Marketplace Lending, Information Aggregation, and Liquidity

Abstract

The remarkable success of online lending platforms, which directly connect retail lenders with corporate or consumer borrowers, has led many to believe that a radically new model of credit finance is on the rise: decentralized, transparent, disintermediated, and more information efficient. That is, there is the potential for a major disruption of traditional business models, particularly banks, and a redrawing of the boundary between markets and institutions.¹

Supporting this view is a growing body of evidence showing that much information is dispersed across retail investors and that, if aggregated properly, it can significantly improve the pricing of consumer and corporate loans. For example, Iyer et al. (2015) study auction prices from Prosper, a peer-to-peer (P2P) platform, and demonstrate that they predict default better than credit scores do. Hence, they conclude that “our results highlight how aggregating over the views of peers and leveraging nonstandard information can enhance lending efficiency” (p. 1). It is, therefore, puzzling that many lending platforms, including Prosper, have abandoned their original auction design in favor of posted prices (see Wei and Lin 2017).² If auctions enhance information efficiency, why were they abandoned?

We investigate this puzzle using hand-collected data from the United Kingdom’s leading peer-to-business (P2B) platform, Funding Circle (FC). The data contain detailed information on |$7,455$| multiunit auctions of small and medium-sized enterprise (SME) loans executed online between 2010 and 2015. We observe every order that was submitted to the platform (⁠|$34$| million in total), whether ultimately accepted or rejected. Investor ID numbers allow us to study individual bidding patterns and how they are incorporated into the price discovery process via an open order book. Like Prosper, in 2015, FC abandoned its auction design in favor of posted prices, and, in September 2017, it further limited investors’ choice to platform-selected loan portfolios. FC is a highly successful operator, and, by 2017, its market share exceeded 50% of a P2B industry that was funding around |$29.0\%$| of new U.K. SME loans (see Zhang et al. 2018).

We suggest a resolution to this puzzle using three interrelated findings. First, in spite of enhanced predictive power, auction prices often deviate, temporarily but significantly, from information efficiency. We provide evidence that the source of these random deviations are liquidity shocks arising from mismatches between flows of funds into and out of the platform, in line with Duffie’s (2010) slow-moving capital. Second, to assist a substantial number of investors who lacked the time and expertise to actively participate in the auctions, FC operated an algorithmic “autobid” to which “passive investors” could delegate the submission of their orders. However, our results indicate that FC was not successful in optimally calibrating the autobid, leading to a decrease in information efficiency. Third, the quality of pricing deteriorated over time. We associate this development with a “vanishing crowd” phenomenon, whereby large investors and the autobid came to dominate the auction process.

To the first finding, we estimate an efficient market hypothesis (EMH) equation in which the dependent variable is a credit default dummy, and the explanatory variables include the borrower’s credit scores and the loan interest rates, as determined by the auction. We adjust the specification to deal with the truncation of our performance data so that some loans mature out of sample. We find that the interest rate coefficient is strongly significant, indicating that prices enhance default predictability, over and above credit scores, similar to Iyer et al.’s (2015) findings. At the same time, we reject key EMH predictions. First, we reject the hypothesis that the interest rate, adjusted for the loss given default, reaches the information-efficient benchmark. Second, we find that the credit scores retain predictive power, contradicting the EMH prediction that all of their information should have been absorbed into the interest rate.

In order to identify the source of deviations from information efficiency, we augment the EMH regression with various proxies for liquidity shocks. Under the null EMH hypothesis, these proxies should have no power to predict default because all relevant information should be absorbed into the price. Hence, once the EMH regression controls for the price, any other variable, whether originally containing information or not, should prove insignificant. In fact, we find that liquidity shocks still have predictive power, which can be interpreted as deviations from information efficiency.

To deepen our understanding of the way in which liquidity shocks can drive prices away from fundamental values, we analyze two liquidity events that, although known in advance, still result in considerable mispricing of loans. The first event involves the March 2014 termination of a £20 million U.K. government program, which contributed a fixed |$20\%$| stake in each auction. Consistent with slow-moving capital, the price effect was felt 10 days after the event, once existing market liquidity had dried up, and persisted for an extra |$20$| days, until new capital started flowing into the platform. A second event relates to the predetermined closing hour of auctions. We show that auctions closing outside of the peak hours, |$4$| p.m. to |$7$| p.m., closed at interest rates above the information-efficient benchmark. This is in spite of the fact that an auction’s closing time is perfectly anticipated, conditional on the randomly allocated opening time.

The second finding is that FC was not able to optimally calibrate the autobid. Over the sample period, on average, |$48\%$| of the funding was allocated via the autobid. This, in itself, should not have been a problem: if investors have no information, it is better to delegate bidding to an algorithm so that it can play a role similar to that of a market maker or an initial public offering (IPO) arranger (cf. Kyle (1985) or Biais and Faugeron-Crouzet (2002)). To perform this function effectively, the autobid has to be calibrated so that it prices in information contained in some funding flows while smoothing out the liquidity shocks. A detailed analysis of the autobid’s activity reveals that this was not the case. On the one hand, the autobid insufficiently smoothed time series fluctuations in aggregate supply and demand for funding, thereby allowing liquidity shocks to drive the price away from information efficiency. On the other hand, the autobid overly smoothed cross-sectional variations by channeling excessive funding when low levels of active bidding should have signaled a higher default probability, thereby preventing relevant market information from being adequately incorporated into the interest rate. This result highlights the difficulties that FC designers faced: an information-efficient autobid required detailed knowledge of the joint distribution of funding flows and default probabilities, so as to “signal extract” the information from the funding flows. The source of such knowledge is difficult to guess, given the relatively small sample of loans, the small magnitude of the event to be estimated (the annualized default probability of an A-scored loan is just |$2.9\%$|⁠), and the considerable lag with which the default and recovery takes place.

Deviations from price efficiency explain at least |$80\%$| of the price variance. Interestingly, the nature and magnitude of the problem are not dissimilar to that found in mature corporate bond markets. A study by Collin-Dufresne, Goldstein, and Martin (2001) reports that “regression analysis can only explain about 25 percent of the observed credit spread [monthly] changes.” In addition, they find that “the dominant component of changes in monthly credit spreads in the corporate bond market is driven by local supply/demand shocks that are independent of both changes in credit risk and typical measures of liquidity.” ³

Regarding the third finding, we document a sharp drop in the predictive power of the interest rate over the sample period, which we relate to the changing composition of the investor population. We document a “vanishing crowd” phenomenon; that is, through the combined effect of declining active investment and increasing activity of large investors, only |$25\%$| of the funding was crowd-allocated by the end of the sample period. To interpret the result, we draw on recent literature on household finance, which suggests that large investors do not increase information efficiency. Using Swedish data, Bach, Calvet, and Sodini (2020) find that large investors earn higher returns that can be fully explained by greater risk-taking, rather than by “informational advantages or exceptional investment skill, [which] contribute only marginally to the high returns of the wealthy.” Similarly, we document that large investors earn an extra |$1\%$| return on their loan portfolios relative to small investors, but that a substantial part of that extra return is generated simply by higher risk-taking. Integrating the effect into our EMH framework provides evidence that the decline in the quality of the price is correlated with the changes in investor composition.

While sophisticated investors can diversify away deviations from information efficiency and might even profit from the arbitrage opportunities that they create, a nondiversified SME borrower would have to bear the cost of overpricing for the entire duration of the loan, which, in our sample, has a median length of 3 years. This observation is consistent with the reasons given by FC on their website⁴ for its decision to replace auctions with posted prices: “(i) businesses are put off by a lack of certainty around the cost of their loan, which is important to them; (ii) the price of each loan will now be based on the risk (and term) of the loan, rather than the availability of investor funds; and, (iii) borrowers will know how much their loan will cost before the funding process, attracting more businesses to Funding Circle, which will create more lending opportunities for you.”

We believe that our results can provide insights into the development of online marketplace lending. They shed light on common concerns of regulators about liquidity provision and the changing nature of the investor population across lending platforms in the United States and Europe. A U.S. Treasury (2016) white paper on online marketplace lending specifically addresses their potential liquidity risks and notes that “ongoing research will be necessary to monitor the liquidity of online marketplace lending and its impact on credit markets.” Across lending marketplaces, the issue of liquidity is intertwined with the phenomenon of “vanishing crowds” documented in our paper.⁵ Our paper, therefore, connects these global trends with the design of online marketplaces.

While our analysis explores the comparative advantages and disadvantages of the P2B innovation relative to traditional models of credit intermediation, we believe that it is too early to conclude that P2B lenders will displace traditional commercial banks. In their favor, banks bridge the liquidity shocks that affect P2B pricing, but, in doing so, they have to engage in liquidity transformation. However, liquidity transformation is accompanied by credit risk and costly regulation, including capital requirements. In contrast, the instant clearing of investors’ funds against accepted bids avoids FC’s exposure to credit risk – i.e., “securitization at origination,” which virtually eliminates any burden of regulation.⁶ Substantial improvements and refinements of the new model are required before its disruptive potential can be fully assessed.

The first strand of the literature relevant to our work highlights the diversity in current platform designs, typical of an innovative industry in the early stage of development. Like FC’s platform, these designs experiment with alternative ways to combine algorithms and human investors. Vallee and Zeng (2019) explore, theoretically and empirically, the connection between the information provision on P2P platforms and rents extracted by sophisticated investors. Their study is motivated by the decision of Lending Club, a P2P platform, to remove half of the 100 variables on borrower characteristics that it previously provided to its investors. D’Acunto, Prabhala, and Rossi (2019) study the implications of robo-advising for the portfolio choices and performance of investors in the Indian stock exchange. They document that the adoption of the delegated investment mechanism has heterogeneous effects across investors, with benefits decreasing in the amount of portfolio diversification. Grennan and Michaely (2019) study the operations of FinTechs that aggregate and synthesize public data. They find a reduction in the quality of information produced by online financial analysis and, as a result, a deterioration in information efficiency. Finally, several studies show how the design of peer-to-peer marketplaces affects the matching between borrowing households and contract terms (Hertzberg, Liberman, and Paravisini 2018; Cespedes 2019; Liskovich and Shaton 2017).

The second strand of literature documents the significant amount of “soft and nonstandard” information dispersed among investors about borrower default prediction. Lin, Prabhala, and Viswanathan (2013) demonstrate that, due to their lower default probabilities, borrowers with friends are more likely to have their credit application approved and on better terms. Lin and Viswanathan (2016) explore the effect of physical proximity – home bias – on the propensity to lend. Butler, Cornaggia, and Gurun (2017) explore the relationship between the availability of local bank credit and the propensity to borrow from a P2P network. Ravina (2019) finds that “beautiful borrowers are 11.7% more likely to get a loan, pay similar interest rates as average looking borrowers with the same credentials, but default more often.”

1. The Data, the Platform, and the Auction

In this section, we provide a description of our data, the design of the FC auction and the price discovery process. We also place the auction design in the context of the theoretical literature.

1.1 The data

Our data include all loans, 7,455 in total, auctioned between 2010Q4 and 2015Q1. During that period, the aggregate value of FC’s loan book grew at a weekly rate of |$2.4\%$|⁠, albeit erratically, with a standard deviation of |$1.2\%$|⁠. For our sample, loan maturity ranges from 6 months to 5 years with a median of three years (see Table 1). Loans are amortized in equal monthly payments. We track these repayments to the end of 2016. While we lack full performance data for |$42\%$| of the loans in our sample, we do have at least one-and-a-half years of performance data on each loan. Section 2 describes the method we employ in order to resolve the sample truncation problems that arise in the estimation of default probabilities and LGD.

Loan size varies from £5,000 to £|$0.52$| million, with a median of £50,000. According to the borrower’s own report, the main purpose of the loan is to fund working capital, growth or new investments. The vast majority of borrowers are organized as limited companies. They come from all regions of the U.K. and from all sectors of the economy. The data set contains no information about borrowers’ bank relationships at the time of the auction. However, since the median SME borrower in our data is nine years old, it seems safe to conclude that he did have substantial banking relationships at the time. Baeck, Collins, and Zhang (2014) report survey results regarding borrowers’ perception of the advantages offered by P2B funding: more willingness to take risk; higher speed of approval; and greater access to funds, with a marginally lower cost of borrowing. It is widely believed, however, that banks have failed to provide adequate funding to the SME sector (see the Breedon Review 2012). In response, the U.K. government decided to support the development of alternative models of financial intermediation. In March 2013, the Department of Business Innovation and Skills (BIS) created a £20 million fund to be allocated via the FC platform.⁷ This experiment provides an event study opportunity, which will be described in Section 3.3.

A borrower’s loan application is first evaluated by FC’s own credit department. Some borrowers are rejected immediately due to suspected fraud or an unacceptably high default probability. Those that are approved are assigned a credit score: A+ for the lowest risk and D for the highest. The analysis is based on hard information, including the borrower’s Experian credit rating, its credit history and its financial statements. The credit scoring process allows for a certain level of discretion by FC’s analysts. Borrowers are asked to provide a “prospectus,” which is made public on the platform’s website. While the auction is running, the platform usually opens a ledger where investors can post questions or ask for additional information. Borrowers are encouraged to respond fully to questions.

Orders submitted to the platform must be backed by funds that are transferred in advance to a special FC account. If a bid is accepted, that account is immediately debited, while the purchased loan parts are simultaneously transferred to the investor’s custodian account. Such instantaneous settlement implies that at no stage in the process does FC have ownership of any part of the loan, and, hence, there was no need to allocate any (regulatory) capital against credit risk. The process might be described as securitization at the point of origination.

Once the loan is issued, FC collects the monthly repayments and distributes them among the investors who funded the loan. In the event of default, FC acts as a Diamond (1984) delegated monitor to manage a resolution, including, if necessary, litigating on behalf of the investors; the mean number of investors per loan is |$571$|⁠, with a median of |$459$| (see Table 1). For these services, investors are charged a |$1\%$| annual fee on the outstanding principal of performing loans, while borrowers are charged a one-time issuance fee that is a function of loan maturity and the borrower’s rating.

1.2 Auction design and function

In a multiunit auction, each investor submits orders for a part of the total amount of the loan. The minimum loan part that the platform accepts is £20. The mean number of loan parts per loan is 805, and the median investor stake, as a proportion of the value of the loan, is |$0.48\%$| (see Table 1).

Only limit orders are accepted; each order must specify a quantity and a price. Auctions run continuously, day and night. Typically, auctions are scheduled for 7 days (⁠|$168$| hours), but a few auctions are allocated more time. It follows that, conditional on the opening hour, there is perfect foresight of the closing hour, an important fact in the event study analysis of Section 3.3 below. However, borrowers may terminate the auction prematurely, in which case the terms of the loan are determined by the orders submitted at that point.

FC auctions price discriminate, so that each accepted order pays the submitted interest rate. At any time during the auction time, orders submitted up to that point can be sorted by price to form an increasing, stepwise supply curve. Figure 1 presents such supply curves, for an arbitrary A-scored loan |$24$|⁠, |$96$|⁠, and |$168$| hours from the open. Orders are normalized by the size of the loan, in this case £15,000. By construction, given that the size of the loan is fixed in advance, the normalized demand curve is just a vertical line at one unit. When the auction closes, the intersection point between supply and demand, in this case |$6.6\%$|⁠, defines the marginal closing interest rate, namely, the highest interest rate of any accepted order. All orders submitted at an interest rate below the marginal rate are accepted, while orders submitted at an interest rate higher than the marginal rate are discarded. Orders tied at the marginal rate are prioritized by submission time. It follows that the “borrowing rate,” that is, that which is charged to the borrower, is a weighted average of all the accepted orders, in this case |$6.49\%$|⁠. By construction, the borrowing rate always lies below the marginal closing rate.

Figure 1

Supply curves, normalized by loan size, 24, 96, and 168 hours after open

Orders, submitted up to the respective points in auction time, are normalized by loan amount, £15,000 in this case, are sorted and then plotted against the submitted interest rate. By construction, demand is represented by a vertical line at one unit.

Open in new tabDownload slide

Another important property of the FC design is that orders are not retractable: once submitted, an order cannot be withdrawn. It follows that the downward shift of supply curves over auction time, as depicted in Figure 1, is a necessary consequence of the design of the FC auction. Nonretractablity plays an important role in limiting investors’ ability to engage in manipulation (see Section 1.3 for a detailed discussion).

Many investors, particularly small ones, may not have the required time or expertise to actively participate in the bidding process. For such investors, FC operates an algorithmic bidding device, the “autobid.” On average, about half of the platform’s funding is provided by “passive investors” (see Table 1). The autobid selects a diversified portfolio to match the risk profile chosen by the investor. The number of “active” (non autobid) investors per loan remains very high, with a mean of |$200$|⁠. It is worth noting, however, that active investors do not participate equally. The mean share of the largest investor is |$8\%,$| while the mean share of the top-5 and the top-20 investors is |$18\%$| and |$29\%$|⁠, respectively. Section 4 provides a detailed analysis of the investors’ composition across size categories.

While the auction is running, the order book is open for the inspection of all interested parties, with the platform computing and reporting the marginal interest rate, continuously. The order book provides information on the identity of the investor⁸ who submitted the order and his bid in terms of the interest rate, the amount, and the time. The platform does not report whether the order is placed directly or via the autobid, although sophisticated investors can infer that information with some degree of confidence.

Table 2 reports mean daily order flows for both the autobid and active investors within |$24$|-hour intervals of the auction time, that is auction “days.” Only auctions that complete the full 7 days allocated to them are included. Daily orders are normalized by loan size and, then, averaged across auctions. Active investors are sorted by the total amount of their daily orders into size groups, small, medium, and large, £100, £100–£1,000, and £1,000+, respectively. The autobid is highly active in the opening hours, and it is quite common for it to submit orders in excess of the amount of the loan. However, |$61\%\left(=1-0.25/0.65\right)$| of the day one autobid orders (by value) are eventually rejected (see variables Flow and Exec.,Table 2). It follows that the autobid is programmed to submit, on day one, a large number of small orders at different interest rates, whereby higher ones are ultimately rejected, as they end up above the marginal interest rate at the close of day seven. Sorting these day one orders by the submitted interest rate, from the lowest to the highest, we interpret the autobid policy as the submission of an upward-sloping supply curve, as shown in Figure 2. The slope, the intercept and the exact shape of this autobid supply schedule had to be calibrated by the designers of the FC platform. Much of the EMH analysis, in Section 3, addresses the question of whether that calibration was done in a manner that maximized information efficiency.

Figure 2

FC auction design

The diagram depicts a stylized exposition of FC’s platform design. Table 2 documents that autobid activity is largely concentrated on day 1; that is, it submits orders at different prices so as to form an upwards sloping supply curve. When we normalize all volumes by the size of the loan, the price insensitive demand curve is a vertical line at one unit. Orders submitted up to any point in auction time, by passive investors (via the autobid) as well as by active investors, can be crossed against demand to derive the marginal interest rate for that point in time. The shaded area represent the borrowing rate, which is a weighted average of all accepted orders. Since the order book is open, active investors who want to compete for an allocation have an incentive to undercut the marginal rate by just a few basis points. For more details, see Section 1.

Open in new tabDownload slide

Nothing compels active investors to place orders at the early stage of the auction: they can wait and bid just prior to the close, free riding on information of other investors without contributing any of their own. Clearly, if all investors were to adopt such a strategy, the process would degenerate into a sealed-bid auction. Table 2 demonstrates that this is not the case: by the end of day one, active investors have submitted orders totaling |$34\%\left(=1.11+0.18+0.05-1\right)$| in excess of the size of the loan. As for the autobid, most of these orders are ultimately rejected. The implication is that a substantial proportion of active investors voluntarily participate in the price discovery process. Their reason for doing so is beyond the scope of this paper.⁹ Suffice it to say that similar results are documented in the literature (see Biais, Hillion, and Spatt (1999) and Bellia et al. (2017) on participation in pre-open bidding on the Paris Bourse).

However, not all investors adopt such a cooperative bidding strategy. Variable New in Table 2 reports the new component of the order flow, namely, orders submitted by an investor who did not participate in the previous days of the same auction. It follows that |$49\%\left(=0.23/0.47\right)$| of the orders placed by large investors on the last day are new in that respect. The numbers for medium and small investors are |$61\%$| and |$79\%$|⁠, respectively. Hence, the pattern of entry into the auction is U shaped in auction time for all size groups: strong at the open, weakening in days 2 to 6, gaining strength again towards the close.

Given the open order book, the investors who participate in the price discovery process face a simple decision: whether or not to undercut the marginal interest rate, at that point in auction time, so as to guarantee that their bid is accepted. If many of them do, the marginal interest rate will fall. As a result, at the close, active investors’ accepted orders are clustered at or just below the marginal interest rate; the mean difference between accepted active bids and the marginal closing rate is |$37$| bps (see Table 1). In contrast, passive investors’ orders are spread out below the marginal closing interest rate, with a mean difference of |$101$| bps. For a diagrammatic representation of the borrowing rate, see the shaded area below the autobid’s supply schedule and the marginal rate in Figure 2.

1.3 Discussion: FC’s auction design and the EMH

Unfortunately, except for the largest platforms, the academic literature does not provide a comprehensive survey of the auction mechanisms used in online lending marketplaces. However, a leading FinTech blogger (“p2pmoney”) constructed a classification of U.K. and European online marketplaces and found that of 36 platforms, 14 employed nonuniform pricing in their auctions. In addition, price discrimination is a prominent mechanism in other financial auctions, notably the sale of Treasury bills. Brenner, Galai, and Sade (2009) survey 48 OECD countries and find that 24 use price discriminatory auctions; nine use uniform price auctions; nine use both; and six use other designs. Interestingly, empirical research fails to find significant revenue differences between discriminatory and uniform price auctions. Using occasions in which U.S. Treasury bills were sold simultaneously, via both uniform and discriminatory auctions, Nyborg and Sundaresan (1996) find only a modest difference in outcomes. Hortaçsu and McAdams (2010) take a structural approach, estimating bidding strategies in Turkish Treasury auctions and then using the results to simulate the effect of a regime change. They conclude that “by our point estimate, the switch from a discriminatory to a uniform price auction would lead to a gain of expected revenue that is at most 0.12 percent of the realized revenue.... However, taking sampling variation into account, we cannot reject the hypothesis that such a switch would lead to no difference in revenue” (Hortaçsu and McAdams 2010).

Quantity discrimination is standard in initial offerings of shares for public listing (IPO), which Biais and Faugeron-Crouzet (2002) analyze as an auction derived from a mechanism design problem. Their solution sets the price at a “significant discount relative to the market clearing price,” while rationing particular investors, so that the price does “not adjust to demand too strongly” (Biais and Faugeron-Crouzet 2002). Benveniste and Wilhelm (1990) model a closely related problem and derive a solution that price discriminates. One interpretation is that price and quantity discrimination are close substitutes, used in order to incentivize revelation of information but, at the same time, to guarantee a fair return to the uninformed so as to elicit their participation. In that respect, the FC design can be rationalized: while price discrimination delivers higher returns to active (and better informed) investors, the autobid caps these returns so as to protect the uninformed.

Price discrimination also can be used to avoid collusion in Walrasian auctions. Adapting a well-known argument by Wilson (1979) and Back and Zender (1993) to our context, investors can collude by submitting steep supply curves that clear the market above the competitive interest rate. Collusion is sustained because any unilateral deviation from that strategy results in a sharp fall in the interest rate, to the detriment of all investors, including those who have deviated, placing such behavior off the equilibrium path. Price discrimination breaks such collusion by imposing a heavy cost on the submission of steep supply curves: since each order that forms the supply curve is executed at a different interest rate, those orders that are executed at a low interest rate impose a heavy cost on the strategy.¹⁰

Collusion is closely related to the problem of price manipulation. Chakraborty and Yilmaz (2004) analyze the problem using a market micro structure setting. They describe manipulation as follows: “An insider who knows that the prospects of a certain asset are not good, might actually start buying the asset in order to drive its price up and then sell it without its price falling too fast.” Clearly, such “pump and dump” strategies require that investors can place both buy and sell orders. It seems that FC was aware of the problem and blocked such manipulation by making orders nonretractable. Thus, an investor cannot deter other investors from participation in an auction by lending heavily early on, thereby depressing the interest rate, but then hiking it up by retracting the early orders when the auction approaches closure. In other settings, similar strategies are implemented through short selling. Goldstein and Guembel (2008) provide a rigorous analysis of manipulation strategies and conclude that “manipulation is profitable only via sell orders... [which] is implicitly understood to be relevant, for example, by regulatory bodies [and platform designers]... who introduced restrictions on short sales.”

Market micro structure, as in Kyle (1985) or Glosten and Milgrom (1985), differs from the mechanism design approach in several respects. Notwithstanding the differences, two important commonalities emerge. First, to operate effectively, markets need intermediaries, such as IPO arrangers or market makers. Interestingly, Biais and Faugeron-Crouzet (2002) recognize that the role of such intermediaries could be automated, although they admit that it may prove difficult to “translate into [the] explicit computer algorithms the rather implicit rules” that human specialists have developed over many years of experience. A second commonality between the two approaches is that prices either reach or near the EMH price. Since matching the FC design with a structural model goes well beyond the scope of this paper, the EMH provides us with a convenient reduced-form conceptual framework to analyze the data. We will work out the precise specification in Section 3.1.

The main role of the (algorithmic) market maker is to provide liquidity, which requires that it has two attributes. First, the autobid must hold sufficiently large reserves, which it can inject when gaps arise between the flow of funds into and out of the platform. Second, the market maker must be able to separate random funding gaps from informative supply fluctuations. While the former requires the injection of more liquidity, the latter requires that market conditions are passed through into the price. That is, the autobid needs to be calibrated so as to smooth out liquidity shocks but allow flows that contain information to influence the price. However, to do so, autobid designers need to know the joint distribution of order flows and default probabilities in order to signal extract any information in those flows.

Failing to have sufficient liquid reserves results in deviations from EMH pricing until new capital flows into the platform, an effect that Duffie (2010) calls slow-moving capital.¹¹ Such temporary deviations from information efficiency are even more likely to occur in a rapidly growing platform, such as FC’s, which had to expand simultaneously into both the investor and the borrower populations. Perfect synchronization between these two processes is highly unlikely. However, we interpret slow-moving capital somewhat more broadly: not only is the platform growing rapidly, but the composition and the characteristics of the investor population are changing fast. It is difficult to know how prior knowledge of the joint distribution of order flows and default probabilities could be obtained in such circumstances.

Finally, large, wealthy and sophisticated investors could also provide liquidity to the system, side by side with the market maker. Indeed, in theory, it is difficult to make a clear distinction between the two, a point made in Kyle (1989). It follows, however, the these large investors are likely to face problems similar to those of the market maker that is, slow-moving capital and imperfect knowledge of the joint distribution of order flows and default probabilities.

2. Default Probabilities and LGD

The fact that our loan-performance data stop at the end of 2016 raises a truncation problem: a loan that does not default within the sample might still default subsequently, when we no longer observe it. Since FC loans are amortized in equal monthly payments, the timing of default plays an important role in our analysis, as it determines the pre-default recovery rate. A similar problem arises for post-default recoveries: the shorter the time we observe a loan post-default, the less likely we are to observe substantial recoveries. After correcting for these problems, we combine pre-default and post-default recoveries into the LGD rates, conditional on the credit scores. These corrections play an important role in adjusting the interest rates in Section 3’s EMH analysis.

To resolve the truncation problem in the estimation of the probability and the timing of default, we use a stacked regression methodology; see Cameron and Trivedi (2005) and Soyeshi (1995) for a more comprehensive discussion of the methodology and its advantages. The idea is simple: rather than estimating the default probability per loan, we estimate the periodic default probability, which we condition on the “life cycle” of the loan. Hence, we use 81,049 loan-performance (3 months) quarters for the 7,455 loans in our sample. Loans drop out of the panel, either when they mature or one quarter after defaulting; note that performance data for the entire sample end in late 2016. We estimate the transition probability from performance to nonperformance using OLS, with the default indicator as the dependent variable. In Section 5, we report the robustness of our results relative to a logit/probit specification of the stacked regression, as well as to the Cox proportional hazard duration model.

Column 3 of Table 3 presents the results. To derive the annualized unconditional default probability for an A-scored loan, we add the quarterly default probabilities across the four stages of the loan’s life cycle: |$0.007+0.01+0.008+0.004$|⁠, equal to |$2.9\%$|⁠. To find the annualized probability of default for an AA-scored loan, we subtract |$1.6\%(=4\times0.04)$|⁠. Column 1 of Table 4 presents the annualized unconditional default probabilities across credit scores.

Applying Bayes’ Law to the unconditional probabilities we derive the conditional probabilities of default given the stage in the loan’s life cycle. That is, for an A-scored loan, we normalize the vector of unconditional life-cycle probabilities, |$\left(0.007,0.01,0.008,0.004\right)$|⁠, by the overall probability of default, |$2.9\%$|⁠, which yields the life-cycle pattern of default probabilities, |$(0.23,0.35,0.28,0.14)$|⁠. That is, conditional on default, there is a probability of |$23\%$| that the loan will default during the early stage of its life cycle and |$35\%$| that it will default during the mid-early stage, etc.

Figure 3 plots the conditional default patterns for all credit scores. Given that FC loans are amortized in equal monthly payments, these life-cycle patterns determine pre-default recoveries. That is, an A-scored loan that defaults early has already repaid between zero and |$25\%$| of the debt, capital and interest; the corresponding figure for a loan that defaulted mid-early is between |$25\%$| and |$50\%$|⁠, etc. Note that, conditional on default, there is a |$42\%(=0.28+0.14)$| probability that the loan will not make it to the midpoint of its life, and so its pre-default recovery is below |$50\%$|⁠. For the conditional recovery, prior to default, of an A-scored loan, we multiply the vector of conditional default probabilities (above) by a vector of conditional recovery rates |$\left(0.125,0.375,0.625,0.875\right)$|⁠. Column 2 of Table 4 reports the results for all credit scores. Evidently, in line with the argument above, conditional recovery rates to default are between |$40\%$| and |$50\%$|⁠.

Figure 3

Conditional default probabilities over the loan’s life

Conditional default probabilities, given the stage of the loan’s life cycle (Early, Mid-early, Mid-late, and Late) and its credit score, are derived from the unconditional probabilities as estimated in Table 3, using Bayes’ Law. For more details, see the discussion in relation to Table 4 in the body of the text.

Open in new tabDownload slide

The last step is to combine recoveries pre- and post-default in columns 2 and 3 of Table 4, respectively, so as to construct an overall estimate of LGD. Column 2 of Table 4 shows that an A-scored loan in default has already repaid |$45.7\%$| of the loan by the time of default, and is expected to pay off |$37\%$| of the remainder post default. It follows that total recovery, conditional on default, is |$65.8\%$|⁠, so that LGD is |$34.2\%$|⁠.

FC’s recovery rates are remarkably high compared with those of unsecured loans and trade credit for U.K. SMEs (see Franks and Sussman 2005). These can be seemingly explained by FC’s innovative use of certain properties of English bankruptcy law. Although FC’s loans are unsecured, they are usually guaranteed, personally, by the owners of the SME. It follows, that corporate default can lead to personal bankruptcy, which has more severe consequences in the United Kingdom than in the United States. First, protection for personal assets, including homes, is nonexistent. Second, many restrictions apply to bankrupt individuals. For example, while in bankruptcy a person cannot “borrow more than £500 without informing the lender,... act as a director of a company without the court’s permission,... create, manage or promote a company without the court’s permission.”¹³. It is common practice for British banks to freeze the bank accounts of bankrupt individuals or to refuse to open new accounts for them. The United Kingdom, unlike the United States, has a centralized corporate register that facilitates the tracking of individuals’ credit histories.

According to Jackson (2016), head of recovery at FC, the platform used the above properties of English law in order to innovate a “survival for revival” recovery strategy. While FC does not agree to haircuts on defaulted loans, it is prepared to show great flexibility in rescheduling loans in default, with a tough position on triggering personal bankruptcy if the borrower fails to meet the deferred payments. As a consequence, FC placed |$25\%$| of the owners of defaulting SMEs into personal bankruptcy. Jackson argues that “for Funding Circle, |$90$|-|$95$|% of recoveries come through the personal guarantor.... A conservative estimate of recovery is 40p in the £ over a five-year period post default” (Jackson 2016), remarkably close to our estimates.

3. Market Efficiency, Mispricing, and Liquidity

Our EMH analysis provides three main results. First, because FC auctions aggregate some private, possibly nonstandard, information dispersed across the investor population, the interest rate better predicts the loan’s probability of default, over and above the credit scores. Second, although such information is priced in, the price still fails the test of information efficiency, as the price deviates from its information-efficient benchmark. The problem is exacerbated over time. Third, these deviations from efficient pricing are due to funding shortages and a poorly calibrated autobid.

3.1 The EMH specification

Information efficiency implies that the augmented interest rate enhances default predictability such that a |$1\%$| increase in |$r^{*}$| predicts a |$1\%$| higher default probability, and that since |$r^{*}$| absorbs all the information contained in |$Dscore$| and |$Liq$|⁠, these variables lose their predictive power.

Under the null EMH hypothesis, since the adjusted interest rate, |$r^{*}$|⁠, incorporates all relevant information, public as well as private (to the extent that it is revealed in the bidding process), it should not be correlated with Equation (4)’s error term. Otherwise, |$r^{*}$| predicts |$\varepsilon$|⁠, which, in turn, predicts default. Rational investors, including algorithmic ones, should use that information in order to revise their bidding strategies. By doing so, they drive the interest rate towards the information-efficient price, up to the point at which all of that information is priced in and the correlation between |$r^{*}$| and |$\varepsilon$| vanishes. An important implication of this argument is that |$r^{*}$| can be included on the right-hand side of Equation (4), though it is determined by the actions of the investors.

By a similar argument, the credit score coefficient, |$\theta$|⁠, as well as the liquidity shock coefficient, |$\eta$|⁠, should lose their economic and statistical significance.¹⁴ The argument for |$\theta$| is straightforward: investors should extract all the information contained in the publicly observed credit scores, discard any “noise” that the credit scores might have, and price it into the interest rate, so that the credit score is left with no predictive power. Likewise, the regression should reject any predictive power in the liquidity shocks. This is obvious when the shock is related to random funding gaps that contain no “fundamental” information about default probabilities. Even if this is not the case, any information that is included should be absorbed into the price through the price discovery process, leaving the funding gap with no predictive power.

Finally, to relax the risk neutrality assumption used in the derivation of Equation (5), and in order to account for aggregate macro risk, we include in Equation (5) the systematic risk variable, |$Sys$|⁠. Since our SMEs are not listed, we have no company-level measures of systematic risk. Instead, we use industry-level betas as reported in Demoderan.¹⁵

3.2 Baseline results

Table 5 presents the baseline regressions; columns 1 to 3 use the borrowing rate (i.e., the weighted average of all accepted bids), and columns 4 to 6 use the marginal interest rate at the close (i.e, the highest interest rate of all accepted bids), both adjusted for LGD, as derived in Section 2.

Columns 1 and 4 report baseline results, without including any measure of liquidity shocks. They already carry the main message of our analysis: while the evidence does not reject the hypothesis that the market aggregates dispersed investors’ information, information efficiency is rejected. That is, on the one hand, the |$1\%$| statistical significance of the interest rate coefficient implies that its inclusion in the regression enhances default predictability over and above the credit scores; on the other hand, the borrowing rate coefficient is significantly smaller than one, at |$0.33$|⁠. That is, a |$1\%$| increase in the borrowing rate predicts only a |$33$| bps increase in the default probability. Similarly, the joint F-test on the estimates of the credit score fixed effects rejects the null hypothesis that they lose all of their predictive power.

Both of these findings are consistent with sizable random deviations of the interest rate from its information-efficient level. Such deviations in the pricing of loans can lead to a standard errors-in-variables effect that flattens the regression line and delivers an interest rate coefficient smaller than one. Similarly, when credit score information is priced into the interest rate with an added random term that diverts the price away from the information efficiency, the credit score FEs regain their statistical significance, as they help in filtering out the deviation.¹⁶ In Section 3.3.1 below, we derive a measure of the magnitude of these deviations as a proportion of the overall variance of the interest rate.

To track the origins of these deviations, we augment the baseline regression in columns 2 and 4 of Table 5. The first of these proxies, Aggregate Weekly Borrowing, captures variations in aggregate demand for funding. For any given loan |$i$|⁠, we measure the total value of loans auctioned off during the 7 days that the loan’s auction is open, and we normalize it by the initial size of FC’s loan book at the beginning of that week; that is, the variable is the growth rate of FC’s loanbook.¹⁷ The validity of our test does not rely on such fluctuations in aggregate borrowing being uncorrelated with loan |$i$|’s default probability. In the event that such a correlation arises, under the EMH, investors should incorporate the relevant information into the interest rate.

The |$-0.002$| coefficient implies that an auction running at a time when there is a one-standard-deviation (equals |$1.3\%$|⁠) surge in platform-wide demand for funds is associated with a reduced annual default probability of |$1.0\%(=1.3\times0.002\times4)$|⁠. We use a simple “correspondence argument” in order to identify the relationship between the shock and the deviation of the price from information efficiency, in both its direction and magnitude: since a demand surge is associated with a lower probability of default, it follows that, in order to restore information efficiency, the interest rate should be adjusted downward. Or, to put it differently, the demand surge drives up the interest rate on loans auctioned during that period over and above their information-efficient level.

We use another proxy to capture variations in the composition of the funding supply. By design, a drop in active funding automatically injects, more passive funding along the autobid supply curve (see Figure 2). We decompose loan |$i$|’s autobid funding into its aggregate time series component and its auction-specific, cross-sectional component. Again, we measure loan |$i$|’s aggregate component of autobid funding as the average share of autobid funding across loans that are open during the 7 days that loan |$i$| is open. We then measure the cross-sectional component of autobid funding as the difference between loan |$i$|’s share of autobid funding and the aggregate component as defined above. This results in two variables, Aggregate Weekly Autobid and the Loan-level autobid.

The estimated coefficient of Aggregate Weekly Autobid suggests that interest rates are above the information-efficient price at times of high aggregate autobid activity. The |$-0.016$| implies that an auction running at a time when there is a one-standard-deviation (equal to |$7\%$|⁠) surge in aggregate autobid funding is associated with a lower annual default probability of |$45$| bps |$(=7\times0.016\times4)$|⁠. Using the correspondence argument above, we conclude that periods of high autobid activity– thus low active investing– are associated with high interest rates relative to the level of the information-efficient price. This result also suggests that, even though the autobid channels in more passive funds, the interest rate still remains too high. It can be argued that the autobid is undersmoothing the interest rate.

In contrast, the estimated coefficient of Loan-level autobid suggests that interest rates are below the information-efficient price at times of high loan-level autobid activity. The |$0.05$| coefficient implies that an auction in which the autobid is one standard deviation (which stands at |$18\%$|⁠) more active is associated with a higher annual default probability of |$36$| bps |$(=18\times.005\times4)$|⁠. Again, we use the correspondence argument above to conclude that an auction with a low level of active participation tends towards a low interest rate relative to the information-efficient price. In terms of the cross-sectional allocation, the autobid channels in an excessive amount of passive funds, resulting in an interest rate that is too low. In that respect, the autobid is oversmoothing the interest rate.

Another source of mispricing on the platform originates in the borrower’s ability to terminate the auction early. Table 5’s Early Termination variable is a dummy that receives a value of one in the event that the borrower takes such an action (and zero otherwise). Early termination is associated with an annualized default probability |$1.2\%(=0.003\times4)$| higher relative to auctions that run their full course, so that the interest rate should have been set higher by a similar amount. Obviously, investors are not informed about early termination until the event actually takes place, when it is too late for them to adjust their orders. Nevertheless, the FC auction design does price in the event of early termination: interest rates can move only downward in auction time, so that by terminating early, the borrower increases his own cost of borrowing. While early termination is, therefore, priced in, the estimated coefficient 0.003 shows that the “penalty” to the borrower for early termination is not sufficiently high.

Columns 3 and 6 of Table 5 address a problem raised in Iyer et al. (2015), whereby EMH regressions may not be sufficient evidence for information aggregation. Rather, one may argue that the extra predictive power obtained by augmenting regressions containing credit scores with the interest rate may result from investors recovering information lost when the platform converted a continuous credit mark into discrete credit scores, from A+ to D. If this is so, no new information is actually revealed by the auction’s price-discovery process. Like Iyer et al. (2015), we reject this hypothesis but use a different approach. If the credit mark is the only source of information, all loans should be priced within the credit-score band, which, in our case, is |$50$| bps around the band’s midpoint (see Table 3). The variable Above The Band is a dummy that receives a value of one if loan |$i$|’s auction closes above the band. Interacted with the interest rate, the variable is positive, |$0.15$|⁠, and significant at the |$5\%$| level for the borrowing rate. The implication is that the information content of prices is actually higher above the credit-score band.

3.2.1 Discussion: Optimizing the autobid

Given that the autobid allocates 48% of the platform’s funding, it can be argued that it could be used more effectively to improve price efficiency. That is, the parameters of the autobid could be recalibrated so as to refine and optimize the design toward information efficiency.

Consider, first, aggregate liquidity shocks, such as cases of high loan demand or low supply of funds by active investors. Our analysis finds that a high demand for new loans drives interest rates up and away from the information-efficient level for all loans auctioned at the time. A similar result is found in weeks of active investors’ lower participation in the supply of funds. In both cases, the autobid should inject more funding to drive down the interest rate toward information efficiency. However, this policy required a stock of liquidity on which FC could draw. To build such a liquid stock, the autobid would have needed to defer the allocation of passive funds contributed at times of excess aggregate funding. It seems, however, that FC was reluctant to engage in such reallocation of liquidity over time, as it owed the passive investors a prompt allocation of their funds into income-generating loans.

Consider next, liquidity shocks with a cross-sectional effect. Our analysis of the Loan-level autobid variable reveals that the autobid channeled too much liquidity into auctions in which active investors held back funding, resulting in an interest rate too low relative to information efficiency. Another related result is that, while early termination is penalized with a higher interest rate relative to auctions that run their full course, the penalty is not sufficiently high to achieve information efficiency. Both results highlight an oversmoothing effect, whereby the autobid inhibits market forces from having their full impact on the interest rate. The solution would be to make the autobid more price sensitive to the order flow, say, by making the autobid supply function steeper. In this case, no buildup of liquid funds would be required as funds would be reallocated across auctions running simultaneously.

It seems that the FC investor community was particularly aware of cross-sectional faults in the design of the autobid. For example, on February 21, 2014, at 1:28 p.m., a blogger named “aloanatlast” commented on an auction that attracted a low level of attention from active investors: “The autobidder will now be chucking every penny it can into that loan.” He then articulated his own idea about design refinement: “If I were an Autobid user [i.e., a passive investor], I’d want it to buy me a random sample, like a sort of index tracker - not something programmed to soak up the [loans] that manual bidders don’t want.”¹⁸ Notice that this refinement is a neutral allocation of passive funds, without any attempt to use the autobid to deal with liquidity shortages.

It has to be emphasized that our finding of a suboptimal design is not an indication of incompetence on the part of FC. Unlike in a market microstructure textbook problem, the (algorithmic) market maker had no prior knowledge of the cross-correlations between default probabilities and funding flows, upon which the parameters of a policy could be calibrated. Rather, in a newly innovated market, refinement could rely only on a trial and error learning process. That is, the autobid had to be “trained” dynamically on a relatively small sample of just |$7,500$| auctions at the end of the sample period. Moreover, the number to be estimated, that is, the default probability, was relatively small, with default events being realized with a long delay. In addition, any change in the design of the autobid was likely to cause an adjustment of the active investors’ bidding strategies along the lines of the Lucas critique (1976).

3.3 Liquidity events

To deepen our understanding of the way that liquidity shocks can drive prices away from fundamental values, we analyze two such events: the first involves the termination of the BIS program mentioned above, and the second involves a shortage of funding in auctions closing at hours less convenient to retail investors. In both cases, we find evidence of price impact that is not consistent with the EMH.

Starting in March 2013, the U.K. government, via the Department of Business Innovation and Skills (BIS), created a £20 million fund to be allocated via the FC platform. The intention was to support the development of alternative models of financial intermediation, in order to complement and compete with the banking industry, which was perceived to have failed to provide adequate funding to the SME sector (see Breedon Review 2012). While the scheme was running, |$20\%$| of each loan processed by the platform was BIS funded, at a price that was determined by the auctioning of the remaining |$80\%$| of the loan. BIS funding was eventually exhausted, with the last auction to benefit opening on February 28, 2014. The effect was a temporary platform-wide shortage of funding, similar to the effect of a surge in loan demand, as analyzed in the previous section. The termination of the program was triggered solely by the depletion of the original fund and was unrelated to any other policy decision or to the functioning of the platform. It is, therefore, highly unlikely that there was any difference in loan quality before and after the event.¹⁹

To capture price dynamics around the event, we run a regression similar to Equation (1), namely, the borrowing rate on credit scores, but with periodic FEs, each lasting 10 days, replacing the quarterly FEs in Equation (1). The new equation is estimated over a period starting 60 days before the BIS termination date and ending 60 days after. Figure 4 plots nonparametric estimates of the 10-day FEs. For completeness, the figure also plots a polynomial time path for the borrowing rate. The figure reveals a surge in the interest rate of around |$1\%$|⁠, starting about 10 days after the termination of the BIS funding, lasting for about 20 days and then, gradually, tapering off. The delay in the response is consistent with investors, both human and algorithmic, providing the market with liquidity from their precautionary reserves, but once these reserves were exhausted, prices were affected. Figure 4 gives currency to the notion of slow-moving capital. It took some time before new liquidity flowed in, either from new investors or from already active investors, responding to opportunities created by the end of BIS funding. Crucially, the time it took for capital to move in response to such opportunities is measured in weeks rather than hours or even days.

Figure 4

Interest rates 60 days before and after BIS funding

The figure plots borrowing rates of auctions held over a period starting 60 days before the BIS termination date (February 28, 2014) and ending 60 days after. During that time window 1,292 auctions were held with 653 held before the termination, and 639 afterwards. We divide the sample into mutually exclusive bins of 10 days. For each bin, we compute the average and 90% confidence interval of the borrowing rate, and plot these values at the bin’s midpoint. The red line plots a third-order polynomial separately before and after the end of the program.

Open in new tabDownload slide

For a stronger test, which controls for the default information that is incorporated into the price, we draw, again, on the EMH specification. Columns 1 and 2 in Table 6 present EMH regressions with the default dummy as a dependent variable. On the right-hand side, we have the interest rate, the previous set of covariates from Table 5, plus FEs for consecutive 10-day intervals following the event. The main result is a |$70$| bps drop in default probabilities (2.8% annualized), over and above any information that is priced into the interest rate |$20$| to |$30$| days after the event. The result is statistically significant at the |$5\%$| to 10% levels, notwithstanding the sharp drop in sample size. Using the conversion argument once again, we conclude that interest rates surged in the short term following the withdrawal of BIS funding.

The second test exploits the auction’s closing hour. We conjecture that, throughout the day, retail investors have relatively less time to engage with the FC platform during work hours or late at night, as opposed to afternoon and early evening hours. Therefore, we define the peak hours as those between |$4$| p.m. and |$7$| p.m. (henceforth peak hours). If our conjecture is correct, loans closing off-peak would attract les attention from retail investors and, therefore, create a shortage of funds. As the closing hour is publicly known at the open, under the EMH, liquidity providers should inject extra funds to compensate for the shortage off-peak, so as to avoid any deviation from price efficiency. Note, also, that the allocation of the opening hour is random, so that there should be no systematic quality difference between loans, conditional on the closing hour.

To confirm the difference in the flow of funds on- and off-peak, Figure 5 plots mean order flows, normalized by loan size, during the 6 hours prior to the closing of the auction. Only auctions that run for the full 7 days are included, so the closing hour is, indeed, perfectly foreseen. As expected, loans that close within peak hours have a significantly higher order flow during the last hour. Normalized by loan size, last-hour order flow is |$0.25$| for auctions closing in peak hours, compared with only |$0.15$| for auctions closing in off-peak hours.

Figure 5

Mean inflow of funds, according to closing hour

The figure plots hourly mean order flow normalized by loan size, during the 6 hours prior to the closing of the auction. The sample constitutes auctions that run the full 7 days. Off-peak denotes auctions that closed outside of the 4PM to 7PM window, and Peak denotes auctions that closed between 4 p.m. to 7 p.m.

Open in new tabDownload slide

Again, for a stronger test, we augment our EMH regressions with the usual covariates, with a dummy variable for the off-peak closing hour.²⁰ Auctions that closed off-peak have a quarterly default probability |$10$| bps (⁠|$0.4\%$| annualized) lower than that of auctions that close during peak hours. Using the conversion argument once again, we conclude that the interest rate was hiked for auctions closing off-peak, relative to the information-efficient interest rate. The result is statistically significant at the |$10\%$| level.

3.3.1 Economic magnitude of deviation from information efficiency

The two events allow us to clearly trace out the mechanism through which liquidity shocks drive prices away from fundamental values. Next, we turn to a quantification of mispricing. Intuitively, we ask how much of the variance in the lending rate can be explained by expectations about fundamentals.

To obtain an idea of the magnitude of the deviations from information efficiency, we report the R-squared of an equation in which the dependent variable is the borrowing rate, and the right-hand-side variable is the loan’s estimated default probability. The latter is derived as a fitted value from the EMH regression in Column 2 of Table 3. It can be shown that the R-squared provides an upper bound on the information content of the price, as a fraction of the overall variance of the interest rate. It is striking that information accounts for, at most, 20% of the price variance; or, to put it differently, at least |$80\%$| of the variance is nonfundamental “noise” in the form of deviations from the information-efficient price. For full technical details about the results reported here, see Internet Appendix B.²¹

It is noteworthy that not all platform participants were equally affected by this amount of noise in the price. While investors can diversify it away and, indeed, even use it to generate arbitrage profits (see Section 4), borrowers are unlikely to be able to do so. The result is that a borrower who was unlucky enough to auction his loan at a hiked-up interest rate, had to bear the cost for the entire duration of the loan, typically for several years. This result echoes one of the reasons given by FC for abandoning the auction design: “The price of each loan will now be based on the risk and term of the loan rather than the availability of investor funds.”²²

3.4 The declining information content of prices

We might expect that, with the expansion of the platform and the accumulation of data and expertise, the quality of pricing would improve: the autobid can be better trained to extract information from funding flows and to prevent uninformative liquidity shocks from affecting prices. The hypothesized result is that the magnitude of random deviations of interest rates from information-efficient prices would fall, and the interest rate coefficient in the estimated EMH Equation (4) would approach one.

To test this hypothesis, Column 1 of Table 7 augments the baseline EMH regression with a linear time trend, interacted with the borrowing rate, and column 2 interacts the borrowing rate with year dummies. Columns 3 and 4 repeat the test using the marginal rate.²³ We find a secular deterioration in the quality of pricing over time. Moreover, Column 2 reports a benchmark coefficient of |$0.82$| in 2011, which is not statistically different from one; the EMH cannot be rejected for the early phase of FC’s activity. Over time, the interest rate coefficients fall sharply, so that by 2015, the year before FC canceled the auction design and moved to posted prices, the interest rate coefficient was no longer statistically different from zero. That is, toward the end of our sample, prices no longer carry information that can contribute to improved default predictability. We explore reasons for this result in the next section.

4. The Vanishing Crowd

The declining information quality of price is surprising since it might be expected that, with accumulated experience, FC designers could gather more information over time and use it to recalibrate the autobid and improve price efficiency. One possible impediment to such learning might be that the patterns of the funding flows were not stable over time, and were changing faster than FC’s learning process. In this section, we provide evidence that there was a sizable shift from active to autobid investment and, equally significantly, from a large number of small and medium-sized investors to a relatively small number of large investors. That is, to a large extent, the “crowd” simply vanished. We link these new results to the results for the EMH regressions, discussed earlier.

We use our bidding data to measure investor characteristics. Since investors’ entry into (and exit from) the platform is staggered, comparing an investor who was active for several years with one who entered toward the end of the sample may be problematic. To restore a measure of consistency, we track investors’ activity within fixed time intervals. That is, for each investor, we aggregate the value of accepted bids across auctions within calendar quarters. Hence, for each investor quarter, we derive a portfolio, denoted as Total Wealth, which is used to sort investors into size groups of less than £100, £100-500, etc. (see Table 8). Each portfolio may be managed via the autobid (see column 3 of Table 8), or invested directly by the investor, which is denoted as the Active Portfolio. We classify investors as Active even if they partly use the autobid. Summary statistics on their portfolios can be found in columns 5 to 10 of Table 8.