Bitcoin Flash Crash on May 19, 2021: What did really happen on Binance?
Joint work with Tim Baumgartner
Introduction
On May 19, 2021, Bitcoin plunged 30% and the largest crypto exchange Binance experienced a significant outage. Besides offering a broad range of crypto spot markets, Binance provides highly leveraged crypto futures. The exchange compensates the counterparties if futures clients lose more than their account value due to liquidations and slippage. Binance operates an insurance fund to overtake these liquidated positions when their value becomes negative.
The financial news intensely covered the May 19 outage event on Binance. Harmed clients allege that the leading crypto exchange intentionally shut down its platform to have free play when liquidating their clients’ futures positions in order to limit losses in its insurance fund.
We find that Binance stopped transaction-level reporting in the middle of the crash for 40 minutes. A part of the data has been backfilled by Binance later on. We investigate the authenticity of these particular transactions and find that they do not conform to Benford’s Law — a widely used approach to detect fake data which also helped uncovering the Libor manipulation (Abrantes-Metz et al., 2012).
During the outage period we estimate that the insurance fund would have had to trade a volume of around 1 bil. USD. At the time of writing this paper there has been no way to verify the insurance fund’s assets, neither the absolute level (as stated by Binance) nor the liquidity of the underlying assets. Hence, one plausible explanation for the outage could have been the attempt to save the insurance fund from running empty.
For further details and more results, refer to our full academic article (Baumgartner and Guettler, 2022).
Institutional Setup
There are more than 600 crypto exchanges, of which Binance has by far the largest trading volume. The exchanges enable trading different pairs of cryptocurrencies directly against each other, or to swap cryptocurrencies against traditional currencies. In addition, they also offer derivative instruments such as futures, leveraged tokens, or options on cryptocurrencies. Binance allowed even small retail clients to engage in futures trading, where the exposure may reach a factor of up to 125 of the pledged collateral. (Note that the exchange has already decreased its derivative products spectrum in various jurisdictions due to regulatory scrutiny.)
To maintain a levered position, Binance determines the “maintenance margin” being typically about half of the initial margin — the collateral required to enter into the position. When the collateral falls below the maintenance margin, Binance liquidates the position.
Liquidations of futures positions are intended to yield a small positive payoff from the exchange’s perspective. Nevertheless, the trade execution (shortly after the automatic initiation of the liquidation) may take place at a less favorable price if market conditions change fast (“slippage”). Then, the payoff can be negative. In such cases, Binance’s insurance fund takes over the position, such that the counterparty receives the total amount due.
Most traditional futures exchanges maintain such funds. In the case of Binance, this insurance fund is however an opaque vehicle for which little information is known (Binance only publishes the fund’s equity on its website at a daily frequency) and almost none is verifiable. It is no separate (legal) entity with independent audits. Binance also does not publish any specific wallet address which would enable a plausibility check. The fund remains a part of Binance’s equity. Therefore, any realized losses (gains) negatively (positively) impact its bottom line.
The insurance fund assumes to-be liquidated futures positions over-the-counter, i.e., these transfers are not reported as trades. The insurance fund may hold the positions for a longer period, cancel opposite positions against one another, or place a limit order to close positions. When the insurance fund closes positions in the open market, the corresponding transactions receive the insurance fund flag in Binance’s API.
Data
We retrieve individual transactions at millisecond-level from Binance’s API which the data provider Tardis.dev has cached (denoted by Tardis API data). We focus on Bitcoin transactions as it is the most important and largest cryptocurrency. We use the day of May 19, 2021 itself and the ten days before as reference period, as well as the day after the crash. Data points comprise a timestamp (UTC) at millisecond-granularity, the Bitcoin amount, the price for each transaction, and an indicator whether the Binance-owned insurance fund has been one of the counterparties of the trade.
The Tardis API data contain a transaction data gap roughly between 13:00 and 15:00 on May 19, 2021. We fill these gaps with additional data which became available in March 2022 in Binance’s Public Data Collection. This data, however, does not contain a flag on insurance fund trades.
We also collect minute-level OHLCV (open, high, low, close, volume) data from TradingView to which Binance supplies this aggregated data. For that minute-level data there has been initially a gap between 13:16 and 13:56, which Binance has later back-filled this gap. The missing nature of in case of this data provider is quite remarkable given that TradingView is the leading financial analysis platform for traders and investors.
Timeline of the Crash
The price development of May 19, 2021, has the character of a flash crash. There has been no particular news justifying such volatility. On the day before, the Bitcoin futures price experienced a decrease by roughly 5% to 42,500 USD, which further expanded to 40,000 USD in the first twelve hours of May 19 (UTC). The sharp price decline began at around 12:30. Prices reached the maximum drawdown at 13:09, being 33.9% lower than at midnight. However, prices recovered rapidly and volatility decelerated at about 35,000 USD at 14:00. We also observe marked differences to the Bitcoin futures prices at other major exchanges (data for other exchanges is from Tiingo Inc.).
Binance’s platform outage evolved as follows. Clients first complained at 12:02 when trying to deposit funds, referring to a gateway timeout or an unavailable service. More importantly, clients reported the first errors connected to trading at 12:14, showing a gateway timeout when, e.g., trying to purchase cryptocurrencies. At 13:05, Binance tweeted that “ETH and ERC20 withdrawals are temporarily disabled due to network congestion”. Further user complaints include screenshots of a blank user interface in Binance’s mobile app at 13:07, and still around 17:00 some customers posted “server too busy” errors on social media.
Has Binance faked Trading Data?
This section focuses on the most critical time frame in the middle of the crash ranging from 13:16 to 13:56 (UTC). These 40 minutes have been populated with delayed data on the leading financial analysis platform TradingView, casting doubt on the credibility of these transactions. Potential data manipulation could either be due to filtering the actual transactions or fully making up the transaction data. Therefore, we employ forensic methods to investigate the plausibility of the delayed transaction data.
We postulate that these transactions should conform to Benford’s Law, a well respected method which also helped uncover the Libor scandal (Abrantes-Metz et al., 2012). Benford’s Law postulates that certain digits should occur more often than others due to the multiplicative nature of their composition. For instance, the first digit “1” should be more likely than the first digit “9”. The observed relative frequency of first digits should converge to their expected value, while substantive deviations speak against the authenticity of the data.
We consider the tick-level transaction data, that were delivered in a heavily delayed manner by Binance for the 40 minutes in question, containing 373 thousand Bitcoin futures transactions. We investigate the individual Bitcoin volumes of these transactions as the Bitcoin prices lack a sufficient dispersion in its first digits (i.e., they were either 2 or 3 in that period) which is necessary for Benford’s Law to hold.
Figure 2 compares the observed frequencies of first digits (bars) to the expected value according to Benford’s Law (dashed line). The observed frequencies are skewed such that lower digits occur more often than expected, and the frequency of higher digits is hence lower. In particular, the first digit “1” appears almost twice more often. These substantial (and according to further tests statistical significant) deviations are consistent with fake data.
We run the following robustness check to address potential concerns that Bitcoin futures volumes in general do not conform to Benford Law’s distributional assumptions. We hence run the test for reference data comprising the ten-day observation period (May 9 to May 18, 2021) before the crash day. We observe deviations from Benford’s distribution, albeit not as pronounced as in the 40-minute interval with delayed transaction data on May 19.
Figure 3 shows the resulting difference-in-differences of the digit distribution (relative frequencies) between the 40 minutes and the reference data (both versus the expected relative frequencies according to Benford’s Law). The comparison shows that the abnormal effect seems more pronounced in the 40-minute interval of interest compared to the reference data.
Furthermore, we add a bootstrap test by sampling 20 000 observations, half of which stem from the 40-minute interval and the other half from the reference data. We find statistically significant differences between the suspicious 40-minute interval with back-filled transaction data and the reference data. Our results are in line with the notion that the latter may have been manipulated.
We have also simulated faking data based on another crash day: April 18, 2021. We delete 90% of transactions randomly in all minutes where the Bitcoin price of Binance was at least 0.1% lower than on other exchanges. This filtration corresponds to what one would do to reduce artificially questionable data. This filtration results in qualitatively similar results in differences-in-differences.
What Role did Binance’s Insurance Fund play?
To answer this question, we need to model the insurance fund activity, because insurance fund transaction data was missing during the outage. We include Bitcoin price volatility and order book liquidity to predict insurance fund activity in our (Tobit and OLS) regression models.
The upper panel in Figure 4 shows the estimated insurance fund activity. The model gauges the insurance fund’s volume adequately and remains conservative as predicted volumes are notably lower than those reported before (or at) 13:15. For the time after 13:15 we estimate the insurance fund’s volume for the data gap, where no reported volumes are available. The model predicts considerably higher volumes than earlier on the day. Notably, the estimated insurance fund interventions are nearly at the record high of the day when data becomes unavailable. There are several waves of predicted interventions in the following two hours, which are then flattening over time. When data availability resumes at 15:15, the reported and estimated interventions converge to zero.
The lower panel depicts the cumulative sum of trading volume attributable to the insurance fund. Until 13:16, the volume slightly exceeded 500 mil. USD, being already a material volume compared to its balance of around 300 mil. USD. More importantly, the model suggests a vast increase reaching more than 1 bil. USD until 16:00.
From a “back of the envelope” approach, the estimated insurance fund’s trading volume of 1 bil. USD can serve as a factor to reckon the overall cost incurred by the fund. Assuming a spread of 1% between the achievable liquidation price and the market price when the liquidation was triggered, and Binance immediately winding off the overtaken positions, the loss would total 10 mil. USD. That stands in sharp contrast to Binance’s stating a mere reduction of 1.7 mil. USD.
Rather than immediately offsetting these insurance fund positions into a thin order book (recall that most clients could not place orders), Binance could have held these positions on its own books and at its own risk. But given the volume of 1 bil. USD, it is not unlikely that Binance may have run out of buying power at some point because of liquidity issues. Additionally, these numbers only include effects of Bitcoin futures alone and do not comprise the negative effects of futures on other crypto assets such as Ethereum which Binance is offering.
On first sight, these losses of 10 mil. USD seem to be easily manageable given a stated insurance fund volume of 290 mil. USD. However, neither the absolute level (as stated by Binance) nor the liquidity of the underlying assets can be verified externally. Given how easy it were to publish the insurance fund’s wallet address, concerns about the insurance fund’s size and asset structure seem warranted. In contrast to centralized (stock) exchanges that are publicly traded, such as Nasdaq or NYSE, for which detailed financial statements are publicly available, nothing is known about the available capital backing operations of crypto exchanges such as Binance.
Overall, these arguments — whether immediately offsetting the fund’s positions (and realizing losses) or holding the positions in order to avoid short-term losses (but running out of capital) — are in line with deliberately “pulling the plug” and shutting down the futures engine because significant and expensive insurance fund interventions were imminent.
Policy Implications
Our study suggests that crypto exchanges require more consumer protection. If an outage occurs, trading needs to stop for everyone.
Any front-end (and general API) downtime, where clients submit orders and add collateral, requires a consequent stop of the liquidation engine as the resulting liquidations are otherwise unfair. A prioritization of valuable clients in the trading engine is strictly inappropriate.
The high leverage itself is at the heart of the problem. Offering high leverage seems misguided for (retail) investors in the already highly volatile crypto market. Leveraged products should only be offered to selected clients who have sufficient spot trading experience and enough capital. Finally, the crypto exchanges need to reveal the insurance funds’ wallet addresses, put their insurance funds into separate and bankruptcy-remote legal entities, and become audited on a regular basis. Furthermore, implementing a (transparent) circuit breaker would allow time for margin calls to be met and reassess information.
References
Abrantes-Metz, R.M., M. Kraten, A.D. Metz, and G.S. Seow (2012) Libor manipulation?, Journal of Banking & Finance 36: 136–150.
Baumgartner, T. and A. Guettler (2022) Bitcoin Flash Crash on May 19, 2021: What did really happen on Binance?, Ulm University Working Paper
