29Case Studies II

On 1 August 2012 a market maker, Knight Capital, lost more than $460 million in forty-five minutes. Five days later it had sold $400 million of convertible preferred stock to a group of investors, convertible into about 73% of its shares; eleven months later it merged with GETCO under a new holding company. The controls failed in the morning; the decisions that set what the failure cost the firm’s owners came in the next five days, and many years before. This chapter reads four more failures from the public record: a software incident, a family office’s prime brokers, an exchange that was also a counterparty, and an exchange’s decision in a squeeze.

29.1 2012: a software incident and a rescue

The Securities and Exchange Commission’s order of October 2013 sets out the morning: a technician did not copy a new release of the firm’s order router to one of its eight servers, where an old function remained present and callable, and over 45 minutes it sent millions of orders and obtained more than 4 million executions in 154 stocks, for more than 397 million shares. Book 11 read the order for its controls: no automated limit on the firm’s aggregate exposure, and monitoring that did not alert on it. The firm lost over $460 million on the order’s figures; its quarterly report put the realised pre-tax loss at about $440 million, against total equity of $1 497 million on 30 June.

On 6 August the firm sold 400 000 shares of convertible preferred stock for $400 million to a group including Jefferies, Blackstone and GETCO, convertible into about 266.7 million common shares, $1.50 each, against 97.9 million shares outstanding. Issuing that much stock normally needs a shareholder vote; the board’s finance and audit committee relied instead on the exchange’s exception for a delay that would seriously jeopardise the company’s financial viability, and the exchange did not object. Eleven months later the firm merged with GETCO under a new holding company, and each share became $3.75 in cash or a third of a share of the new company.

The table is the firm-level arithmetic of a rescue under time pressure. The loss alone cost the shareholders 29.4% of the book equity; the rescue, priced at $1.50 a share when book value per share was still $10.80, cost them a further $666 million of book equity, transferred to the new investors, who paid $400 million for $1 066 million. The price was set by the days the firm had left: its clients and counterparties, the quarterly report says, had reduced order flow and lost confidence, and a firm without capital cannot keep making markets. What it had not decided in advance (the limit on aggregate exposure, the capital held against a morning like that one, a standing facility that could be drawn in a day) set the terms.

29.2 2021: a family office and its prime brokers

The Credit Suisse board’s special committee published its report on the bank’s losses from Archegos, a family office, in July 2021. On the evening of 25 March 2021, after the value of its largest positions had fallen steeply since the 22nd, Archegos told its prime brokers that it had $120 billion of gross exposure, $70 billion long and $50 billion short, against $9 to $10 billion of equity, and asked them to agree a standstill while it wound its positions down. They declined. Credit Suisse’s own gross exposure was about $27 billion on 23 March and $17 billion on 26 March; it had called $2.8 billion of margin on the 25th; it lost about $5.5 billion.

The report describes both regimes inside one bank. The prime brokerage book was dynamically margined; the swaps were margined statically, the initial margin fixed in dollars at inception, so that as the positions rose in value the margin as a share of them eroded. The average margin on the swaps was about 9.4% on 23 March. A concentrated position that its own holder estimated would take two weeks to a month to liquidate carried the margin of a diversified one.

The report records the race. On 26 March Goldman organised block sales of some of the positions, and Credit Suisse took part in three without knowing how many shares were offered or whose were sold first; it sold just over $3 billion that day, keeping its algorithmic selling within 2 to 3% of average daily volume. Over the weekend several banks, including Deutsche Bank, Morgan Stanley and Goldman, were not interested in a managed liquidation; Credit Suisse, UBS and Nomura agreed one and sold blocks in April.

29.3 Tutorial: the rescue, the race, the run, the squeeze

def exit_race(shares, margins, impact):
    """Brokers holding fractions `shares` of one long position (value 1 at default) sell in turn, in list order,
    under linear permanent impact: the price after a cumulative fraction f is sold is 1 - impact f. Each broker's
    loss per unit of the whole position, net of its margin (fraction of its exposure), floored at zero."""
    out, f = [], 0.0
    for s, m in zip(shares, margins, strict=True):
        avg = 1 - impact * (f + s / 2)                 # average price over its segment
        out.append(max(0.0, s * (1 - avg - m)))
        f += s
    return out


def race_impact(loss_share, share, margin, shares_before):
    """Impact coefficient at which a broker with `share` and `margin`, after `shares_before`
    was sold ahead of it, loses `loss_share` of the whole position."""
    return (loss_share / share + margin) / (shares_before + share / 2)

In the model the first broker to sell loses nothing at any of the three margins: it sells at an average fall of 3.8%, inside its margin. The last loses $6.79 billion at 9.4%, and $4.99 billion even at 20%. The margin that would have left each broker whole is the average fall over its segment: 3.8% for the first, 26.6% for the middle, 49.3% for the last. No static percentage covers the race, because the loss depends on the order, which no broker controls in advance.

The proposition is why a standstill is rational for the creditors together and irrational for each one alone: a managed liquidation leaves the total fall unchanged (or smaller, if slower selling lowers the impact), while selling first moves one’s own share of it onto the others. Dynamic margining is the defence each broker controls: a margin that grows with the position’s size against the market’s volume and with the broker’s ignorance of what the client holds elsewhere, set before the race starts.

29.4 2022: an exchange that was also a counterparty

Book 3 read FTX’s collapse from the debtors’ filings. For a trading firm the lesson is about exposure. A firm that holds cash and tokens on an exchange is a creditor of the exchange; if the exchange lends customers’ assets to an affiliate, the firm is also, without knowing it, a creditor of that affiliate. The debtors’ first interim report found that the affiliate, Alameda, had a borrowing limit set to $65 billion and settings that let it withdraw below it and run a negative balance; the shortfall analysis found that at the petition, for the most liquid tokens, the exchange held assets worth 6.6% of what it owed customers, and 22.6% for all tokens with receivables.

The run was fast: customers withdrew a net $3.3 billion on 7 November, and on 8 November the exchange paused withdrawals. A firm that learned of the shortfall on the 8th had no exit left; the decisions that mattered were the limits set earlier on how much the firm held on any one venue, how quickly excess balances were swept back, and what evidence of segregation it required (chapter 27’s trapped assets).

29.5 2022: the nickel squeeze and the exchange’s decision

On 7 March 2022 the London Metal Exchange’s three-month nickel contract closed at $48 078 a tonne, up from an opening just under $30 000. In the first hours of 8 March it rose to $101 365 at 06:08 and stayed above $80 000 from 07:00. The Court of Appeal’s judgment records what the exchange faced: if the morning’s trades stood, members would have had to post some $19.75 billion of margin by 09:00, and at least five were expected to default. It suspended trading at 08:15 and cancelled every trade made since midnight. Large short positions had been built by several participants, among them Tsingshan on the over-the-counter market. The FCA’s notice adds that margin calls had set records of $3.5 billion on 4 March and $5.1 billion on the morning of 7 March, and that the price bands were switched off from 04:49 to 08:15 on a decision by junior staff.

For the short, the arithmetic is chapter 14’s: a call that grows linearly with the price, against cash that does not. For the exchange it was a choice between a margin call that would have defaulted several members and a cancellation that would take trades from those who made them. The Divisional Court held the cancellation lawful, and dismissed Jane Street’s similar claim; the Court of Appeal dismissed Elliott’s appeal, holding that the only alternative, collecting margin at the previous close, would have put the clearing house in breach of its own obligations. Two firm-level lessons follow for a trading firm: an exchange can cancel trades in extreme conditions, so a profit made in a disorderly market is not yet a profit; and the firm’s margin plan must survive calls several times larger than any seen before, since the record ones came on consecutive trading days.

29.6 What the four share

The four failures differ in cause and share a structure. In each, a counterparty’s terms were fixed before the event on assumptions the event broke: a rescue priced in days, a margin set at inception, a balance left on a venue, a margin call sized for ordinary moves. And in each the firms that did best had decided their response in advance: the brokers that sold first, the exchange with the power to cancel, the firms whose exposures were already small. Chapter 27’s drill is where a firm finds out which of these decisions it has taken.

29.7 Build: the casebook, part 2

29.8 Exercises

29.9 Problem: The Last to Sell

29.10 Interview questions