predictive models that are used to increase engagement and conversion by targeting content.

Lifetime value prediction

A predictive model that has been trained on a customer database could predict how much money any given customer is likely to spend with your business in their lifetime. This can help you to maximise the ROI on your marketing spend.

Personalised product recommendations

Predicting what a customer is likely to buy allows you to upsell just the right product at the right time.

Ad targeting

The backbone of Google and Facebook's 'lookalike' audience feature, predictive models can generate a list of factors common to a person who is likely to buy from you, allowing you to focus your marketing spend on those most likely to convert.

Demand forecasting

Using predictive analytics to forecast when customers are likely to want your products allows you to stay one step ahead of them and keep your inventory fully stocked only when it needs to be.

Closing thoughts

Few marketers would turn down the opportunity to predict the future. Predictive analytics turns your data into a clear strategy for directing your marketing efforts, streamlining your budgets and showing your offering to the right people. Whether you decide to build your own algorithm or make use of existing AI solutions, all you need to see into the future is a dataset and a question.

 

Written by: Mia Hatton

Mia Hatton is the data science apprentice at Nightingale HQ, the catalyst company in helping businesses adopt AI. With an expert matching algorithm, Nightingale HQ offer the support and the connections to get organisations started on their AI journey. Find out more and register your organisation for free today.