CASE STUDIES
What We Can Do For You
Demand Forecasting
Client
UNILEVER
Challenge
Demand forecasting is a resource intensive and a very manual process and Unilever wanted to see if Machine Learning can automate this process with increased accuracy & bias or not.
Solution
- We built demand forecasting models using a combination of various machine learning techniques (pre-processed and handcrafted features, extreme gradient boosting and stacking models with different sets of features) on open source platforms – Python and Hadoop.
- We tested model robustness on 39 weeks of blind period. Proof of concept tested in 3 countries (Finland, UK and Hungary) for four categories: skin cleansing, fabric sensation ,fabric conditioners and dressings.
- This was completed in-house and the actual IP is owned by the client.
Output
Forecast accuracies were improved significantly and the model building process is now getting automated as well. Currently these models are in parallel run for assessing live performances.
Demand Forecasting
Client
Avon Products Inc, a global cosmetic brand
Challenge
Similar to Unilever, Avon Products had issues with their demand forecasting process. The forecasts were generated manually using business understanding and perception rather than any data driven statistical approach. The demand planning exercise involved various iterations of adjusting price and then re-running the forecasts, then assessing the impact and then again re-adjusting. This entire process of planning took 3 months.
Solution
- We built demand forecasting models using a combination of feature engineering and the application of a group of machine learning techniques like extreme gradient boosting.
- Here the approach was to build separate models for each group of products that exhibited similar characteristics, for example building models for pack products vs loose products separately.
- The final step on this was to integrate these models into a simulation tool where the demand planners can play with the inputs to generate expected output and can assess impacts of their changes straight away. This helped the demand planners in executing promotions & campaigns effectively & efficiently.
- This was completed with the help of in-house resources as well, so the actual IP is owned by the client.
Output
Forecast accuracies were improved significantly and the simulation tool helped the clients plan for their campaign effectively. Although we are not involved in the process any more, but currently Avon is rolling out the modelling approach to many other countries with the help of their internal data science team.
If you have any questions about demand forecasting and data analytics in the UK and worldwide, please contact our team.
Anomaly Detection in Emails
Client
Department for Work and Pension, UK
Challenge
Due to increasing risk of account hacking and phishing, the DWP wanted to identify the anomalous emails coming in and out of their network that could threaten their infrastructure.
Solution
- We extracted email content from the preceding year of data and used a word embedded representation of the content which compressed the email content into a smaller dimensional matrix.
- From this data we then applied Long-Short-Term-Memory & GRU’s over an RNN set up to build a sequence to sequence model.
- Reconstruction loss was calculated from the model to assess anomaly in the data.
Output
- Email anomaly detection algorithm was developed, tested and presented back to the stake holders.
- We provided insights on different types of anomalies and segmented them based on what type of a particular anomaly belonged to.
- We provided on-going support on helping the DWP to implement this live in their system.
Market Research Insights
Client
Mix London Market Research Agency
Challenge
Mix conducted a global survey on a popular ‘shoe’ brand with the aim of understanding their brand perception and brand position vs. competition in UK, France, Germany, Hungary, USA, China and Japan.
The challenge was to interpret key perceptions and customer sentiment from the survey responses, and in a scalable way.
Solution
- We identified the key response ‘groups’ using a combination of factor analysis and cluster analysis.
- We separated responses that were globally consistent vs. country specific to enable the client to action correctly.
- A scalable process was developed and then automated end to end.
- The solution was extended to other brand surveys, using <1% of the original processing time – resulting in a high amount of ongoing cost savings for the company.
Output
- Cluster & factor analysis results were presented back to the clients.
- An automated & templated tableau dashboard visualisation tool was built with several factor & cluster scenarios, allowing the user to choose a factor & cluster solution and look at their profile with several dimensions.
Machine Learning Technique to Predict Churn
Client
CP Group Lotus Retail in Shanghai
Challenge
Lotus is one of the biggest retailers in China but their customer retention numbers were decreasing month on month in the preceding year. They wanted to know why their customers were leaving and to also predict which customers were likely to churn, so that they could run targeted campaigns to prevent this.
Solution
- We first created groups of features that linked to each of the areas:
a. Price
b. Promotion
c. Range
d. Channel
e. Demographics - From a panel of customers who were active last year, we created several characteristics from the above groups.
- We used multiple machine learning techniques, and developed a two staged approach:
a. Logistic Regression Model with L1 & L2 penalties that helped in explaining the drivers of churn.
b. Random Forest & Gradient Boosting on selected variables which helped in having a high accuracy of churn prediction.
Output
- Insights regarding drivers of churn which were shared with the client.
- Automated churn prediction modules were deployed in client servers.
- Automated churn driver alerts were created that indicated red-yellow-green alert on each of the drivers.
If you would like further information about machine learning and data analytics in the UK and worldwide, please contact our team to hear how we can help your business.
Personalisation
Client
CP Group Lotus Retail in Shanghai
Challenge
Lotus wanted to build an automated personalised recommendation system for their instore promotion. They had large screens set up in different areas of the store and upon scanning the loyalty card, six of the current promotions were to be displayed which was based on an algorithm that is real time and relevant to that customer.
Solution
- We developed an algorithm that recommended products of two types based on the pool of promotion products.
a. Products that were bought before by the customer.
b. Products they might like but which had never been bought by the customer. - We used several features regarding past promotion, customer attributes, product characteristics, offer value, then developed a probabilistic ranking of the offers for each customer that was specific to their need.
- Together with this, we applied Matrix Factorisation techniques to recommend products that the customer had never bought before but had a high likelihood of buying.
Output
- Automated real time recommendation systems were integrated with the servers in store.
- Everything was deployed in-store with automated control selection for campaign evaluation.
- A tableau dashboard tool for campaign evaluation analysis was also developed for the clients.