swahareddy/prospekts

A tool (built for a hackathon) to assist wealth managers identify, research and contact clients

Generate potential HNI clients for private banking and wealth management

This was a team project for the FoB Hackathon 2020, ML Challenge. We were a group of 5.

The Use Case

• HNIs are defined as individuals having net worth of > 5 million USD. These high profile individuals are still significant in number and are often in need of wealth management for their businesses or family offices.
• Create an AI/ML based platform to enable identification of potential prospects for Wealth Management leveraging News & Social Analytics.
• Using NLP & Clustering Techniques to identify topic of interests for a group of prospects
• Profiling the prospects using Publicly Available Social Media Data.
• Identify the Degree of Affinity of the prospects to the Trends.

Demo application - https://datastudio.google.com/s/qRyMPSH5YFQ

The ML Motivation

• Utilize the affluence of data & soaring social footprints of wealth creators, machine learning can be the motive force to gauge Social, Cognitive, Behavioural & Cultural elements of individuals who could be our future prospects.
• Drastically reduce the manual task of identification, analysis, profiling and segmentation of the leads generated.

The Data Strategy

• Global Knowledge Graph (GKG): Starting point to filter data as per time series, countries, themes, etc eventually
• Data Segmentation: Confirm data metrics, data scale-up and segment variable definition. followed by Profiling and interpretation.
• Profiling: Building a profile map and perform social and news analytics
• Network Graph: Networkx – Network Cascading Algorithm to simulate link associations in the network graph of prospects that have a social or business connection and a potential lead.
• LinkedIn, Wikipedia, Twitter and Instagram were used for profiling and key value extraction.

The ML Model

• NLP : Stop Word Removal, Tokenisation, Stemming, Lemmatisation, N-Gram Modelling, TF-IDF to find out high frequency n-grams
• Latent Dirichlet Allocation (LDA): Generative statistical model that allowed us to use sets of observations in order to explain similar parts of data by unobserved groups thus facilitating Topic Modelling.

The Pipeline

A loooot more could have ben done. Some half baked ideas of ours are here
Inspiration for LDA topic modelling. See Stages 1-3. https://github.com/mukut03/ESG-BERT