25Circuit Breakers and Halts

On 9 March 2020 the S&P 500 fell seven per cent minutes after the open and all US equity trading stopped for fifteen minutes; it happened three more times that month. Whether stopping a falling market calms it or makes it fall faster toward the stop is an old argument with new data. This chapter sets out the designs, builds them as venue-side policies for the exchange simulator (firm.halts), and runs a crash through each, measuring in particular whether sellers rush toward a band before it stops them: the magnet effect.

25.1 Designs: bands, pauses, halts, limits

A band stops trades from printing outside a range; a pause stops them altogether for a while; a halt (One Quant Book 8, chapter 17) stops them until the venue or the regulator decides; a price limit (One Quant Book 1, chapter 12) is a band that holds all day. They differ in what they protect against: a band against an erroneous or runaway order, a pause against a price moving faster than liquidity can return, a halt against news the market has not digested.

firm.halts implements them as policies that firm.exchsim loads (the halts field of the venue’s configuration): the policy watches the venue’s engine every half second through a scheduled call and sends the engine’s own controls, R to set a band (orders priced outside it are rejected and matching stops at it) and P to change the phase (halted, reopening call, trading). Leaving the call uncrosses it with chapter 10’s auction.

        if self.trades:
            ref = sum(p for _, p, _ in self.trades) / len(self.trades)
        else:
            ref = self.ref if self.ref is not None else st.ref_price
        if self.ref is not None and abs(ref / self.ref - 1) < self.update:
            ref = self.ref                                # the reference moves only by update_pct or more
        self.ref = ref
        self._set_band(t_ns, ref)
        lo, hi = self.band
        bid, ask = st.book.best(1), st.book.best(-1)
        if bid is not None and ask is not None:
            self.dist.append((t, (bid - lo) / st.tick, (ask - lo) / st.tick))
        limit = (ask is not None and ask <= lo) or (bid is not None and bid >= hi)
        if self.shadow:
            if limit and self.limit_since is None:
                self.log.append((t, "limit", lo))
            self.limit_since = t if limit else None
            return
        if limit:
            if self.limit_since is None:
                self.limit_since = t
                self.log.append((t, "limit", lo if ask is not None and ask <= lo else hi))
            elif t - self.limit_since >= self.ls:
                self._send(t_ns, CTL["P"](self.locate, "U", "LULD"))
                self.paused_until = t + self.pause
                self.limit_since = None
                self.log.append((t, "pause", self.pause))
        else:
            self.limit_since = None

25.2 Single-stock and market-wide mechanisms

In the United States, single stocks trade inside limit up-limit down bands (One Quant Book 1, chapter 31), and the whole equity market stops when the S&P 500 falls far enough from the previous close.

The simulation scales the mechanisms down to its five-minute sessions: a stock at USD 100.00 with a one-cent tick, a crash in which a seller sends 40 000 shares as sixty market orders over a minute, and four regimes: no mechanism; a dynamic band of 0.15% around the last minute’s average trade price, with a limit state (the best offer at or below the lower band, or the best bid at or above the upper) that pauses the stock for 20 seconds if it lasts 3; velocity logic (below); and a market-wide halt of 30 seconds at falls of 0.1%, 0.2% and 0.3%.

25.3 Futures: velocity logic and stop logic

CME’s Globex describes velocity logic as detecting market movement of a predefined number of ticks up or down within a predefined time, and moving the instrument into a reserved state in which customers can submit, modify and cancel orders. Stop logic does the same when the triggering of stop orders (chapter 2) would move the price by more than a set amount. The simulation’s velocity policy pauses for 10 seconds after a move of 8 ticks within 2 seconds (firm_halts.Velocity).

Each mechanism shortens the fall by taking orders out of the market (Table 25.1). The band does it without a single pause: market orders stop at its lower edge and the rest of each order is cancelled, so the seller sells fewer shares at better prices. Velocity logic pauses often (79 seconds of five minutes) and helps least: each reopening auction meets the same selling. The market-wide halt cuts the fall most and the seller’s volume most, which is the policy question in one line: a mechanism that protects prices does so by refusing trades.

25.4 Magnet effects in the evidence

Subrahmanyam (1994) showed in a model that circuit breakers may cause agents to advance trades in time, and so increase price variability and exacerbate the moves they are meant to stop; Cho, Russell, Tiao and Tsay (2003) looked for this magnet effect of price limits in high-frequency data from the Taiwan Stock Exchange. The simulation adds the mechanism behind the argument: sellers who, when a band is published and enforced, sell faster the closer the best bid is to its lower edge (up to twice their base rate within one tick of it), for fear of being unable to sell once it binds.

The measure is a first passage: for each session and each distance $x$ from the lower band, the first time the best bid comes within $x$ ticks; did it come within one tick of the band in the next 30 seconds? Three conditions, 80 sessions each: the band enforced with the anticipating sellers, the band enforced without them, and no band (the same band computed but not enforced, and no anticipation).

Figure 25.2 separates two effects. The magnet effect, the anticipators’ contribution (band and anticipators minus band alone), averages $+0.03$ over distances of 3 to 7 ticks: at 5 ticks, 0.66 against 0.63, with standard errors of 0.05 and 0.06. It has the sign the theory predicts at nine of the eleven distances and is too small to tell from zero in 80 sessions. The band’s own effect (band alone minus no band) averages $-0.12$, and is negative at ten of the eleven distances: at 5 ticks, 0.63 against 0.75. In this market the band repels more than it attracts, because it takes the sellers’ market orders out of the book at its edge, which is stronger than the extra selling it invites. A real magnet needs more traders who anticipate, or a band that is not also a barrier; the simulation says how to look for one: condition on the distance, compare with a market without the mechanism, and separate the behaviour from the mechanics.

25.5 Reopening

A pause or halt ends in a reopening auction, and its design decides whether the pause did any good. Chapter 10’s auctions apply: an indicative price published during the call, collars that stop the reopening price from jumping too far, and extensions when imbalance is too large. Brugler, Linton, Noss and Pedace (2018), with London Stock Exchange data, found that single-stock circuit breakers improved the liquidity and reduced the volatility of the other stocks that kept trading: the pause’s benefit may be felt outside the paused stock. In the simulation, the velocity pauses’ reopenings met the same selling and the fall resumed: a pause that ends before the other side of the market has arrived only postpones.

25.6 Tutorial: stopping a crash

25.7 Build: halts

25.8 Exercises

25.9 Problem: Fifteen Minutes of Silence

25.10 Interview questions