Machine Learning @ CARS24 — algorithms coming together in a C2B transaction

This blog is an attempt to explain how different independent ML modules come together to inject data science driven decision making across a single transaction in our core C2B business

Supply first & transaction led business model of CARS24

In 2015, CARS24 pioneered the ‘supply first & transaction led’ business model in India’s used car industry when rest of the organized players were still solving for demand side (through classifieds based lead generation model).

This difference in our core business model (C2B vertical, which continues to account for ~70% of our current revenue) implied that we had to conduct thorough inspections of cars and furnish all that information in a good format (and in near real time) to the channel partners participating in live auctions and spread across India ; and now having done that for ~600K cars and transacting ~25% of them we have huge leverage in solving the inefficiencies of used car ecosystem.

https://www.cars24.com/

We know the true price at which a transaction happened for a given car with all its attributes (~150 inspection parameters) and transient demand-supply conditions which allows us to do a robust job of estimating true price of any inspected car in future — refer the blog on Profecto, for more on this.

Also, we have very deep understanding of the requirements of our ~10K strong base of channel partners spread across India (harnessing their click-stream behavior), which helps us drive higher participation in auctions through real time relevant targeting and catalog personalization — refer the introductory blog on Auctoris, our recommendation engine for more details.

C2B transaction — high level flow

A typical C2B transaction starts when a prospect seller gets to know about us through any of the marketing channels (digital, TV/radio ads, word of mouth, billboards etc).

In an ideal world, prospect seller then goes to our website(or app), evaluates the expected price (web quote) and books appointment. Seller then visits the branch for physical inspection and auctioning of the car, gets the desired price from auction, sells the car and gets instant payment from us.

Different stages in a typical C2B transaction at CARS24

Of course, in real world things move somewhat differently!

Not all customers who land on our site end up booking appointment, and not all those who book appointment show up on our stores on the day of appointment (aka ‘funnel dropouts’). Also, among those who come for inspection, not everyone gets the fair price which the car deserves (as per CARS24 estimates) due to transient market forces and/or auction inefficiencies, and even among those who do get the fair price not everyone is willing to sell the car at that price! In case we fail to fetch the true price in first attempt or when the seller expectation is higher than true price, we need to go through the follow up processes to help convert the customer — almost, one third of the cars transacted by CARS24 are procured post follow up!

ML algorithms involved across a transaction

Let me start with a quick introduction to our key ML modules

• Profecto (meaning ‘assured’ in Latin) is the pricing engine which establishes true price of a car accounting for make, model, age, mileage, inspection parameters, documentation and localized demand-supply factors.
• Auctoris (‘organizer’ in Latin) is the buyer recommendation engine, currently focused on channel partners personalizing auction catalog, notification and calling from sales team.

One of the applications of Oraculum — enabling our retail team with buyer acceptance likelihood using a color coded slider.

• Oraculum (‘oracle’, one who delivers prophecy) focuses on Post Auction optimization in the core C2B business i.e. tools to rationalize seller expectation, models driving decisions around re-auctioning vs one-click-buy sale vs inventory buying as well as dynamic margin control after accounting for confidence in Profecto for the given car and seller conversion likelihood at different price points.

More recently we have kicked off 2 more modules

• Profundus (meaning deep/profound in Latin) is the deep learning module focused on image and sound processing & classification to address human subjectivity in our inspection process

Profundus — deep learning for image processing & classification at CARS24, fledgling but with very bright prospects!

• Magneto (no, not Latin this time and no marks for guessing what it means) is focused on optimizing our pre-inspection leg (the journey from session to ‘online evaluation’ to ‘booking of appointment’ to ‘footfall’) — this includes lead prioritization models, web-quote logic, web/app feature optimization through A/B tests, CRM fine-tuning etc.

So, if I were to establish a chronological order, ML intervention begins with Magneto focusing on arresting the pre-inspection funnel drop. Post inspection, Profecto (supported by Profundus) ensures accurate pricing (true price) in the backend followed by Auctoris ensuring strong auction participation needed to fetch bids from market above the true price. Finally, Oraculum (tech-enabled API integrated features in our internal systems) comes into play to optimize the follow up process where we attempt to align buyers & sellers close to the true price of the car.

Profecto (now being supported by Profundus, deep learning module fine-tuning inspection parameters), Auctoris, Oraculum and recently kicked off Magneto work in perfect harmony to drive efficiencies across different stages of a C2B transaction

Next Steps

These are still very early stages for ML @ CARS24 and we have a long way to go. We are constantly iterating our models, innovating their applications & use cases and injecting them in as many core functions as possible.

We are betting heavy on Profundus, the deep learning module focused on classification of images and sound. Lot of this optimism is driven by the huge collection of labelled images (~30 images per inspection) and engine sound files we have collected over the last many years through our evaluation engineers. Robustness of Profundus would help us address human subjectivity in inspections, which shall essentially further strengthen Profecto, the pricing engine which depends heavily on inspection parameters.

Also, equally critical is the success of Auctoris & Oraculum. However thorough the inspection and accurate the pricing, end of the day we need to get the relevant buyer participation in auctions to drive bidding and we need to get the buyers and sellers aligned to close the transactions.

Last but not the least, ~90% of those evaluating our web-quotes drop off the funnel before coming to the store for inspection. There is huge headroom available to improve this funnel which shall be the key focus of Magneto — we have just scratched the surface!

We are always seeking outstanding people to join our data science team working across complex problems like Pricing, Recommendation Engines, Auction Scheduling/Strategy, Marketing Optimization, Retail/Sales efficiency projects and Risk models for our lending vertical.

Please reach out to me directly at naresh.mehta@cars24.com or drop a mail to datascience@cars24.com or hiring@cars24.com for more details.