Machine Learning

Machine Learning

We help customers with applied machine learning models built into their business as usual processes. We do segmentation, customer and offer level affinity models, campaign response models across various channels and next best offer recommendation models to increase EBIT from X Sell  and customer loyalty. With our machine learning based targeting our customers reached 2-6 times higher CTR and booking rates vs control group. We do value based churn modelling to support efficient value based retention efforts.

Increase Cross-Sell efficiency with Machine Learning

Marketers holy grail is to find right answers of basic questions of targeting: To whom, what to offer, when to offer at what price. A popular saying illustrating how difficult it was to qualify the response to advertising is attributed to John Wanamaker:“Half the money I spend on advertising is wasted; the trouble is I don’t know which half”. According to our experience X-Sell campaign response rates are actually much lower. Vast majority of campaign response rates -not supported by machine learning models- are between 1%- 5% depending on industry.
One of the most common goal of our Machine Learning projects is to increase efficiency of X-Sell campaigns. In such projects we do analyse data from past campaigns and build response models or  in case of campaign based analysis is not possible, we build look-a-like models based on portfolio overlap. Then we do score customers based on their affinity of given offer. We build profit charts and optimize campaign audience based on customer affinity score, cost per targeting (call, email), estimated revenue from successful X-Sell. We apply test and control methodology to validate model and campaign success results. Results are highly dependent on industry, market environment and available data quality. In our projects we have seen 2X-6X higher response rates increase for X Sell campaigns where audience selected based on Machine Learning models versus control group. With our solution partners we also help our customers to integrate machine learning results into their core BAU processes via automating entire process.

Next Best Offer Recommendation Engine with Machine Learning

Machine Learning Recommendation engine Geo-Demoegraphics Collaborative filtering Content Based filteringOne of the biggest challenge for most of marketers across industries being it retail, e-commerce what service or product to offer on the next campaign or at next customer touch-point. Common approach especially when potential offer palette is relatively small to go for highest revenue regardless customer need. In more matured advance stage organizations tend to fine tune their campaigns based on earlier results with descriptive or predictive models as described in chapter “Increase Cross-Sell efficiency with Machine Learning”. That can lead to strong results as long as the offer palette is a few dozen.
We do help our customers with combined machine learning algorithms to find next best offer that is personalized to the customer level even when there are hundreds thousands of customers and products varieties are over ten thousands threshold. Our solution combines geo-demographics with collaborative filtering and content based filtering to find best offer for the customers customer.
With our tailored recommendation engine we managed to triple click thru rates in e-commerce.
Read more about our customer story how we optimized with recommendation engine email Open Rate and Click Thru Rate at Netpincér.

Machine Learning based Churn Models & Attrition Models

Primary objective – from analytics perspective – of churn management projects is to identify high probability to churn or attrition segment within Portfolio. While we build machine learning, churn predictive models we do identify and share with the customers key drivers of attrition, churn. We also recommend and implement action steps based on findings. We build descriptive dashboards to measure Churn Management Performance based on saved customer value.

Machine Learning in Location Based Offers, Geo-targeting

Customer whereabouts Geospatial information such as home address, frequent path, transaction location can significantly increase targeting accuracy. With our GIS partners we can geocode and enrich customer attributes with public GIS information such as income proxies, property price index,  population density, population demographics, location surrounding residential type, town center index, tourism index and couple more that boosts machine learning models accuracy especially when customer level data availability is limited.
Combining GIS public information and our semi automated machine learning approach helped us to predict which customers will visit which bank branch how many times just within 6 days.
Machine learning combined with GIS data not just greatly improved efficiency of cross-sell campaigns of our customers but significantly boosted offline Door to Door and billboard campaigns efficiency.

Leveraging Machine Learning in Micro Location Analysis

On top of location based targeting with GIS and Machine learning Holistic and GeoX consultants can support business development where to open next branch, restaurant, kiosk. Where they can expect the most customer visit from their preferred segments. We do analysis based on historical performance if it is available, if past performance data is not available for given business we can analyze and predict best performing locations based on competitive environment location analysis.

Big Data and Machine Learning

Sales and Marketing managers expect answers to “simple” core questions from Marktech vendors: What to sell? To whom? When? At what price? On which channel? With which feature?
Big Data usually sold to business decision makers as philosophers stone as an ultimate solution to understand their customers needs.
Most of the Technology Vendors Big Data efforts tend to focus on storing and processing large volume of data in a cost and time efficient way on premise or preferably in the Cloud.
There are a couple of use cases where data collected by big data solution can have a significant impact on Machine Learning algorithm performance.
Machine Learning can help with sophisticated answers for most of core questions with or without data from a Big Data source. We have achieved up to 6 times response rate increase via ML driven campaigns compared to control group only on data from core and CRM systems without any big data source. We have achieved 30% response rate via combining Big Data event driven campaign solution with Machine Learning.

Contact us for more details:

Our consultants have decades of cumulative experience in machine learning, we have built 100+ machine learning models across multiple industries from banking to e-commerce across multiple geographies.


Machine Learning Frequently Asked Questions


There are no “guaranteed” uplift or benchmark KPIs available (yet) for Machine Learning model based cross-sell campaigns vs educated guess based targeting. It depends on the data sophistication maturity of existing campaign management environment and given business environment. Machine Learning Modelling has to be done first. Based on the propensity to buy we can tell what will be the expected response rates for campaigns. Based on cost estimations we can also define ideal audience size for the given cross-sell campaign. Holisticrm consultants have achieved 2x to 6x times higher response rates in X-Sell campaigns with Machine Learning based propensity to buy models.

Different tools languages have their strength in different area. Some might fit well into on premise infrastructure some are cloud only. Some have the latest algorithms implemented others are more scalable. Some have strength in data processing, others are better for visualization.

At Holistic we select the right tool for the right task that fits customer environment keeps operationalized machine learning framework total cost of ownership at minimum level. Our data-scientists have measured different tools and methods performance.

Different tools in terms of model predictive accuracy tend to converge to each other. On the other side there are still significant differences in speed and performance. Just different implementations of R language have significant performance differences as it is shown on chart below

Machine Learning R vs R Open

Read more about Machine Learning tools usage including R vs Python trend changes base on data scientist poll here.

Bigger players in Machine Learning providers like Google, Microsoft provide multiple choice of solutions. Ivan Kosyakov article and mindmap below gives better understanding of variety of tools just from Microsoft for Machine Learning.

If we go by Total Cost of Ownership approach we can distinguish “Inital” or set up Machine Learning (project costs) and “maintenance” (score and model refreshment). And there are some “related costs” such as data-warehouse and system enhancement both for model input and score result automation into day to day operations. Each have different Human resource and computational resource requirements. Projects costs can start at few thousands USD range and the limit is the sky (especially if it is in-sourced) depending on the complexity data and level of automation.

There is no silver bullet machine learning algorithm one for all solutions. There are algorithms that are hyped by media and community such as extreme gradient boosted decision trees (xgBoost), Neural Networks, Deep Learning algorithms like Convolutional Neural Networks (CNN), Recurrent Neural Network (RNN). In most of the cases it depends on the type of the business goal and available data set what algorithm works the best for a given problem. Here you can find a good summary cheat sheet for different machine learning algorithms on AzureML. At Holistic we prefer to test several algorithms and select the champion to score or select set of algorithms and ensemble scores in a final model.


Machine Learning model’s shelf-life varies between few days up to 18 months. Shelf-life really depends on business data variety, industry and market environment dynamics. According to our experience Ecommerce models tends to have shorter shelf-life while retail banking models can provide stable results without retraining for several months. In Holistic machine learning projects we automate both scoring and model retraining on need basis keeping associated costs at optimal level.


By continuing to use the site, you agree to the use of cookies. more information

The cookie settings on this website are set to "allow cookies" to give you the best browsing experience possible. If you continue to use this website without changing your cookie settings or you click "Accept" below then you are consenting to this.