Product × Finance × AI: Financial Products in the Age of AIPart Two Changing one parameter中文
Chapter 17

Agents become the main market participantsExecution costs, crowding and moving in the same direction

More and more of the buying and selling in the market square is done by scribes. They find counterparties for their principals, split large orders, and quote buying and selling prices at any moment. Finding a counterparty costs almost nothing any more, so spreads narrow and ordinary households trade more cheaply. But the scribes come from a few operators, read the same reports and use the same methods, so they reach the same judgment at the same moment. A single fall triggers many scribes to sell at once, and the selling deepens the fall; the scribes that quote prices for others withdraw to protect themselves, precisely when they are most needed. Execution becomes cheaper, yet the market may become more fragile, and so it needs new designs.

1. Scribes place orders directly

Finding a counterparty willing to trade used to take time and require concessions on price. Now only one thing changes: more and more of those placing orders for principals and quoting prices in the market are scribes. They can watch many trades at once, quote at any moment, and split a large order into many small ones to sell gradually.

Work through how the spread changes. The spread is the compensation that the party standing ready to trade charges for the time and risk of holding inventory. A scribe quotes at almost no cost and turns its inventory over quickly, so it holds inventory only briefly; many scribes are willing to quote, and they compete with one another, so the spread is pushed very narrow. On calm days, the cost for an ordinary household of buying or selling one ship share falls markedly.

Today's counterpart. The U.S. Securities and Exchange Commission's 2010 concept release on market structure stated that estimates of high-frequency trading's share of stock market volume varied widely but were typically 50% or more.

Narrow in calm times, wide in a hurry

Drag the number of scribes quoting prices in the market, and see how much someone in a hurry to sell one ship share must give up, in calm times and when risk suddenly rises.

Illustrative: when only one household quotes, the spread is 10%. The more quoters there are, the more they compete, and the calm-time spread narrows as 10% ÷ (number of quoters to the power 1.2), to a minimum of 0.2%. When risk suddenly rises, 90% of scribes stop quoting, and the spread is calculated from the quoters that remain.

2. Scribes that resemble each other

Many households' scribes come from a few operators. They read the same reports, score with the same methods, and even take each other's trades into account. In the past, each household in the market square had its own judgment: some expected prices to rise, others expected them to fall, and their trades offset one another. Now the same news leads many scribes to the same conclusion at the same time.

Even when each scribe has a little information of its own, it may abandon it. Work it through: the first scribe buys on its own information. The second scribe's own information also points up, so it buys. The third scribe's own information points down, but seeing that the first two have both bought, it judges that others know more than it does, and buys as well. From this one on, every later scribe looks only at what those before it did, and its own information no longer enters the price. A long run of buying may rest on nothing more than the little information held by the first two, and a single new piece of news can reverse the whole run.

Today's counterpart. A 2024 report by the Financial Stability Board noted that artificial intelligence may amplify correlations in markets: many institutions using similar models and data may react in the same way to the same shock.

Following those ahead

A certain ship's prospects are in fact poor. Twenty scribes decide in turn whether to buy. Each has a clue of its own that is right 70% of the time, and each can see what the scribes before it did. Switch the source of the clues, try a few sets of clues, and see how judgments form.

Do not buy (correct)BuyDrops its own clue and follows those ahead

Illustrative: each scribe infers its predecessors' clues from their actions and adds its own clue to reach a judgment. When the inferred favourable and unfavourable clues differ by two or more, its own clue can no longer change the conclusion, so it follows those ahead (a simplification of Bikhchandani, Hirshleifer and Welch, 1992). With "the same source", every scribe receives the same clue.

3. Crowding through the same door

Scribes that resemble each other also hold similar ship shares. In normal times this is no problem. But once one of them has to sell for some other reason, the price of the ship shares falls, the other scribes holding the same shares also lose money and cut their positions according to their own rules, and the price falls further. The door is only so wide, and everyone wants to get out at the same time.

Khandani and Lo (2011) studied the collective losses of quantitative hedge funds in August 2007. Many funds used similar long-short equity strategies. The authors infer that the rapid unwinding of one or a few large portfolios pushed down the prices of similar positions, so that other funds holding similar positions suffered large losses at the same time.

Today's counterpart. The more similar the strategies and the more concentrated the holdings, the wider the damage spreads when one participant is forced to sell.

Crowding through the same door

One of ten portfolios is forced to sell all its holdings. Drag how much the portfolios' holdings overlap, and see the losses of the other nine portfolios, and the second-round losses after they follow their rules for cutting positions and sell too.

First-round lossSecond-round loss after following the selling

Illustrative: the forced sale makes the ship shares it held fall 10%. Other portfolios' first-round loss = 10% × degree of overlap. Portfolios whose loss exceeds 3% sell half their holdings, and each portfolio that does so pushes the overlapping part down a further 1.5% (a simplification of the unwinding process described by Khandani and Lo, 2011). Columns are numbered portfolios 2 to 10.

4. Falls that cause falls

Many scribes follow risk rules: cut the position when losses exceed a certain percentage; sell when margin runs short. Each rule is reasonable on its own, but together they form a loop. Prices fall, scribes lose money, and the rules require selling; selling pushes prices down, more scribes lose money, and more selling follows. Scribes that borrowed to buy ship shares are hit harder still: when prices fall, the lender demands more margin, and the scribe has to sell some more.

Today's counterpart. When prices fall sharply, margin calls and forced liquidations tend to cluster, which deepens the fall.

Falls that cause falls

After an initial fall of 2%, scribes' losses, tightening margins and rules for cutting positions bring new selling. Drag how much buying is done with borrowed money, and see the additional fall in each round and the final total fall.

Prices fall→Scribes lose money→Margins tighten, rules require cuts→Selling↺

Illustrative: the selling caused by each round's fall makes the next round fall by k times this round's fall, where k = 0.15 × leverage. When k is less than 1, the total fall is 2% ÷ (1 − k); when k is 1 or more, the spiral does not stop by itself, and the chart shows only the first ten rounds (a simplification of the liquidity spiral described by Brunnermeier and Pedersen, 2009).

5. Liquidity vanishes when it is most needed

The scribes that quote prices in normal times are the first to withdraw when risk suddenly rises. They do not know why prices are falling, the risk of holding inventory has suddenly grown, and the safest course is to stop quoting or to pass on the ship shares they have just bought as quickly as possible. So the spread, narrow in normal times, becomes astonishingly wide just when trading is most needed.

The joint report by the U.S. Securities and Exchange Commission and the Commodity Futures Trading Commission on the "Flash Crash" of 6 May 2010 records the following. A mutual fund used an algorithm to sell 75,000 S&P 500 index futures contracts, worth about $4.1 billion. The algorithm sold at 9% of the previous minute's trading volume, taking account only of volume, not of price or time. High-frequency traders first bought, then rapidly passed contracts among themselves, trading more than 27,000 contracts within 14 seconds while buying only about 200 on net. Buy-side order book depth fell to less than 1% of its level that morning, and futures prices fell a further 1.7% in 15 seconds. After the exchange triggered an automatic mechanism that paused trading for 5 seconds, selling pressure eased and prices began to recover. Using that day's trading records, Kirilenko, Kyle, Samadi and Tuzun (2017) found that high-frequency traders did not trigger the crash, but their responses amplified the volatility.

Today's counterpart. After the Flash Crash, the U.S. Securities and Exchange Commission approved the "limit up–limit down" mechanism in 2012 (in effect from 2013) and revised the market-wide circuit breaker rules.

A quarter of an hour on the afternoon of 6 May 2010

The S&P 500 index futures market as described in the joint report of the U.S. Securities and Exchange Commission and the Commodity Futures Trading Commission. Click a moment to see the details.

Source: U.S. CFTC & SEC (2010), Findings Regarding the Market Events of May 6, 2010. Times are p.m. U.S. Eastern Time.

6. Giving the market time to breathe

Since spirals come from selling in the same direction at the same moment, the way to break them is to slow time down. When prices fall too far within a very short time, trading is halted, giving scribes time to reassess and giving those willing to buy at low prices time to arrive. Continuous matching can be replaced by matching in a batch at very short intervals, so that being a moment faster loses its advantage (Budish, Cramton and Shim, 2015). Another approach is to keep participants diverse: the more scribes there are with different methods, the less likely a single piece of news is to set off action in the same direction.

Halts have costs too. Those who want to sell cannot sell for a while, and prices may jump further when the halt ends. The threshold and length of a halt must balance breaking the spiral against obstructing trade.

Today's counterpart. The current market-wide circuit breakers in U.S. stock markets trigger when the S&P 500 index falls 7%, 13% and 20% from the previous day's close. Under the single-stock limit up–limit down rule, trading pauses for 5 minutes when the price stays at the limit band for more than 15 seconds.

The circuit-breaker threshold

In a market where half the scribes come from the same operator, the same large order sets off a plunge. Drag the circuit-breaker threshold and see when the breaker triggers and how large the biggest fall is.

Illustrative: the model is the same as in "A market of scribes", with half the scribes coming from the same operator; each halt lasts 3 steps. A threshold of 12 points is equivalent to having no circuit breaker.

7. A market of scribes: Civilization Simulator, version 17

The simulator unlocks "Agents as market participants". Forty scribes buy and sell the same ship share in the market square; some of them come from the same operator and use the same method. All of them follow a rule of cutting positions when losses mount. One day, a large holder's scribe must sell a big block of ship shares. You can adjust the share of scribes from the same operator and switch the circuit breaker on or off, then watch how low the price falls that day and how quickly it recovers.

Civilization Simulator · Chapter 17: a market of scribes

Forty scribes buy and sell the same ship share, whose fundamental value is 100. From step 5, a large holder's scribe sells a big order in each of ten consecutive steps. Adjust the share of scribes that come from the same operator and use the same method, switch the circuit breaker on or off, and watch the price over the day.

MarketParameter: price of cognitive labourGuaranteeing agentsAnalysts for allAgents as market participantsAgents and human bias · Ch. 18
PricePrice when scribes are not alikeCircuit-breaker halt

Illustrative model: every scribe follows the rule "sell once losses reach a certain percentage". Scribes from the same operator share the same threshold and trigger at the same time; the other scribes' thresholds are spread between 4% and 16%. One third of the scribes quote prices in the market, and when the price falls more than 1.5 points in one step, they withdraw three quarters of their quotes. Households willing to buy at low prices buy more the lower the price. During a halt, the price recovers one fifth of the way towards fundamental value, and selling rules do not trigger.

8. What this step leaves open

Once scribes become the main market participants, several questions remain unresolved:

  1. Human bias. Can scribes help principals stick to their plans and not sell along with everyone else in a panic? Or will they be used to exploit human biases?
  2. Sources of diversity. Diversity among participants reduces action in the same direction, but who has an incentive to maintain that diversity?
  3. Coordinating rules. Each scribe's risk rules are reasonable, yet together they form a spiral. Who looks at the whole picture?
  4. What does not change. Ships can still sink; the risk of a fall does not disappear because of scribes.

What this step leaves open

Four questions that remain unresolved once scribes become the main market participants. Click a question to see the assumption it hides.

A market of scribesCheap but fragile

9. Implications for products

As more and more of users' trades are executed by agents, a product faces not only the risk to each individual user but also the risk of many agents acting at once. Three implications follow. First, the fall in execution costs is real, but it shows up mainly on calm days; the product should let users know that in violent swings the price they get may be far from the usual one. Second, an agent's risk rules should not look only at a single user's account; they should also avoid firing at the same moment as large numbers of similar agents. Building delays, batching and price protection into the rules is safer than a hair trigger. Third, the more similar agents' methods and data are, the more fragile the market becomes. For a product, preserving each user's own agreement and differences, rather than putting every user on the same strategy, is itself a contribution to the market.

Chapter summary: agents become the main market participants

Four conclusions, each corresponding to one mechanism in the chapter.

Rerun: the spreadCheaper in normal times

Quoting and holding inventory cost scribes almost nothing, so spreads in calm times narrow sharply.

Information cascadesScribes that resemble each other act at once

The same methods and data mean judgments are no longer independent, and uniform action may rest on very little information.

Liquidity spiralsFalls that cause falls

Rules for cutting positions and for margin trigger at the same time, and quoting scribes withdraw just when they are most needed.

Market designGive the market time to breathe

Halts and circuit breakers break the spiral, and diversity among participants reduces action in the same direction.

References