Optimise search results on your site to provide a better shopping experience.
Although our products are easy to use, we offer a wide range of services to help you succeed even more in using our software.
Predictive AI is a specific form of artificial intelligence, different from generative AI. With generative AI – such as ChatGPT – you can create content. With predictive AI, you create predictions based on a probability calculation.
Predictive AI uses historical data, algorithms and machine learning models to predict future outcomes or trends. It analyses patterns in existing datasets and extrapolates them to future scenarios. Let’s explore how that works and what you can get from it.
For marketers, predictive AI is a game changer. Every marketer has the core task of understanding their target group(s), increasing the demand for products or services and ultimately driving more sales. Predictive AI can help you in all of these areas and tasks:
With predictive AI you gain insight into customer loyalty and you can make great churn analyses. With the models that the AI uses, you can make predictions about purchasing habits such as order value and purchase frequency. This gives you the insight you need to split your audience into useful, effective segments.
Predictive AI can be used to improve your A/B testing and to optimise the channels you use. For example, predictive AI can help you choose the moment in the campaign where a series of paid ads would be most effective.
If you have insight into potential future conversions, predictive AI can use that information to classify leads, which can then provide information for meaningful segmentations.
Many organisations want to base their pricing on the demand for products or services. If you can predict how that demand will develop, you can always adjust your prices accordingly at the right time. Predictive AI makes that possible.
Predictive AI is particularly helpful in recommendation systems. Large organisations, in particular, use this. For example, Amazon works with recommendations for the right clothing size depending on customers’ styles and brand choices. This is simply because not every clothing brand uses the same size chart. If you choose a certain piece of clothing as a customer, you will hear from other customers via reviews and feedback whether it is smart to order the item of clothing a size larger or smaller. Customers love this.
But an organisation like Netflix also fully uses these smart forms of personalisation. We would like to give an example of this. It clearly shows what predictive AI makes possible.
Netflix wants a lot of viewers on its platform. Fortunately, Netflix marketers are sitting on a goldmine of data. After all, customer preferences are widely available. Netflix uses all that data in a smart way. Netflix’s recommendation systems are personalised by using predictive AI. But what does that look like in practice?
Netflix personalises movie posters based on previously viewed movies, a focus on certain actors and preferences for certain movie genres:
Netflix tries to attract more customers for this specific film or series with this smart recommendation technique. This is a great example of the power of predictive AI.
It should go without saying that the benefits of predictive AI are only possible if your data is in order. After all, the predictions are made on the basis of historical data. If you do not store price developments and times of purchase in your systems, something like dynamic pricing becomes difficult. And if Netflix personalises movie posters, you have to store how often which customer prefers to watch which type of movie.
The data you need to enable Predictive AI, therefore, depends on your business and mainly on what you want to do with predictive AI.
We want our customers to earn and save money with our software. For that reason, we focus on improving the Customer Lifetime Value (CLV). The data points required for this are: customer_id, transaction_timestamp, invoice_id, item_number and order_type. This data is loaded into a specific statistical model that is frequently used by data scientists to make predictions about order value and transaction frequency.
In addition, we use the Random Forest model. We feed that model not only with the above data, but also with things like email behaviour and website interactions.
Below is a brief explanation of both models. This will give you a good introduction to predictive AI and how it works.
If you want to gain insight into the future development of the CLV of customers, you start with the following historical data: the number of times a customer has bought something from you, the order value of each purchase and the time at which the last purchase took place. This is also known as an RFM model. RFM stands for Recency, Frequency and Monetary Value.
Then the CLV is calculated by multiplying the customer value (based on RFM) by the average duration of a customer relationship. You can calculate the average duration from historical customer behaviour. This calculation does not give you insight into future developments and is deterministic in nature. However, once you know your CLVs, you can start to ask the questions you really want the answers to:
With this knowledge, you can set up great marketing campaigns.
If you want to answer the above three questions, you are making a prediction about possible future behaviour. To do that, you need a mathematical model that calculates that for you. Long story short: the model with the best track record is the BG/NBD model including a gamma distribution of the data (GG model).
BG/NBD + GG-model stands for Beta Geometric / Negative Binominal Distribution + Gamma Gamma Submodel. This sounds very complicated, but what you need to know is that it’s a model for probability calculation. The model is not that old. It was developed by Fader, Hardie and Lee in 2005. A short explanation:
English-speaking marketers and researchers have reduced the complicated abbreviation of the model to ‘Buy Till You Die’. The model calculates the possible amount of transactions in a predetermined period (this is the ‘Buy’ part). For example: how many transactions will take place in the coming month? But also: which customers will buy something in the coming month?
The model is ‘negative binomial’. After all, if you know which customers will make a purchase in the coming month, you also know which customers will not. In addition, you almost certainly have a number of customers who no longer buy anything from you or who have been inactive in your database for months (this is the ‘Till You Die’ part).
When you put all that information about active and inactive customers into the BG/NBD model, you gain insight into which customers are likely to churn, which ones are most likely to make a purchase, and how often a purchase occurs.
Furthermore, the GG model provides you with insight into the average order value you can expect per customer.
Remember, the GG model looks at the total order value that each customer has and then randomly distributes that value over the total number of transactions. This way, the average order value can change over different periods. Finally, a gamma distribution of the average order value over all customers takes place.
Or, to put it more simply, you get a prediction of the average order value per customer by looking mathematically at the average order value for all customers.
As you have read, Spotler’s Predictive AI model is fed with RFM data. We also use the BG/NBD + GG model for CLV predictions.
Organisations not only use the predictive BG/NBD + GG model, but also refine the outcomes by looking at email behaviour, website interactions, responses to seasonal promotions, event attendance, and so on.
It is perfectly conceivable that a customer scores low based on the number of transactions, but still looks around your website a lot and has visited your last event. Every prediction becomes better if this data is also taken into account. You have different AI models for that.
There are several machine learning models that are used in Predictive AI. Think of: Gradient Boosting Machines such as XGBoost, Logistic Regression, Recurrent Neural Networks (RNN) and Random Forest. If we were to explain all these models in one article, it would be roughly the length of a PhD thesis, so we’ll focus on Random Forest because Spotler uses that model itself.
Keep in mind that generative AI models provide a probability calculation as output. If a customer buys exactly one item from you every month, you can expect this customer to repeat this action the following month. The chance is high, but it is not a law. The outcome is not determined.
AI models are probabilistic and not deterministic in nature. Or, more simply, AI models take into account the factor of ‘uncertainty’ in the predictions they make. People do the same thing all day long when they make decisions. Very often, those decisions are based on incomplete and uncertain information.
We even have a nice word for it: intuition. And there is even an equation that says ‘thinking + feeling = knowing’. Now, AI models don’t have feelings, but they are probabilistic. We asked Chi Shing Chang, managing director of SPARQUE.AI, for a quote.
“Suppose you have a pet shop and you want to send visitors a targeted offer. In a deterministic model you say: everyone who visits the category ‘dog’ will receive an email with an offer for their dog. But what if a visitor visits multiple categories? A probabilistic model takes this into account by calculating probabilities. For example: the visitor’s behaviour can be divided into 60% dog, 25% cat, 10% bird and 5% rodent. If you have a nice offer for cat owners and not for dog lovers, you can still send something relevant to a visitor with the aforementioned search behaviour.”
Chi Shing Chang, SPARQUE.AI
Take a look at the following table:
Customer ID | Last purchase | Purchases per year | Average order value | Email opens | Email clicks | Promo |
---|---|---|---|---|---|---|
100 | 13-12-2024 | 10 | 15 | 6 | 4 | winter |
101 | 09-10-2024 | 2 | 70 | 3 | 2 | october |
Suppose that on January 1, 2025, you want to predict which of these two customers – with customer_id 100 and 101 – is most likely to make a new purchase within 30 days. How would you approach this? There is a good chance that you will set up decision trees. Maybe not literally, but at least in your mind. Visually, you can think of the following:
From this simple decision tree, you could predict that all customers with a value of 1 are likely to make a new purchase within 30 days. Of course, this is not guaranteed; such a decision tree can be expanded with multiple decision moments, and datasets based on multiple decision trees can become complex. Before you know it, you have a whole forest of decision trees. The AI model is not called Random Forest for nothing.
But why do you need a whole forest? This has to do with a key part of the AI model: bootstrap aggregation. This needs some explanation.
The term bootstrap aggregation refers to a statistical method that allows you to collect data. To keep things simple, think of a group of musicians playing together in an orchestra. Then only one musician, for example, an oboist, plays his part in the piece of music. It would be difficult to guess the piece of music based on just one instrument. Such a prediction becomes easier (or more accurate) when multiple musicians play their part in the piece of music.
Scenario 1:
You are looking for a new car, and you ask a good friend which car you should buy. He says to you: “Buy the new Ford Explorer.” How big are the chances that you will run to the showroom now?
Scenario 2:
You are looking for a new car, and you ask a good friend a few questions:
After an inventory of all the answers, you conclude that a new Ford Explorer is the best option. Is the chance of purchase in scenario 2 greater or less than in scenario 1? Exactly: much greater. This is, in layman’s terms, how bootstrap aggregation works.
Bootstrap aggregation involves creating subsets from the original dataset. Or, in the car example, sub-questions are created from the main question. Each subset contains the same amount of data as the original set. It is even possible that the different subsets contain the same data. After all, the new Ford Explorer must appear in different subsets.
Such a subset is also called a bootstrap sample. These samples are randomly drawn from the original set. Hence “Random Forest”. This is a powerful technique. You can make statistical predictions without having to rely on assumptions.
Of course, the reliability of this method has been tested. A cool example is a test in which the model had to predict the ground cover of a piece of land in Colorado, USA. The model was fed with data such as the number of hours of sunshine, the presence of water, etc. Random Forest scored 94%.
Predictive AI, like generative AI, has the property that it ‘learns’ or trains itself. The question is of course: what does that look like in an AI model like Random Forest?
Training Predictive AI is done by splitting data into training and test sets. For example, 80% training and 20% testing. The test data is then used to check the training data. To explain:
Suppose you want to predict what the temperature will be in Liverpool on December 28, 2025. How do you do that? One: you know it is December and not high summer. You know that Liverpool is in the UK. And you know today’s temperature. You also know that last year, on December 28, you were still shivering in the Peak District. In short, 5 degrees Celsius is a reasonable estimate.
As humans, we use a lot of knowledge of the world to make predictions. An AI model does not have that knowledge, but it does need it. The automatic human connection between ‘early December’ and ‘colder than high summer’ must be learned by an AI model. An AI model does this by assessing historical data in the training set (for example, temperatures at different times): “1 = yes vs 0 = no” or “true versus false” for the outcome we want: “What is the temperature on 28 December 2024?” The decision trees that ultimately lead to the correct result are then tested with comparable data (test set).
Remember: the AI model has no idea that it is working with ‘temperature’. ‘Worldly’ data (clicks in an email or temperature in Liverpool) are entered as numerical data. If you were to set the AI model to work with the question, “What will the weather be like in Manchester on 28 December 2025?” then the AI model would have to recalculate everything, while we as humans can say, “That won’t be much different from Liverpool.”
In other words, people learn through experience plus knowledge of the world, and a Predictive AI model learns by testing decision trees. Of course, an AI model can calculate with a lot of data, which ultimately makes the prediction better than we as humans can achieve with our worldly knowledge.
The output of the Predictive AI model is a probability calculation. The model predicts for each customer the probability that a purchase will take place within the time frames you have specified. The following could be a possible output of the model:
Customer ID | Chance of buying (30 days) | Chance of buying (60 days) | Chance of buying (90 days) |
---|---|---|---|
100 | 85% | 92% | 98% |
101 | 20% | 50% | 65% |
Probability calculation runs -theoretically- from 1% to 100%. The subdivision in percentages can be specified into deciles (divided into tens). If you score 85% or an 8.5, you’re doing well. In Spotler Activate, we do something similar. Customer 101 comes in the second decile for the chance of ‘buy 30 days’.
You can use predictions to plan personalised actions. You can exclude customers with a high chance of buying from certain promotional actions. Then you will not be spending marketing money on customers who were going to buy anyway.
For example, you send customers with an average chance of buying an email with a targeted product recommendation and give customers with a low chance of a discount. In this way, you target very precisely to achieve the conversion—more revenue for a lower spend!
Predictive AI has four key benefits:
Cost-effectiveness
This should be clear by now with the explanation above. With predictive AI, there is less wastage of marketing budget because you can target your marketing efforts very precisely.
Efficiency improvement
You can put predictive AI to work by automating tasks like segmentation and targeting. But even if you don’t automate these kinds of tasks, the output of predictive AI will give you a lot of information about customer segments. After all, the models are probabilistic. You don’t just have static data on which to base your decisions, but also predictions about possible future behaviour.
More customer engagement
With predictive AI, you can put customers into targeted campaigns and provide a highly personalised experience. If you can accurately predict what your customers need at what point in their customer journey, you increase customer engagement.
Improved decision making
Predictive AI is not only a boost for your database, but for your whole organisation. After all, you use advanced data analysis for strategic marketing choices. Every decision to use or not to use a certain campaign is better substantiated with predictive AI.
With predictive AI, companies can act proactively instead of reactively, leading to more efficient processes and better results. With predictive AI, Spotler customers can save money and make money.