A Customer-Based Valuation Analysis of Overstock and Wayfair

More and more people are talking about the health of the customer base in the context of company valuation for subscription firms. Blue Apron is a notable recent example, but it was also heartening to hear the CEO of Raised Real emphasize his company's impressive retention rate as a sign of the health and viability of the business (link). It was also heartening to hear the head of a marketing firm emphasize the importance of customer acquisition and retention to Stitch Fix's valuation in their upcoming IPO (link). One common denominator behind all of these examples though is that they are all subscription-based firms. As this article notes, PE and VC investors love subscriptions, and part of the reason why is that it is easier to measure customer value in subscription settings.

But what about non-subscription firms? There are far more non-subscription firms than subscription ones, and the health of the customer is no less important to the viability of the firm. While getting a pulse on the customer is as or more important for non-subscription businesses, we have seen a lot less progress here model-wise, and for good reason -- it's harder. We don't get to observe when customers churn, and customer ordering and spending per order can be a lot more variable across customers and over time.

Wharton professor Peter Fader and I have been hard at work to come up with a customer-based valuation model that is the non-subscription analog to the methodology that we had built to value and analyze subscription firms. We're very excited to finally share the results with you -- our paper proposing a method for customer-based valuation for non-subscription firms is available here. We apply the methodology to two publicly traded companies, Overstock.com and Wayfair, valuing both of them and then comparing the unit economics of newly-acquired customers.

While I think you might find the whole paper to be interesting, it is over 50 pages long, so I thought it would be helpful to quickly summarize our main findings here. To sum it up in one sentence, we found that Overstock seems a bit undervalued (we estimate a fair value of $16.88, ~9% above their then-current stock price of $15.50, with a 95% valuation interval between $7.2 and $26.9) while Wayfair appears heavily overvalued (we estimate a fair value of $10.24, ~84% below their then-current stock price of $64.16, with a fair value above $45 being highly unlikely), and the red herring is very challenging unit economics (for customers acquired in Q1 2017, Overstock earns a profit of about $9 per acquired customer, while Wayfair loses about $10). I unpack this sentence below.

What we did

We fit our statistical model to publicly available customer data from both companies (see the Appendix of the paper for the OSTK and W datasets), which drove our forecasts for future revenues at both firms. We then drove discounted cash flow valuation models off of those revenue forecasts, turning those revenue forecasts into stock price estimates. In the spirit of full transparency, we list out in detail every assumption we made in both DCF models.

While those point estimates would be our best guesses of what OSTK's and W's respective valuations should be, we want to embrace the fact that our point estimates could be wrong. We used a bootstrapping procedure to create alternative revenue projections which account for uncertainty in our models' fits, which provided us with valuation distributions. The more uncertain we are about our forecast (e.g., if our data is limited, as it is with Wayfair), the wider and more disperse the valuation distribution.

Valuation Results

We want to make sure our model fit the observed data reasonably well. Overstock's expected and actual quarterly customer acquisitions (top row), total orders (middle row), and total revenues (bottom row) are shown below:

These are the corresponding plots for Wayfair:

Do we capture every single wiggle and jiggle? No. But we capture the baseline trends across all metrics very well, which is all the more impressive because the only covariate we allowed in the model is for Q4 seasonality. Occam's razor applies in spades when the data is so limited -- it is highly questionable whether our out of sample predictive accuracy would improve if we encumbered the model with a slew of additional covariates, even though they would surely make the plots above look (marginally) better.

Overstock's revenue forecasts are shown on the left below, while the resulting stock price distribution is shown on the right:

These are the corresponding plots for Wayfair:

We estimate that Wayfair's fair value is $10.24, 84% below its then-current stock price. At the same time, there is no doubt that the future is more uncertain at Wayfair. While our baseline expectation is that quarterly revenues peak at $2.3B in Q4 2022, more than double its current level, there are bootstrapped realizations of peak quarterly revenues exceeding $12B. Even in this most optimistic revenue scenario, however, the resulting fair valuation for Wayfair’s stock is $57, which is still approximately 11% below Wayfair’s observed stock price. These results are also robust to assumptions regarding Wayfair’s future margins (our base case scenario is that their margins will rise to the midpoint of long-term margin goals that Wayfair management has provided on recent conference calls within five years) and discount rate (our base case scenario is equal to Bloomberg’s estimated Q1 2017 WACC figure).

In contrast, using the exact same methodology, we find that Overstock's fair valuation was 9% above its then-current stock price.

Comparison of unit economics

While we have done the best job we can to incorporate uncertainty into our projections in a data-driven way, I am sure that there will be many people who are still skeptical of these projections... or any projections that don't align with their investment theses. An important diagnostic that makes no assumptions whatsoever about future customer acquisitions is simply looking at what these figures imply for the profitability of newly acquired customers. Properly accounting for the time value of money, this gives us what the lifetime value is of customers today. We compared Overstock and Wayfair in this way, looking at the initial customer acquisition cost (CAC) to acquire the customer compared to the ongoing stream of variable cash flows after the customer has been acquired, which are estimated in the following way:

• We assume that CAC is equal to the trailing 12 month ratio of total advertising expense to the expected total number of customers acquired, consistent with statements made by both firms and with prior academic literature. This also ensures consistency in CAC across both companies, as they both disclose these figures.
• We operationalize marginal profit after the customer has been acquired to be equal to earnings before interest, taxes, depreciation, amortization, and acquisition expenses, or "EBITDA-CAC." These cash flows are easily derived from our model by obtaining the expected quarterly revenues associated with one customer acquired at the end of Q1 2017, then multiplying these future expected QREV figures by the firm’s overall ratio of expected EBITDA-CAC to total revenues.

The results are shown below:

Wayfair customers generate more profits after they are acquired – the net present value of future profits after acquisition are $59 and $47 per customer at Wayfair and Overstock, respectively. However, Wayfair spends far more than Overstock to acquire new customers. Wayfair’s CAC is $69, nearly double the $38 that Overstock is spending to acquire its customers.

While one could argue that some of Wayfair’s advertising expenses are earmarked for customer retention, the proportion is likely to be small, and because Overstock is a relatively more mature business, its corresponding proportion is likely larger than Wayfair’s.

If Wayfair were able to reduce its CAC to Overstock’s level, we estimate that Wayfair’s expected valuation would more than double, all else being equal. Of course, all else is not equal – Overstock is pursuing a more conservative customer acquisition strategy, acquiring a smaller number of higher lifetime value customers. We estimate that Overstock earns approximately $9 per acquired customer, while Wayfair incurs a loss of approximately $10 per customer in Q1 2017. While we anticipate that the unit economics of Wayfair’s newly acquired customers will improve in the future as their variable contribution margin is expected to expand, challenging unit economics are a reality for the business, and are an important driver behind their relatively modest valuation.

Wrapping up

This sums up the most important aspects of the analysis that we performed. For all the gory details, please see the paper. The health of the customer matters and the markets are waking up to this fact -- Wayfair's stock is down 11% (or $800MM in equity value) since we posted this paper with no other news except for a couple of tweets from Citron Research about this work (link, link). See chart below. Their stock fell by 8.8% today, which is the biggest one-day stock price decline in over a year. But as with the customer, it's the long term, and not the short term, that matters most.

Miscellaneous Comments

This is an FAQ of sorts as we've gotten a flood of great questions and comments. This is an evolving section, so please check back for updates!

Baseline revenue expectations: We do not suggest that the online furniture retail category will peak in 2022. A few key points:

• Uncertainty: Our baseline revenue expectation is that revenues peak at $2.3B, and if there is one thing that is certain, it is that this figure will be wrong, necessitating the revenue distribution. There are "worlds" where quarterly revenues peak at $12B, and the stock is still overvalued in those worlds, but by a lesser amount -- we estimate fair value to be more $45-50 in very high revenue scenarios (versus the then-current $64 price and the current price of $75). Similarly, the quarter of peak revenues is uncertain -- there are worlds in which peak revenues is all the way out in 2029. It is helpful to think in terms of distributions and not point estimates.
• Customer retention: (a) Retention/development is perhaps the biggest “blind spot” for models that are not at the customer-level. Our statistical model explicitly estimates retention/development. Once you’ve acquired a cohort of customers, they are “leaky bucket.” Good businesses have smaller holes in that bucket. (b) Case in point: Amazon. For their retail business, Amazon’s customer acquisitions have almost surely already peaked (at least in the US). What is keeping their retail sales so strong is retention and sales for retained customers. Amazon has done an absolutely stellar job on retention/development. I became a customer in 2006 and still regularly buy 11 years later. They have a huge number of customers like me. (c) Purely looking at Wayfair's acquisitions and active customer counts in conjunction with revenues, the model suggests that Wayfair is very far from Amazon-like in terms of retention/development. I have read many qualitative arguments for why this might be (e.g., good for discovery but less good for distribution; SKU count vs Amazon; pricing). One simple but poignant data point - all the way back in Q1 2000, Amazon had 76% of total orders coming from repeat customers despite the fact that Amazon was experiencing rocket ship acquisition growth at the time (check out this image of the reference from an old SEC filing). At Wayfair, even here in Q2 2017, that proportion is still only 61.3%. So once customer acquisitions hit peak and then fall over, there is relatively little to stop revenues from falling. Which brings us to customer acquisitions...
• Customer acquisition: (a) The standard way that a statistician like me models long-term customer acquisition is as a “diffusion process,” using the same sort of models that an epidemiologist might use to model the transmission of a disease across a population (but here, it is “transmission of a company”). The data tells us how this evolution will occur through the model. (b) To keep the disease analogy going, the size of the population is not infinite, and there are a number of people who are going to be immune to the disease. The "effective ceiling" for a disease can be more limited than one might think. The effective ceiling at Wayfair is very high for an e-commerce retailer -- our model implies that 41% of the combined US, UK, German, and Canadian labor forces will eventually be acquired. (c) Just like a disease, the more contagious it is, the faster it rips through the population, the faster it drops as it begins to saturate its potential market. Wayfair's acquisitions data suggests that adoption is very "contagious" (i.e., adopters are heavily influenced by the number of people who have previously adopted), which implies that acquisitions are likely to fall quickly when it starts saturating its market (somewhere between 2022 and 2029). (d) This is not to say that those people who have already been acquired will stop buying it (see the retention bullet point above this one), just that the number of newly acquired customers, and the revenue associated with them, will.

Margins: We project margins to be the midpoint of Wayfair's long-term margin forecasts, not half of their margin forecasts. In other words, we project that their Adjusted EBITDA margin will rise to 9% of sales and then remain there... indefinitely. This is not a pessimistic assumption.

Bias: The cries of "you're biased" were probably inevitable given the circumstances. A few points on this: (a) Peter and I have never had a long or short position on either company, so we are not "talking our books". (b) Peter and I have also received no payment, direct or indirect, from anyone with a bearish opinion on either stock. Happy to do an audit if you would like to confirm it. (c) The underlying paper is being sent to a well-known academic journal (the Journal of Marketing Research), where there is no incentive to make extreme stock price projections... in fact, the opposite incentive may exist (note too that 3 of the 4 valuations Peter and I have done in academic papers were quite consistent with then-current stock prices).

Horizon for CLV Calculation: To calculate CLV, we get the net present value of all the future cash flows associated with a newly acquired customer, and those future cash flows extend out indefinitely (this is also consistent with the unit economics chart from above -- there are future cash flows in many future quarters, not just the first). It would heavily undervalue CLV to only consider CAC and the cash flow in the next quarter -- we certainly do not assume this.

Our paper versus Gupta, Lehmann and Stuart ('GLS'): GLS's undervaluation in previous work is one of the big reasons why we wrote this paper in the first place. More work is needed in this non-subscription setting! We stress exactly this point in the introduction of the paper. There is further evidence of this point in our comparison against alternative models (Figure 3, line for GLS) -- once again, GLS implies an extremely pessimistic valuation relative to ours. Pointing to GLS as a reason to question our results is like pointing to a car with unappealing features as a reason to question a car that fixes those features.

Wayfair's CAC / LTV Calculation. Wayfair offers their own LTV-like calculation (see Slide 36 in this presentation). While our CAC figures are largely in agreement, their annual revenue per customer figure is not appropriate / not the right figure to be comparing CAC to, to come up with a view on LTV versus CAC. That revenue per customer figure is not total revenue from a newly acquired customer over the next year - it is total direct revenues divided by total annually active users (AAU). Total revenues includes a lot of revenue generated from existing customers, and likewise, AAU is every single customer, new or old, who placed 1+ order in the previous 12 months.