A leading car maker and financier in the fast-growing Indian market improved the performance of their leasing business. We helped them by predicting residual value of their cars.
Context
The ‘residual value’ or RV of a vehicle in plain words is its ‘worth to a buyer’ when the lease term ends. Vehicles do depreciate. How much, really depends on a host of factors. The business uses this important number to arrive at monthly lease payments. They also use it to set purchase prices if the lessee decides to buy the car at lease end. The factors that influence RV include the make and model, market conditions, mileage, ownership, and wear and tear.
Challenge:
The business would like to get the RV as accurate as possible. Overestimating it means lower lease payments but at lower profits. Underestimating it makes the business un-competitive. Even so, very few companies build predictive models of RV. Our client was in fact the first to try this in the Indian market. Initially they tried using off-the-shelf models from other geographies, and soon realized the limitations. That’s when the idea of custom, home-built models came about.
Solution
Dhurin mobilized a three-member team, well versed in modelling, python, and Django. They first established critical data sources including internal sales data, dealer data, re-sale price data scraped from the web, and some more. Synthesizing this programmatically and at times manually with the product and sales managers help got us covered on the data front. The method we adopted was indeed adventurous and pushed the boundaries of modelling skills. For the statistically minded, we built an OLS model with over 70% efficiency reflecting a good strong model.
What made this special?
1. No benchmark of any kind.
2.Unavailability of data – which is why web-scraping and other sources were important.
3) Intensive research on market practices and data harvesting.
4) high level of customization with a custom user interface.
Benefit
The model is already in production. The leasing plans are already utilizing the residual value predictions from the model. Initial estimates indicate a 7% gain in pricing and 30% more accuracy in predicting residual value.
