9Dark Pools and Blocks

A pension fund wants to buy two million shares of a stock that trades five million a day. Shown on a lit book, the order would move the price before a tenth of it traded; sent to a dark pool, it waits for a seller, and every venue it touches learns a little of what it wants. Dark venues exist to let large orders meet without announcing themselves, and they work only as well as they keep the secret. This chapter builds a dark pool and a block venue beside a lit book in the exchange simulator, measures who fills in the dark, and works one large order four ways to price what leaks.

9.1 Midpoint crossing and minimum quantities

A dark pool (One Quant Book 1, chapter 9) displays nothing and usually crosses at the midpoint of the lit market’s best bid and offer, so both sides save half the spread against trading on the lit book: between 20.00 and 20.02, a cent a share each, USD 100 on 10 000 shares. The price is borrowed: the pool needs the lit quotes, and its crosses are only as current as its copy of them. firm.exchsim now gives a dark venue that copy: a reference quote control message (N) carries the lit venue’s best bid and offer to the dark engine after each change, and midpoint pegs price off it (midpoint_dark_pool(reference=…)).

A resting midpoint buy of 5 000 shares with a minimum of 1 000 ignores a 300-share sell and trades with a 1 500-share one. The minimum is a filter on counterparties: small orders that probe the pool, and the small orders of fast traders generally, cannot reach it. It is also a filter on fills: the large order trades less.

9.2 Conditional orders and block venues

Conditionals let one order rest on several block venues and algorithms at once without risking a double fill, since only a firm-up commits it. Their weakness is in the definition: an invitation tells both sides that a large contra exists. A counterparty that never meant to trade gets the information for free, which is why the firm-up rate is policed (Figure 9.1). firm.darkpool.ConditionalBook implements the matching, invitations, firm-up timers and rates.

    def invitations(self, t_ns: int) -> list[tuple[int, int, int, int]]:
        """Match free conditionals in time priority: (invitation id, buy id, sell id, qty)."""
        out = []
        free = sorted((o for o in self.orders.values() if not o.busy), key=lambda o: o.t_ns)
        buys, sells = [o for o in free if o.side > 0], [o for o in free if o.side < 0]
        for b in buys:
            for s in sells:
                if s.busy or b.busy:
                    continue
                q = min(b.qty, s.qty)
                if q >= b.min_qty and q >= s.min_qty:
                    b.busy = s.busy = True
                    iid = self._next_inv
                    self._next_inv += 1
                    self.open[iid] = _Inv(b, s, q, t_ns)
                    for o in (b, s):
                        self.stats.setdefault(o.owner, [0, 0])[0] += 1
                    out.append((iid, b.id, s.id, q))
                    break
        return out

9.3 Who trades in the dark, and the evidence

Zhu (2014) explained who goes dark. Informed traders trade in the same direction as one another, crowd on the heavy side of the market and so face a higher risk of not being filled in the dark, where the other side must come to them; the exchange suits them better, the dark pool suits uninformed traders, and a dark pool can concentrate information on the exchange. Comerton-Forde and Putnins (2015) found in Australian data that dark trades are less informed than lit trades, that low levels of non-block dark trading are benign or even beneficial for informational efficiency and high levels harmful, and no evidence that block trades in the dark impede price discovery.

The chapter’s market reproduces the selection. Its efficient price $P^\ast$ moves by exogenous jumps and by the permanent impact of uninformed lit market orders, $\lambda=0.0005$ ticks a share (informed orders reveal $P^\ast$, they do not move it). Liquidity providers quote on LIT; noise and informed takers try the dark pool first half the time and send what is left to LIT; patient uninformed traders rest midpoint pegs in the dark. Over four seeds of 45 minutes, informed takers who tried the dark were filled for 6.3% of their shares, uninformed ones for 17.5%: the informed buy when everyone informed buys, and the dark sellers run out. Resting in the dark gives up the spread, but its fills were followed by a move of $+0.11$ cents in the resting trader’s favour over 30 seconds, against $-0.13$ cents for lit providers, who capture 0.51 cents of spread and pay adverse selection on it.

9.4 Information leakage

A fund buys 20 000 shares from ten minutes into the session over 30 minutes, four ways: lit, a market order of 200 every 18 seconds; dark, one resting midpoint peg for the whole order; dark with a minimum, the same peg with a minimum of 1 000; block, conditionals in a block venue where natural sellers post conditionals of 5 000 to 15 000 shares every five minutes on average, 30% of whom never firm up and buy 2 000 shares on LIT instead. Whatever is left after 30 minutes is bought on LIT at once. A pinger probes the dark pool every two seconds with 100 shares, alternating sides; when two probes in a row on one side fill within five seconds, it buys (or sells) 500 shares on LIT ahead of the order it has found.

            if s == PINGER and venue == "DARK":
                self.pings += 1
                fund = self.dark.trades[-1][4] == FUND       # whose order the probe hit
                self.fund_pings += fund
                last = self._last_ping.get(side, (-1e9, False))
                if last[0] > self.t - 5.0:                   # two fills in a row: a large order
                    self._market(PINGER, -side, 500, False)  # trade ahead of it on LIT
                    self.front_runs.append((self.t, 500, fund or last[1]))
                self._last_ping[side] = (self.t, fund)

The shortfalls are noisy, so the simulation also accounts for the two costs it knows. Each lit buy of $q$ moves $P^\ast$ by $\lambda q$ and raises the price of every share the fund buys afterwards: summed over the fund’s own lit children this is its own-impact cost, $\lambda Q/2\,(1-1/n)=4.95$ basis points for 100 children; summed over the pinger’s and the leaky sellers’ trades ahead of it, when they were triggered by the fund, it is the leakage cost. Over sixteen seeds with the same exogenous events for every strategy (Figure 9.2):

• Lit: shortfall 3.8 basis points (standard error 1.4), drift of the mid over the window 7.3 (2.3), own impact 4.95, no leakage.
• Dark: the peg completes in about 360 seconds, but the pinger finds it: 88 of its probes fill against the fund, it buys 44 000 shares ahead, and the leakage costs 10.9 basis points (0.3). Shortfall 12.5 (0.5), drift 26.1 (2.2).
• Dark with a minimum of 1 000: no probe can fill it and there is no leakage, but only 46% of the order fills in the dark; the rest is bought at the end, at 1.4 basis points of own impact. Shortfall 9.5 (1.6), drift 11.9 (1.6).
• Block: 79% crosses in blocks, 0.75 leaky invitations per run cost 0.6 basis points. Shortfall $-2.9$ (1.9), drift $-1.7$ (2.2).

The lessons hold beyond the model’s numbers. A resting dark order is visible to whoever probes for it, and the probe costs the prober 100 shares at the midpoint; the leak can cost more than the impact the dark venue was chosen to avoid. A minimum quantity above the probe size stops the leak and costs completion. And a midpoint fill saves the spread but not the impact when the counterparty is flow that would otherwise have sold on the lit book: taking it out of the lit market moves the price the same way. Blocks are cheap in the simulation because their sellers bring size that would not otherwise have traded: the cost of a trade is the imbalance it creates, wherever it is printed.

9.5 Dark pools in enforcement records

The secret is the product, and enforcement records show what happens when operators do not keep it. On August 12, 2015 ITG and its affiliate AlterNet Securities agreed to pay USD 20.3 million, admitting that an undisclosed proprietary desk, Project Omega, had accessed live feeds of its dark pool subscribers’ orders and executions and traded against them in POSIT. On January 31, 2016 Barclays agreed to pay USD 70 million, half to the SEC and half to the New York Attorney General, for misrepresenting how its Liquidity Profiling policed its LX pool, and Credit Suisse USD 84.3 million over its Crossfinder pool, where it misrepresented its Alpha Scoring of subscribers’ flow, accepted over 117 million sub-penny orders and alerted two high-frequency firms to customer orders through its Crosslink service. Each is a case of information leakage by the venue itself.

9.6 Tutorial: the block that leaked

9.7 Build: the dark venue

9.8 Exercises

9.9 Problem: The Block That Leaked

9.10 Interview questions