DoorDash: great unit economics, but many unanswered questions

A lot of people have a lot of different opinions about DoorDash in light of their pre-IPO S-1 filing. Many reporters and pundits are concerned about profitability in the restaurant meal delivery category as a whole, wondering whether any of these firms could ever sustainably turn a profit. Many others are more optimistic, pointing to some of the (many) customer-related disclosures that DASH scattered throughout its S-1. And of course, all of this uncertainty is only further compounded by COVID, which was and has continued to be a big shot in the arm for everyone in the category, but introduces yet more uncertainty about how the world will look post-pandemic.

Given my own academic research with Elliot Oblander and Kivan Polimis into the industry, which I will hopefully be able to link to over the next couple of months (if you're interested, shoot me an email and I'll make sure send it to you as soon as it breaks!), and given the volume of disclosures that DoorDash provided in the S-1, I took a closer look at the company to uncover what its underlying unit economic condition currently looks like, how COVID has impacted it, and what their "normalized" economics may be after stripping out pandemic-driven behavioral changes.

This piece is long, so if you don't have much time, these are the main highlights to focus on:

1. DoorDash has exeptional repeat buying -- for example, I infer that customers do about 2.6 orders in year one, 3.5 in year two, and even more in year three. Very back end loaded.
2. DoorDash's variable margins are thin but expand significantly over time (both calendar time and as a function of customer tenure), rising from 0% or less in year one to 10% by year three. The thinness of these margins makes their valuation very sensitive to margin assumptions, which is only further compounded by how back end loaded they are.
3. Even when we focus on short horizon customer valuation estimates, unit economics look exceptionally strong. I infer that DoorDash gets a return on customer acquisition investments of 250% and ~900% over three and five year horizons, respectively. This is very good, and significantly de-risks their unit economics, despite how back end loaded ordering and margins are.
4. COVID has been a huge shot in the arm for DoorDash, but as with other companies, it has been a bigger boost for acquisition than for retention. I infer about a 200% increase in business from newly acquired customers relative to baseline due to COVID, versus a 70% increase to repeat purchasing from existing customers.
5. If their customer acquisition data can be trusted, DoorDash is spending about $6 to acquire customers, and making back $21-60 after acquisition over a 3-5 year horizon.
6. But I think there may be some sort of data error in a key chart in the filing which could change these inferences. As a result of this, everything (CAC and CLV) could be 4x larger, implying customer ROI that is directionally the same, but with much bigger numbers in the numerator and denominator, so to speak.
7. It is hard to come to a conclusion one way or the other about their valuation until this data issue is clarified. If their customer acquisition data is to be trusted though, this analysis might bode less well for their overall valuation as it would suggest that they've already acquired more of their TAM, which would create more significant headwinds for future customer acquisition, and/or revise downwards our expectation for the post-acquisition goodness of customers due to possible rampant abuse of new customer promo discounts.
8. A more detailed analysis that is significantly more comprehensive and applies to the entire restaurant meal delivery category will be coming soon.

Below, I unpack how I reached these preliminary conclusions. I first take a step back to summarize the model I ran, how well this model fits their disclosures, and what the data and model imply for DoorDash's unit economic position and where it may go in the coming years.

12/3 UPDATE: If you are short on time, I would also recommend the S-1 Club's analysis of DoorDash. A summary of my thoughts on their unit economics are included, but it covers a lot more ground. Highly recommended.

The Model

The underlying series of statistical models I used is nothing new. The plain vanilla version of the model is what I wrote about in my paper in the Journal of Marketing Research. It is identical in structure to other analyses that have been performed on non-subscription companies such as Lyft, Farfetch, and Revolve through Theta Equity Partners, which regularly does CBCV like this work (if you are looking into an investment that has end customer visibility, reach out!). Here's the reader's digest version of what I do:

1. I propose statistical models the flow of new customer acquisitions over time, how long those customers are retained, how many orders they place while alive, and how much they spend on each of those orders, which drives marketplace gross order value (GOV).
2. I allow for customer acquisitions and existing customer purchase frequency to be higher or lower during Q2 and Q3 2020, because of COVID. Of course, we suspect higher, but let the data determine exactly what the magnitude is.
3. I allow order rate to vary as a function of customer tenure, to allow for the possibility that customers who live longer buy more as they age, but do not allow for this variation to persist for more than three years. This prevents order rates from shooting off to the moon, and accounts for the fact that we basically have no more than 2-3 year's worth of purchase data for the vast majority of DoorDash customers, making longer horizon projections fall outside the range of the data. I also allow spending to trend in calendar time, because of inflation and pricing power.
4. I train these models on all relevant customer-related data that underlies these processes: namely, the quarterly evolution of GOV, broken down between new and existing users, GOV by cohort indexed to year one, orders and GOV over time, and the number of active customers in September 2020. In English, I'm finding the behavioral patterns that are most consistent with all of the disclosed data.

That's it - no surprises.

Once I've forecasted out customer purchase behavior, I then incorporate the cohorted margin data that DoorDash disclosed to infer their (profit-based) unit economics and CLV-related measures.

Model Validation

It's hard to trust an empirical model -- any model, not just mine -- that doesn't fit and forecast the data well. While not perfect, the model validates well in general, especially when we consider how simple the proposed model is. Let's take a look at those fits...

Actual versus expected total orders:

Actual versus expected total marketplace GOV:

Actual versus expected GOV from new customers:

Actual versus expected GOV per order:

Actual versus expected cohorted GOV (indexed to year one) and active customers:

Are the fits perfect? No. There is evidence, for example, that COVID's impact on new customer GOV was different in Q2 versus Q3. And while we generally capture the pattern of expanding monetization as a function of tenure, its not perfect. But all in all, this is a pretty darn good fit for a simple model, which makes it more likely that the results are valid and would generalize well into the future.

Now that we've checked the box on validation, let's see what the model implies for customer behavior...

Key Insights

Repeat purchasing: exceptional

As we can see from the charts above, there weren't a whole lot of dynamics in GOV per order over calendar time. And when we look to alternative data sources, we can directly observe that there also isn't much in the way of dynamics in GOV per order for initial versus repeat purchases. They're very comparable.

All the action here is in the repeat purchasing.

Here are some of the dynamics that stood out to me:

1. The primary reason that GOV improves as a function of tenure in the indexed cohorted GOV table above is because of customers who live longer also buy with a higher frequency as they age. For example, I infer that customers do on the order of 2.6 purchases in their first year and 3.5 purchases in year two.
2. I know I'm repeating myself, but I can't stress enough how strong the repeat purchasing is here. Here's one other way of framing it. DoorDash's new customer acquisition was down 3% from Q1 2019 to Q4 2019. Over that same period of time, revenue growth was still 100%+. That means all that growth, and then some, was purely due to strong repeat purchasing.

This is exceptional repeat buyer behavior. It implies open ended upside should DoorDash be able to continue developing these customers over time. Again, I didn't allow purchase frequency to grow as a function of customer tenure for more than three years, but this could prove conservative.

COVID: big tailwind for both repeat buying and acquisition

COVID has been a boon for many digital business (although the extent of the tailwind varies by company, as I've noted previously).

COVID significantly increased purchase activity. Averaging across the tenures of the alive customers as of September 2020, the weighted average increase in order frequency during COVID is on the order of 70%. That's a lot more repeat buying because of the pandemic...

... but the impact on customer acquisition has been even bigger still -- I infer an increase of 200%+ relative to what I would have inferred the company to do had we not experienced COVID. This is fully consistent with my prior note on COVID bumps, that the impact has been larger for acquisition than for repeat buying. To visualize the extent of the effect of COVID on customer acquisition, below is a chart that shows GOV from new customers -- actual, expected given COVID, and expected when we "shut off" the inferred boost to acquisitions due to COVID. The "COVID boost" is basically the uplift from the grey dotted line to the red dotted line:

Remember, before the pandemic, GOV from new customers was flat to falling.

New customer data - strange... and possibly incorrect?

The insights above could be skewed somewhat, though, because there is something off with this chart:

At first glance, it seems like a perfectly normal chart. It's one of the most informative of all the charts that DoorDash included in their S-1. But there are a couple of aspects of it that don't pass the smell test.

ISSUE ONE: Did DoorDash get 15% or 22-23% of GOV from new customers in Q3 2020?

First, there is the fact that they say 15% of GOV came from new customers in Q3 2020, which is in conflict with this chart. The easiest way to "see" this is to note that DoorDash did $7.25B in GOV in Q3 2020. 15% of $7.25B is $1.1B. It is visually obvious that the height of the rightmost light green bar is above this (see below).

Taking this logic a step further, we can easily show by comparing relative heights of light green bars in that chart (i.e., GOV from new customers) to the corresponding overall heights of the bars (i.e., GOV in total), that while it is indeed the case that 32% and 23% of total GOV came from new customers in Q3 2018 and 2019, respectively, 22-23% of GOV came from new customers in Q3 2020, not 15%. The figure below was my way of making both of the aforementioned arguments at once (underlying file here):

You might be thinking "okay, okay, I get it -- sales from new customers didn't go from 32% to 23% to 15% of GOV over the prior three years. So what?" And on some level this is fair - what we can clearly see from the customer acquisitions validation chart above is that DoorDash got a huge boost of GOV from new customers due to COVID. No kidding that could increase the proportion of sales from new customers -- that's a very benign explanation if I ever saw one!

What I don't understand is that this is not the argument that DoorDash is giving in its filing. In its filing, it plays up the story that a growing proportion of sales from repeaters is driving the percentage of sales from new customers down, when that does not appear to be the case during COVID. This has the potential to confuse. It also creates more issues that I expand on below.

ISSUE TWO: The new customer GOV figures don't make sense

This also segues me into my second issue with this chart -- it cannot possibly be right.

Everyone in the industry -- including companies such as Wayfair and DoorDash's competitor, GrubHub -- defines sales from new customers in a period to be equal to total sales from the very first order that newly acquired customers during that period place.

It then follows that this equation must be true:

[Total sales from new customers] = [Total customers acquired] * [Average GOV per initial purchase from acquired customers]

Has to be. That's just math.

Well, we know total sales from new customers -- if you use pixel tracking software, you can easily back this out from that chart. This means that if we can get a read on average GOV per initial purchase, we can back out what total customers acquired was.

We can easily see through alternative data sources that GOV per purchase from customers' very first purchases are more or less the same as for repeat purchases, so it would not be unreasonable to assume that initial and repeat AOV's are about the same. If you did though, you get an almost implausible estimate of total customer acquisitions over time:

Remember, this is a company that gets 99.8% of its sales from the US only. How could they have acquired 209M customers in total since Q1 2018? There are only 328M people in the US, 209M of whom are 18 years and over.

Even when we allow for the possibility that some customers have multiple accounts, this doesn't pass the smell test. There is some evidence of multiple accounts with the Dashers. Could it be that there is this much abuse of "new customer referral program" discounts, with people referring themselves via different emails? Possible, but this would be a heck of a lot.

One way of thinking about just how rampant such abuse would have to be, focus on Q3 2020 for a moment. They did about $1.6B in GOV from new customers this quarter. This implies ~52M new customers acquired. But in their S-1, they say that active customers in the month of September 2020 was "over 18 million." Despite all that aforementioned repeat business from existing customers, with 52M customers being acquired from July to September 2020, only 18M customers were active in September? Even if we were to assume that customers acquisitions happened equally throughout the quarter, that would imply 17.3M new customers acquired (and thus active) in the month of September, before taking into account anyone else who might have made a purchase that month. This just doesn't pass the smell test.

I would love to understand the explanation for this from DoorDash.

If these figures are anywhere remotely close to the truth, this could imply that DoorDash has penetrated a significant fraction of their target market. This would be consistent with the trends that we saw pre-COVID, with new customer acquisitions falling throughout 2019. It would not bode well for new customer acquisitions post-pandemic. It would also imply that new customers acquired may be more price sensitive than the average customer that they have acquired thus far to date, because they have probably already acquired the less price sensitive people (and that even these customers have only been brought in, to some extent, through more than one new customer promo). Of course, it would also imply that there are a lot of unprofitable "one and done" customers that aren't ever coming back.

We need to get to the bottom of this -- it has significant implications for their future growth prospects.

CLV - strong ROI, middling dollar amount

Holding all of the above aside, it is still clear that repeat buying is excellent at the company, and we can still form assessments about the overall health of their business from a CLV and marketing ROI standpoint.

These are the main data points that I infer:

1. DoorDash spends about $6 to acquire each new customer. [ Note: DoorDash notes themselves that they spend upwards of 10% of GOV on sales and marketing in year 1. As noted above, initial AOV is around $30, but this is a mix of new and repeat purchases, which is why CAC must be higher than 10% - $6 is 20% of that initial order. ]
2. When I reverse out the beneficial impact of COVID, I infer that customers place about 14 orders on average in their first 3 years with the firm, but again, these orders are back end loaded because order frequency goes up with customer tenure.
3. DoorDash's gross margin has been improving over calendar time (i.e., all customers' gross margins are expanding over time) and as a function of customer tenure (i.e., customers' gross margins are increasing as customers age). While their gross margin in year one is ~2%, it rises to about 12% by year three. Again, very back end loaded.
4. DoorDash spends about 2.3% of GOV on sales and marketing for repeat orders. This has been fairly consistent over time.
5. Netting gross margin against sales and marketing for repeat orders, this implies DASH's contribution margin grows from about 0% in year one to about 10% by year three.
6. Present valuing this stream of variable profits, I estimate a three-year "post-acquisition value" of about $21. Subtract off the $6 of CAC, and we get a $15 CLV, implying a marketing return on investment of 250% in just the first three years.
7. When we extend the horizon on our customer profitability calculation from three years to five years, my estimate of their post-acquisition value goes up to ~$60, implying a CLV of ~$54, or a ~900% marketing ROI.

This is a very strong return on investment. And if DoorDash can continue holding onto and monetizing these customers in future years, those PAV estimates continue to go up and up.

That said, I would be remiss if I did not mention that the absolute dollar amount of this return is not all that large. A five-year CLV of $54 is not a whole lot of value in an absolute sense. This is primarily due to the very thin margin profile of the firm. What sort of overall corporate valuation does this imply? To justify a valuation of $25-30B+ when a lot of the core market has already been acquired, it would seem like we may need to see something more -- a transformative new vertical, or a way to dramatically expand TAM.

To their credit, we would fully expect some of these positive dynamics to play out -- adding more consumers, Dashers and restaurants to the platform will only further expand their lock-in with all of them. That is basic marketplace economics 101. Also, DoorDash is moving into adjacent verticals like groceries, which could further bolster retention and order frequency for everyone in the ecosystem.

Finally, if that new versus existing customer GOV chart is wrong, and DoorDash has acquired a far smaller number of customers, then it could very well be that each customer is a lot "bigger" -- bigger CAC and (potentially even bigger) PAV, implying a larger CLV. My back of the envelope math would suggest cumulative acquisitions may be closer to 50-55M. If so, then DoorDash's CLV figures are 4+ times higher (e.g., a 5-year CLV of $216 or more).

In sum, there is a lot to like here, but a lot of open questions, some of which revolve around the possibility of a data error in the S-1 filing.

Acknowledgments: shout out to Travis Mays, who called this out as we were collecting their SEC filings data.