Why are we so excited about Liquidity ML Modeling Studio

By Oron Maymon — Chief Science Officer, Liquidity Capital

Liquidity’s data science team is tasked with designing and deploying models that enable investors and lenders to augment their investment decision making process, giving them a competitive advantage by surfacing insights and projections that display the risk reward ratio of the various financial instruments they wish to offer to their clients. Working closely with our investment analysts, product and development teams we were able to introduce a novel solution for managing debt for the private tech markets.

We started our path in developing models focused purely on private SaaS and eCommerce companies who are in rapid growth stages, a huge underserved market for quick working capital needs. As demand for our platform keeps growing, we are finding that as can be accepted, that applying our models to new verticals requires adjusting and adopting new and different risk profiles.

While this keeps our intellectual curiosity at peak high, we are also, as we deal with predictions and their implications on a daily basis, are aware of what it means about our future business- Would we come to a point that customizing and configuring models to meet different business verticals become cost and time prohibitive ?

This notion itself leads us to understanding what we need to build. Also, we came to realize the challenges we are trying to tackle are actually one of the critical challenges that holds back productivity from a large portion of the global enterprise investment in ML and analytics. Scale.

If we look into the core challenges of data science teams in enterprises, one key repeating issue is that data science deliveries do not really scale. There are many anecdotes about improving customer loyalty or improving key isolated metrics, but there’s a violent consensus that most organizations can’t really continue and justify or measure the return on investment in analytics and data science infrastructure and cannot embed ML capabilities into the delivery of the value chain.

Trying to address those problems, which will likely worsen if the prediction of a global shortage in talent in the data science space were to come true, a breed of AI and ML software was developed, widely addressed as Auto ML.

Those solutions are mostly focusing on automating the data science processes of building training and deploying machine learning models. Their premise is to increase the productivity of the data science teams by automating a large portion of repetitive tasks — which is what machine learning is all about really. A newer breed of Auto ML, are taking the “No Code” agenda a step further by seeking to empower data analysts to be able to perform some of the more generic mundane tasks of data science.

As we looked into those “no code” solutions, we were inspired by some of the advances made in the field. Yet, we think there are some key elements missing in the current applications, as it pertains to our need to quickly deploy models for new business verticals. They are missing significant components from the value chain of data science driven features. Between getting the business requirements right, gathering the data, and engineering the features — much time is lost, not to mention controlling deployed models in production and embedding the model effectively in the business flow, to drive efficiency and business actions.

Trying to sort out the approaches and solutions out there, vs where we think we need to be, we found several sources offering a rank approach to the maturity of ML solution, based on the Levels of Vehicle Autonomy.

We saw 5 levels practiced and applied in real life circumstances. We created the following breakdown (below) in order to make it clear what we are after in realy building an Domain specific solution for automated ML level 5.

Level 1

No automation. You code your own ML algorithms.

Level 2

Use of high-level algorithm APIs. Sklearn, Keras, Pandas…

Level 3

Automatic hyperparameter tuning and ensembling. Basic model selection.

Level 4

Automatic feature engineering and feature selection, technical data augmentation, GUI.

Level 5

Automatic domain and problem specific feature engineering, data augmentation, and data integration

Liquidity business model and target market, is putting us in a unique position where the domain is large enough to require such an effort. The commonality in basic practices, addressed by the existing Credit Dynamics flow for predicting debt defaults, financial debt structuring and risk management, enable us to design a system that would introduce a new way of harvesting the power of ML for business, addressing the even larger blue ocean of corporate debt.

A certified corporate credit analyst, will be able to offer financial solutions to a newly engaged business vertical, in a short time, making Credit Dynamics value proposition to be vertical, agnostic. A debt provider can then take over the debt using Credit Dynamics risk taking, devising his own investment thesis.

Like every journey, we started small, with making sure our parameters are tuned towards the right business segmentation. Engineering features and re-wiring modules within the platform is our next goal. The end game is to have the entire ML value chain performed with Liquidity Modeling Studio.