# Machine Learning in your marketing mix strategy: taking the guesswork out

*This blog post has been written with the collaboration of Gonzalo Marín*.

Designing a pricing solution involves shuffling a fundamental piece in the company’s marketing mix puzzle. While the ** price** may look like the most fundamental piece to maximize profit it can rarely be defined independently of the others.

This is why, in this post, we will dive into how Machine Learning systems can be beneficial to automate and optimize other marketing tasks, and how globally they can help improve your company’s pricing strategy.

At the end of the day, it is not enough to define at what price you should sell your items but it is also necessary to define:

- what items to sell
- to whom
- where
- and when to do so

## Machine Learning Models vs. Machine Learning Systems

Let’s begin with an important distinction: ML models vs. ML Systems since both terms are often confused. **Machine Learning models** allow us to **generate predictions** and also allow us to anticipate or determine how something will look like given certain conditions or variables. These models learn from experience, gathered from past data.

We could build a Machine Learning model to predict:

- Customer Segmentation
- Customer Expected LifeTime Value
- SKU Expected Demand
- Items categorization
- Customer propensity to different actions, as: trying a new product, category expansion, propensity to buy more, to churn, to engage, to change shopping habits, and more.

The **machine learning model** per se does not tell you what are the best actions or decisions to take. This is where the **machine learning system** as a whole comes into play.

The machine learning system is fed with one or several different machine learning models and includes a certain engine or **optimizer** that defines the best action to take given the predictions obtained by the machine learning models.

In order to create the optimizer, one needs to mathematically formulate **which is the objective function to minimize or maximize** and take into consideration the existing constraints (could be operational or business ones). Based on the objective function, one will be able to define which machine learning model/s and variables are best suited for the business case.

Finally, the machine learning system must be designed in such a way that one will be able to evaluate its performance not from the technical point of view but from the business one. The machine learning system must include an **experimental design** capable of answering the following kind of questions:

**Has this ML System improved my performance?**Did we acquire new profitable customers? Have we increased the gross margin? Have we reduced churn? In how much?

**So which machine learning system could you incorporate at your marketing department?**

This question is a tricky one since there is no such thing as a general out-of-the-box plug-and-play machine learning system suitable for all businesses. These solutions to be successful must be built and adapted for each company’s specific needs and characteristics. This is why it is fundamental to have on the same team **marketing experts**, who understand the specificities of your own business, and **machine learning experts** who will understand the technical and data requirements needed to make the solution possible.

No matter whether you develop the machine learning capabilities in-house or you hire third-party experts, the key for a successful solution is to have these people working side-by-side.

**3 Machine Learning systems which can revolutionize your marketing department:**

Here we exemplify 3 Machine Learning systems that take in several different Machine Learning models and go inline with the consumer’s life cycle, i.e., **acquisition, optimization, and retention.** With these examples, you will be able to better understand what we mean by integrating the **Machine Learning models**, with an **optimizer** and an **experimental design**.

### 1. **Machine Learning Acquisition System**

Imagine we are standing at the initial point of your customer’s journey and there are consumers who initially do not interact with your brand or product, and prefer other brands or totally different product categories. The main goal for you at this phase is to **acquire new customers.**

Your acquisition system will very much depend on your business model. For instance, it won’t be the same acquisition system built for a grocery shop than the one built for a hotel, airline, university course or conference which have limited seats or stocks. Moreover, the acquisition system will also differ if we are dealing with a subscription-based product.

Since it is important to target customers with high value for your business and with high probability of responding to your marketing campaign, you will have to:

- Identify quality potential customers
- Predict their customer marketing campaign response

In other words, the objective is to *maximize the expected profit by finding the subset of customers that are likely to respond in the most profitable way.*

Nevertheless, it is important to also consider the probability of your customer to respond without the treatment, as we do not want to target customers who would buy your product even without the promotion or discount. This is formally known as maximizing the **uplift metric** or the incremental response.

The overall system overview would look something like this:

The **customer segmentation model** could be a **look-alike model** in which we learn from historical data how our target customer looks like or could also be a **clustering model** in which segments are identified in an unsupervised way. Or even more, we could segment our customers based on their expected **LifeTime Value** thinking of it as a regression problem —with a continuous assigned value— or as a classification problem in which we discretize the segment e.g., high, medium, low value.

Once we have the segments defined we need to estimate how these different segments will **respond** to the different campaigns. This could be a regression problem if thought as the probability of buying a new product or if thought as the amount of money spent given a certain discount.

Finally, the **Customer Acquisition Optimizer** will be fed with the Machine Learning models and will maximize the expected profit by finding the subset of customers that are likely to respond in the most profitable way, taking into consideration the operational and business costs and constraints for each business campaign. Once the optimal acquisition strategy is defined we need to randomly or quasi-experimentally divide the customers into two groups, i.e., treatment and control, in order to be able to measure the system’s causal gain in a statistically significant way. This last step is the one that enables us to determine the what-if scenario.

**target my prospect customers optimally?**

### 2. **Machine Learning Optimization System**

Imagine now that your customers are already inside your store (could be your online store or a brick-and-mortar one). At this point, you want to incentivize your customer to spend more. Promotion campaigns for these customers typically follow **up-sell** or **cross-sell** methodologies. In any case, the most relevant variable you can play with is the price for each product and a good approach is to dive into a **price optimization system**.

This system will have many things in common with the already mentioned Machine Learning Acquisition System but the most important and particular aspect of the price optimization system is the need and usefulness of generating a **Demand Forecasting Model.**

In our previous post, we explained how Machine Learning is reshaping price optimization and we won’t enter into many details here, but here we show the system’s general overview.

In this case, we will start by estimating the demand. The level of aggregation of the demand estimation will be defined based on the company’s needs and the available data too. For example, the demand forecast could be for each SKU, location, store per day, or week. Once we have the Machine Learning model trained with several different variables we can build the demand curves for each SKU in each moment in time. We would do so, by changing the input price and observing the output demand estimation.

**How many units will I sell if I lower my price by $10?**

With the demand curves and the price elasticities of demand estimated, we can feed our optimizer and get the **optimal price list which maximizes our objective function**. As mentioned before our optimizer will also take into consideration several different business and operational constraints. The available stock could enter our equation as a restriction for the optimizer or we could take a proactive approach and generate stock replenishment recommendations given the demand forecasted. Finally, we will have to define a treatment and control groups in order to be able to properly measure the impact of our price optimization solution.

This kind of Machine Learning Systems and design allows us to answer questions such as:

**How much did my gross margin increase**with the implementation of the price optimization system?

### 3. Machine Learning Retention System

Finally, our focus may not be on acquiring new customers nor optimizing our existing ones but could be saving customers who are likely to leave. This problem is many times modeled as *identifying customers with a high probability of churn.*

These churn prevention systems —or retention systems— are widely used in:

- Telecommunications
- Insurance
- Banking
- and other subscription-based domains where the continuity of a relationship is critical

Nevertheless, the problem of customer churn **is relevant for most non-subscription businesses as well, including retail**. Since acquiring new customers can be much more challenging and expensive than the retention of existing ones.

In fact, research done by Frederick Reichheld of Bain & Company (the inventor of the net promoter score) has shown that by increasing retention by as little as 5%, profits can be boosted by as much as 95%.

In any Machine Learning Retention System one needs to identify those customers with a high probability of churning **but also identify those who are worth investing in retaining**, since retaining low-value customers would be meaningless. Therefore, we could think of two important machine learning models to include:

- Churn Estimation Model
- LTV Estimation Model

These two models combined could be understood as the **expected loss**:

loss = p_{churn} \times LTV

The **Churn Estimation Model** could be thought of as a classical **propensity model** in which the output to be estimated is the probability of each customer to churn in X periods of time from now. However, we could also frame the problem as a **Survival Analysis** problem in which we would like to learn what is the customer churn probability in each period of time in the future. This Survival Analysis approach is useful from an explainability point of view where we could learn which are the main drivers of churn in each moment in time.

The **LTV Estimation Model**, as we mentioned before, could be thought of a regression problem in which we would estimate the exact value of each customer along with the whole relationship with the brand, or could be thought of as a classification problem if we would like to categorize customer in different proxy buckets (let’s say high, medium and low-value customers).

Finally, we would use both Machine Learning models to feed our **campaign optimizer** which would target customers with the highest expected loss. The campaign optimizer would use the campaign templates and would aim to maximize the Campaign’s ROI. In other words, the optimizer needs to define the optimal number of customers to include in the targeting list, or equivalently, finding the threshold score that separates these top customers (with high expected loss) from the low ones.

All in all, we would implement this solution (execute the campaigns) in a targeted and control group in order to measure the gain, and we would be able to answer this question:

**How much did your churn rate decrease**due to the implementation of the Machine Learning Retention System?

### Conclusions

Machine learning models are extraordinary tools to integrate into any marketer’s toolkit. They allow you to make accurate predictions using structured and unstructured data. However, in order for these models to deliver meaningful results, they must be integrated into a Machine Learning System which is fed with the Machine Learning model’s predictions and performs some sort of optimization.

**The key value of Machine Learning for decision-makers does not lay on the predictions itself but lay on the ability of learning which actions are best to take.** Therefore, these systems must be designed with an experimental mindset in order to be able to clearly estimate the impact of the actions taken.

To sum up, Machine Learning systems allow you to:

- Automatically compare several thousand scenarios
- Scale up your capabilities
- Adapt and learn from the latest trends and data
- Design easy to evaluate campaigns and actions

And finally: learn what to sell, when to do so, to which customers and at which price.

At Tryolabs we are Machine Learning experts eager to partner up with your marketing department in order to build high-value ML Systems. By working together we can crunch huge amounts of marketing data and boost your marketer’s capabilities and return on investment (ROI) faster.

If you have any doubts or ideas just contact us. We are here to help you.

Download here the free ebook with 10 real-world machine learning case studies that we have supported at Tryolabs.

## Like what you read?

Subscribe to our newsletter and get updates on Deep Learning, NLP, Computer Vision & Python.