Selling Used Cars Faster

This is an example project to show you how I could make your business more profitable. (The project is fictional, but the data is real: from a U.S. car seller called Drive Time, downloaded here).

There are five phases:

  1. Oriëntation: we discuss your current business processes to identify opportunities, no strings attached.
  2. Focus: we pick one idea to pursue in more detail. Together we decide (1) how this fits into the existing business processes and (2) how the improvement will be measured.
  3. Implementation: if we both have confidence in the plan, I will implement it. A first version should be ready in a matter of weeks, and a full solution in a matter of months.
  4. Profit sharing: we measure how much value is being added to your business, and you share a percentage of that with me on an ongoing basis.
  5. Maintenance: since we have a mutually beneficial relationship now, I will monitor the system and make improvements where possible.

Phase 1: Oriëntation

Suppose you own a company that buys used cars at auctions, fixes them up, and then sells them to the public in showrooms in various cities. To keep making a profit, it is essential that the cars are being sold quickly. Every slow-selling car takes up space in the showroom as well as your working capital.

Your current policy is to offer a big discount on cars that haven’t sold after 90 days. If a car still isn’t sold 30 days later, it is sold at auction – to recoup at least some of the money invested.

In one of our discussions it is revealed that the routine administration keeps track of all kinds of data about the cars: make, model, colour, the year it was built, mileage, when and where it was bought and sold, etc. Nice! Maybe we can use this information to predict which cars will be “slow sellers”, and avoid buying them in the first place. This would also free up space and money, so that you could sell even more of the “quick sellers” than you do now.

Phase 2: Focus

So let’s build a model that predicts how fast each car will sell. But first: how would this be used in the business, and how would we measure the added value?

These are the current business numbers: 

Current situation Normal sale Discounted Auctioned Total
Average buying price € 3000 € 3000 € 3000
Average selling price € 3600 € 3000 € 2100
Gross margin 20% 0% -30%
Number of cars / month 2335 272 311 2918
Average time in showroom 25 104 121 43
Revenue € 8406 k € 815 k € 653 k € 9874 k
Gross profit / month € 1401 k € 0 k € -280 k € 1121 k
Fixed costs € 1000 k
Net profit / month € 121 k

Table 1: current situation

The goal of the project is to increase the net profit, by:

  1. Increasing the profit per car (by avoiding cars that would sell at a loss).
  2. Selling more cars per month (with the same resources).

The plan is to predict the throughput time for each car, using recent sales data (6 months). The prediction must be made before the purchase, so purchasing officers can adjust their decisions. Of course, they already use their experience and rules of thumb (like “Peugeot is popular”), so the model should be more accurate than that.

How will this fit into current business processes?

We’ll build an online app/website to be used by the purchasing officers:

  • With a few taps on a mobile phone, the purchasing officer will enter some basic info about the car that is about to be auctioned. Then immediately the screen shows whether the normal purchasing process can be followed, or whether this car should be skipped.
  • The web-app is always available via the internet (we don’t need an offline version).
  • The web-app and the existing IT-systems are not interdependent and therefore will not be linked. This significantly reduces the risk of outages.

What about data security?

  • First of all, the data only includes certain features of the cars themselves, not the buyers. So the data is sensitive in terms of competition, but not privacy.
  • Data is shared only incidentally with me, at the start of the project and later for updates every month or quarter.
  • The web-app only contains the predictive model, not the underlying data.
  • Users of the web-app will be required to log in.
  • The web-app will collect data on usage of the app: who entered what information when, and what prediction was returned. This is necessary to keep monitoring the quality of predictions.

In principle, the web-app and predictive model will remain under my control (but with a backup-plan in place just in case). A monthly review will ensure the quality of predictions.

How will we measure the added value?

After introducing the model, some cars will not be purchased that otherwise would have been purchased. If the predictions are somewhat accurate, those avoided cars will be the ones with longer throughput times. So the average throughput time should go down. This means we’ll be able to not just replace the avoided cars with others that sell faster, but there will be some room to buy and sell extra cars.

However, there are limits to this type of growth: will we be able to find enough cars that the model does not reject? And will the staff be able to handle the extra workload? To be safe we assume that the total number of cars sold per month can not grow more than 20% (if it does in practice, that profit is all yours).

Together we decide to measure the added value as follows:

  1. We use Table 1 above as the basis.
  2. We assume the average purchase price, sale price, margins, and throughput time remain the same for each outcome (normal sale / discount / auction).
  3. We assume the fixed costs do not change either (mainly real estate and payroll).
  4. We assume the total number of cars sold can not grow more than 20%.
  5. The only thing that really changes is the number of cars sold normally vs. discounted vs. auctioned. For this we use the actual outcomes in each month after the software has been rolled out. After 5 months we’ll know for each car in month 1 how quickly it was sold.
  6. Using Table 1, we will estimate the net profit each month in the new situation. The difference with the old situation, so the estimated additional net profit, we’ll divide every month: 90% for you and 10% for me.
t

So why all the estimating?

Why don’t we just compare the real increase in profit, and share some of that?

Because I want to be rewarded for only the added value of the model. The net profit of the business as a whole can be higher or lower for all kinds of reasons: expansion into a new region, staff problems, calamities, interest rates, etc.

Phase 3: Implementation

The table below gives an idea of what the implementation phase could look like. The system is operational after month 4.

Month Activities Result
Month 1

Extract and interpret the data.

Build a first model and estimate its added value with Table 1.

A first feasibility estimate.
Month 2

Improve the model.

Consult with purchase officers to build a user-friendly web-app.

Production model ready.

Web-app sketched out.

Month 3

Build web-app interface and test with end-users.

Integrate the model into the web-app, more testing.

A working web-app.
Month 4

Test the web-app with more users, organize training events to get everyone to use it.

Actively monitor app usage.

Web-app actually being used.
Month 5

Keep monitoring app usage.

Build systems for monthly monitoring & updating of the model.

Model monitoring system.
Month 6+ First model-monitoring event (although data is incomplete because some cars are not sold yet, so their final outcome is not yet known). Motivate users by sharing tentative successes.
Month 10+ First complete monitoring event: all cars from month 5 have now been sold, so we know exactly how good the predictions were. Concrete results! Go to Phase 4.

Table 2: implementation

Phase 4: Results & Profit Sharing

Although this project was not implemented for real, we can use cross-validation to get an idea of what results could look like. Table 3 below shows the results from a fairly simple predictive model. The bold numbers indicate what changed compared to Table 1 (old numbers in parentheses). (The 2335 cars “sold normally” is the same as before, but only by chance – it could have been higher or lower).

Current situation Normal sale Discounted Auctioned Total
Average buying price € 3000 € 3000 € 3000
Average selling price € 3600 € 3000 € 2100
Gross margin 20% 0% -30%
Number of cars / month 2335 (272) 214 (311) 207 (2918) 2757
Average time in showroom 25 104 121 (43) 38
Revenue € 8406 k (815) € 643 k (653) € 435 k (9874) € 9485 k
Gross profit / month € 1401 k € 0 k (-280) € -186 k (1121) € 1215 k
Fixed costs € 1000 k
Net profit / month (121) € 215 k
Added value /  month € 94 k
Added value after performance fee (10%) € 84 k

Table 3: new situation

Conclusion

70% more profit

Net profit increased by 70%, thanks to data collection, a small piece of software, and 4-5 months of work.

Without investing

You start enjoying the benefits in month 5, and only start paying in month 10 (and the payments are only 10% of the extra profits anyway).

And no worries

The quality of predictions is monitored every month, and future updates cover the regular software and the predictive model.