<<

Downloaded by guest on September 26, 2021 t ob ae hr r oicnie o rirgus othere PNAS so arbitrageurs, for incentives no are there prof- profits made, no excess are be make there to if to its But, possible information. available not publicly is on it based competition evolve where rapidly that equilibrium to standard an markets is cause to A efficiency should arbitrageurs market efficiency. by profits justify market for to of used theory argument the old in as strat- contradictions and each change, in appear. regulations invested strategies as new wealth and species. time, fail the setting, a strategies through price as of changes via evolves population egy another market the one like the with is and interact strategy strategies cap- a investing, Trading The in others. invested many trading, and ital , technical into statistical classified as making, market similarly, be such can exploit food; categories, to strategies provide distinct evolve Trading that that inefficiencies. specialists niches market are fill strategies spe- to trading are financial animals evolve and strategies that Plants trading species. cialists Financial biological to markets. analogous financial and are biology to from them methods and alternative. applies concepts necessary borrows the approach provides which This the- ecology, the markets market of why beyond ory and go how to understand need to we efficiency malfunction. that market of informa- means new ory This fundamental little from (2). substantially very deviate values is often prices there and outside (1), when factors price tion even Large by otherwise. occur caused suggests movements are evidence Empirical setting market. price the with definition, problems by when Thus, only any values. change fundamental affects and information values new fundamental reflect always Prices W from ecology prices market of deviations and mar- The the how values. price fundamental into others. excess explain insight the gives of ecosystem of and origins market occur returns spontaneously the the inefficiencies increase ket increases of an one dynamics that of wealth meaning wealth relationship, the mutualistic in market a into as strate- have three insight all such gies give equilibrium, efficient ecology, webs, the at from food example, For behavior. and concepts perfect matrix core from community how away the term, time show of We the periods in efficiency. extended Even real spends uncertain. match market would and to the statisti- noisy calibrated market the this is the makes but (which strategy noise, markets) profitability equilibrium, each of in efficient uncertainty in absence an cal the invested toward In wealth evolve time. the slowly given on of any depend consisting at market they dependent; density a is, strongly are of that that strategies show model of We returns toy traders. average noise the a and abundance followers, study the trend , like We value is species. strategy financial a a alternative of in an which invested in biologists, wealth develop by the developed we methods and Here concepts efficiency. on 2020) based based 16, and are September markets review equilibrium for financial (received of on 2021 theory 24, the March to approved approaches and Standard NJ, Princeton, University, Princeton Levin, Asher Simon by Edited 87501 NM Fe, Santa Institute, a malfunction Scholl P. market Maarten explains ecology market How nvriyo xod xodO13D ntdKingdom; United 3QD, OX1 Oxford Oxford, of University nttt o e cnmcTikn,Ofr atnSho,Uiest fOfr,Ofr X Q,Uie Kingdom; United 3QD, OX1 Oxford Oxford, of University School, Martin Oxford Thinking, Economic New for Institute h hoyo akteooyeegsfo h inherent the from emerges ecology market of theory The (3–7) work earlier on build we Here 01Vl 1 o 6e2015574118 26 No. 118 Vol. 2021 yd akt afnto?Acrigt h theory perfectly. the function always to markets According efficiency, malfunction? market markets of do hy | aktefficiency market a,b,1 nsaaCalinescu Anisoara , | gn-ae modeling agent-based ∗ n eeo h the- the develop and b c ahmtclIsiue nvriyo xod xodO13D ntdKndm and Kingdom; United 3QD, OX1 Oxford Oxford, of University Institute, Mathematical n .DyeFarmer Doyne J. and , aktefiinyi eyso.Teepce eitosfrom deviations expected The slow. very is efficiency market causes under evolution time malfunction. time to through market endogenous evolves the their wealth how invested their and wealth how reinvestment, the strategy, on an depend each that strategies in the way the of same how returns study the We average species. in interacting three problem study would the ecologist approach We strategies. of ecology market model predict from and better ideas phenomena a market how behavior. interpret construct show to used to to be rather not can but is markets, goal financial Our understanding for models. mar- framework of conceptual such theory a The provides 9–12). ecology refs. ket (e.g. markets financial than of rather models complement mar- a of as theory naturally viewed the be ecology substitute. than can a market problems and of of efficiency set theory ket different The how a time. and inefficiencies addresses with market market cause change the necessary strategies affects to is between strategy it interactions each functions, the as market how change the understand wealth inefficiencies how to the the understand as changes, To and strategy well. appear each market strategies in exploit new invested strategies under as trading contrast, but, ecology, In inefficiencies, market (8). approximate of efficient some theory perfectly in sug- the be efficient paradox cannot be This they may efficient. sense, markets markets while make that, to gests mechanism no is www2.econ.iastate.edu/tesfatsi/blake.sfisum.pdf 2021. 25, June Published doi:10.1073/pnas.2015574118/-/DCSupplemental at online information supporting contains article This 1 the under Published Submission.y Direct PNAS a is article This interest.y competing research, no declare performed authors paper.y The research, the wrote designed and J.D.F. tools, analytic and new contributed A.C., M.P.S., contributions: Author B eao,Bidn h at eatfiilsokmarket. artificial Fe Santa the Building LeBaron, *B. owo orsodnemyb drse.Eal [email protected] Email: addressed. be may correspondence whom To e.Ti rvdsacneta rmwr htgvsinsight malfunction. markets gives why that reasons framework the conceptual into a provides val- This fundamental and from ues. volatility deviate of prices bursts where fluctuations periods causing large extended state, are efficient there the that from noisy away so is strategies evolution slowly the all happens and this where However, returns. state average same efficient the make an system toward the evolve that should suggest dynamics each deterministic in invested The wealth strategy. the on depending predator– mutualistic, competitive, or be prey, can they eco- that in showing strategies terms, trading logical stereotypical ana- of apply We interactions markets. to the financial lyze how to show ecology from and concepts species standard biological and strategies trad- financial ing between analogy mathematical the develop We Significance eso ht ihraitcprmtr,eouintoward evolution parameters, realistic with that, show We trading three with model market toy stylized a study we Here agent-based on work of body large a on builds here study Our a,c,d NSlicense.y PNAS https://doi.org/10.1073/pnas.2015574118 b optrSineDepartment, Science Computer . y (2002). https://www.pnas.org/lookup/suppl/ d at Fe Santa | f9 of 1

ECONOMIC SCIENCES SPECIAL FEATURE efficiency are, in some sense, small, but they persist even in A the long term, and cause extended deviations from fundamen- tal values and excess volatility (which, in extreme cases, becomes market instability). Our study provides a simple example of how analyzing markets in these terms and tracking market ecosys- tems through time could give regulators and practitioners better insight into market behavior. Model Description The structure of the model is schematically summarized in Fig. 1. There are two assets, a stock and a bond. The bond trades at a fixed price and yields r = 1% annually in the form of coupon payments that are paid out continuously. The stock pays a D(t) at each time step that is modeled as a discrete time-autocorrelated geometric Brownian motion, of the form

D(t) = D(t − 1) + gD(t − 1) + σD(t − 1)U (t), [1] U (t) = ωU (t − τ) + (1 − ω2)Z (t),

where g is the average rate of dividend payments, σ is the vari- ance, ω is the autocorrelation parameter of the process, and Z is Gaussian noise. The auxiliary process U (t) introduces persis- tence. We choose parameters so that one time step is roughly equal to a day. We use estimates from market data by LeBaron (13), taking g = 2% per year for the growth rate of the dividend Trend Follower with a volatility of σ = 10% (see ref. 14, for example, for a review of the empirical evidence on ). The dividend process is ≤ 0% 2% 4%%%6 8 ≤ 10 % positively autocorrelated through the auxiliary process U (t) with return rate (annual) lag τ = 1 d. Trend Follower wealth We use market clearing to set prices. The stock has a fixed 38% 33% 28% 23% 18% 13% 8% supply Q, but the excess demand E(t) for the stock by each B varies in time. We allow the trading strategies 2.5% to take positions and to use leverage (i.e., to borrow in 0% order to take a in the stock that is larger than their -2.5% wealth). We impose a strategy-specific leverage limit λ¯. Because we use leverage and because the strategies can have demand 7.5% Trend Follower functions with unusual properties, market clearing is not always 5.0% Value straightforward—see Materials and Methods. 2.5% The size of a trading strategy is given by its wealth W (t), 0% that is, the capital invested in it at any given time. In ecology, Return Volatility 19% 24% 29% 34% 39% 44% 49% this corresponds to the population of a species, which is also Value Investor wealth called its abundance. Unless otherwise stated, the wealth of each strategy varies proportional to its cumulative performance. Let- Fig. 2. The profitability of the dominant strategy in the wealth landscape. A is a ternary plot which displays the returns achieved by the strategy with ting πi (t) be the return of strategy i at time t, the wealth changes the largest return. The axes correspond to the relative wealth invested in according to each strategy. The top corner is pure trend followers, the left corner is pure noise traders, and the right corner is pure value investors. The color indicates Wi (t + 1) = (1 + f πi (t))Wi (t). [2] the strategy with the highest returns at a given relative wealth vector w. The regions colored in red correspond to the noise traders, blue regions The reinvestment rate f models investor flows of capital. The correspond to value investors, and green regions correspond to the trend default value f = 1 corresponds to passive reinvestment, and f > followers. The intensity of the color indicates the size of the average return. 1 means that profitable strategies attract additional capital and (B) Upper shows the average returns to value investors (blue) and trend unprofitable strategies lose additional capital. followers (green) while holding the wealth at its equilibrium A trading strategy is defined by its trading signal φ(t), which value of 42%. Lower shows the volatility in the returns of each strategy. The can depend on the price p(t) and other variables, such as div- horizontal axis is the relative wealth of the trend follower (top axis) and value investor (bottom axis). idends and past prices. We modify φ by a tanh function to

ensure that the excess demand is bounded and differentiable. A strategy’s excess demand for the stock is

W (t − 1)λ¯  1  E(t) = tanh (c · φ(t))+ − S(t − 1), [3] p(t) 2

where S(t − 1) is the number of shares of the stock held at ¯ Fig. 1. The three trading strategies correspond to noise traders, value the previous time step, and λ is the strategy-specific leverage investors and trend followers. They invest their capital in a stock and a bond. limit. The parameter c > 0 determines the aggressiveness of the The mixture for each strategy changes with time as strategies accumulate response to the signal φ, and is strategy specific. When the sig- wealth based on their historical performance. nal of the strategy is zero, the agent is indifferent between the

2 of 9 | PNAS Scholl et al. https://doi.org/10.1073/pnas.2015574118 How market ecology explains market malfunction Downloaded by guest on September 26, 2021 Downloaded by guest on September 26, 2021 fe 0 .Tesse siiilzda admwt nfrl itiue elhvco n hnalwdt reyeov o 0 .Teclrsdarkness malfunction color’s market The explains y. ecology 200 market for How evolve freely al. to et allowed Scholl then and and y, vector 200 wealth = distributed T uniformly for a system with the random follow density. color at conditions, to same initialized initial proportional the is distributed is with system uniformly trajectories The The from y. wealth. Starting 200 initial time. after the indicates with of “+” increases dynamics marker saturation deterministic The color coded. the color that each so vectors, uncertainty, wealth initial statistical to due noisy change. extremely of rate the denotes to seek They stock. the of value the derive to model a use and investors. using investors Value turn. all in strategy however, representing each fund; describe as single now of We a strategies. thought by these used be only should were these it each though treating as hypothesis, not agent are strategy representative strategies a their use and We is, information, optimal. that limited rational, from boundedly work traders. strategies they noise all and make followers, intentionally trend We investors, value call we which Strategies. process, demand Walrasian excess usual the in i.3. Fig. the is between time equally given portfolio of any its term the splits (hence and two bond the and stock rfi yaisa ucino wealth. of function a as dynamics Profit P au netr bev h iiedprocess dividend the observe investors Value i p E (t i )S (t D esuytretpcltaigstrategies, trading typical three study We ) .Televerage The 1/2). ipasteoetm-tpatcreaino rc eun.Tebakdti h qiiru on from point equilibrium the is dot black The returns. price of autocorrelation one-time-step the displays (t B vralagents all over )/W hw apetaetre o e ifrn nta auso h elhvco,mkn tcerta h rjcoisare trajectories the that clear it making vector, wealth the of values initial different few a for trajectories sample shows (t p = ) (t ) λ| ¯ sstsc htaggregate that such set is ah(c tanh i szero. is A λ(t hw o elheovs naeae hog ieudrrivsmn.Teitniyo h color the of intensity The reinvestment. under time through average, on evolves, wealth how shows · ) φ(t fasrtg at strategy a of As 1/2|. ))+ h icutdepce uuedividends, future expected discounted the their of dividends. historical parameters of on The more based overvalued. estimated hold is are and stock model undervalued the is when it bond when the stock the of more hold inlrssascae ihtesok eflo e.1 n use and data. addi- historical 15 by ref. the implied follow for rate We expect discount stock. fixed model the a our with in associated risks investors tional the of all premium with return, parameter The h udmna value fundamental The A r orapoiain h iulzto ipastredifferent three displays visualization The approximation. poor a are C k ipasadniympo h smttcwat distribution wealth asymptotic the of map density a displays ≥ r k ; sadson aecle the called rate discount a is k V k stesmo h ikfe rate risk-free the of sum the is 0 = (0) 2% = V E ae nteaeaert freturn of rate average the on based , (t " X t ) ∞ =1 https://doi.org/10.1073/pnas.2015574118 ftesokat stock the of 1+ (1 D (t k ) ) t # . A. t eurdrt of rate required 0 = PNAS r sgvnas given is n risk a and | f9 of 3 [4]

ECONOMIC SCIENCES SPECIAL FEATURE Table 1. Estimated community matrix near the equilibrium at This process reverts to√ the long-term mean µ = 1 with rever- w = (NT = 0.43, VI = 0.34, TF = 0.23) sion rate ρ = 1 − 6×252 0.5, meaning the noise has a half life

Gij NT, % VI, % TF, % of 6 y, in accordance with the values estimated by Bouchaud et al. (22). The variable  is a standard normal random vari- NT −0.89 0.89 0.82 able, and γ = 20% is a volatility parameter, chosen so that noise VI 26.6 −10.6 22.4 traders generate volatility in excess of the volatility of dividends, TF 11.1 15.2 −19.3 matching the level observed in the US . NT, noise traders; VI, value investors; TF, trend followers. The parameters of the model are summarized later (see Table 5). We have chosen them for an appropriate trade-off between realism and conceptual interest, for example, so that each strat- V VI(t) The valuation used by the value investors is made by egy has a region in the wealth landscape where it is profitable. We g estimating the mean growth rate from past data using the will see that there are a few ways in which the properties of these [D(t)] = D(0)(1 + g)t classical (16) E . strategies do not match those observed in real markets (the trend ω This ignores the autocorrelation coefficient , and so, in general, strategy is very short term, the value investor uses somewhat high V VI(t) 6 =V (t) . leverage, and, at equilibrium, the noise trader has a surprisingly We define the trading signal for the value investor as the dif- low return-to-risk ratio). Nonetheless, they are realistic enough ference in log prices between the estimated fundamental value to make our key points. V VI(t) and the market price. Results VI φVI(t) = log2 V (t) − log2 p(t). [5] Density Dependence. An ecosystem is said to have density depen- dence if its characteristics depend on the population sizes of the This strategy enters into a long position when the proposed price species, as is typically the case. Similarly, a market ecosystem is is lower than the estimated fundamental value and enters into a density dependent if its characteristics depend on the wealths of short position when the proposed price is higher than the esti- the strategies (both its own wealth and that of other strategies). mated fundamental value. The use of the base two logarithm The toy market ecosystem that we study here is strongly density means that the value investor employs all of its assets when the dependent. stock is trading at half the perceived value (17). When the core ideas in this paper were originally introduced in ref. 3, prices were formed using a market impact function, which Trend followers. Trend followers expect that historical trends translates the aggregate trade imbalance at any time into a shift in returns continue into the short-term future. Several vari- in prices. This can be viewed as a local linearization of market ants exist in the literature, including the archetypal trend fol- clearing. The use of a market impact function suppresses density lower that we use here (3, 10, 18–20). There is evidence to dependence and neglects nonlinearities that are important for suggest that trend-based investment strategies are profitable understanding market ecology. over long time horizons, and ref. 21 argues that investors earn With market clearing, there is strong density dependence. a premium for the liquidity risk associated with with This is evident in Fig. 2, which shows which strategy makes high (momentum trading is a synonym for trend the highest profits as a function of the relative size of each following). of the three strategies. To control the size of each strategy, The trend-following strategy extrapolates the trend in price we turn off reinvestment, and instead replenish the wealth between recently realized prices p(t − 1) and p(t − 2) time steps of each strategy at each step as needed to hold wealth con- in the past. It always buys when prices trend upward and sells stant. We then systematically vary the wealth vector W = when they trend downward. (WNT, WVI, WTF). We somewhat arbitrarily let the total wealth 8 WT = WNT + WVI + WTF = 3 × 10 . For convenience of inter- φTF(t) = log p(t − 1) − log p(t − 2). [6] 2 2 pretation, we plot the relative wealth wi (t) = Wi /WT . The results shown are averages over many long runs; to avoid tran- The trend followers’ demand is a decreasing function of price. sients, we exclude the first 252 time steps, corresponding to one Unlike the value investor, who cannot observe autocorrelations trading year. in dividends, the trend follower can exploit the autocorrelation Roughly speaking, the profitability of the dominant strategy that dividends impart to prices. Because the positive autocor- divides the wealth landscape into four distinct regions. Trend relation in market prices is, at most, ω = 0.1, we multiply the followers dominate at the bottom of the diagram, where their signal by 1/ω. wealth is small. Value investors dominate on the left side of the diagram, where their wealth is small, and noise traders domi- Noise traders. Noise traders represent nonprofessional investors nate on the right side of the diagram, where their wealth is small. who do not track the market closely. Their transactions are There is an intersection point near the center where the returns mostly for liquidity, but they are also somewhat aware of value, so of all three strategies are the same. In addition, there is a compli- they are slightly more likely to buy when the market is underval- cated region at the top of the diagram, where no single strategy ued and slightly more likely to sell when the market is overvalued dominates. The turbulent behavior in this region comes about (this is necessary so that their positions do not diverge over long because the wealth invested by trend followers is large, and the periods of time). price dynamics are unstable. We do not regard this region as The signal function of our noise traders contains the product realistic, except perhaps in rare extreme market conditions. of the value estimate V VI(t) (which is the same as for the value investors) and a stochastic component X (t), Table 2. Estimated community matrix near VI w = (NT = 0.26, VI = 0.55, TF = 0.19) φNT(t) = log2 X (t)V (t) − log2 p(t). [7] Gij NT, % VI, % TF, % The noise process X (t) is a discretized Ornstein–Uhlenbeck NT 0.46 0.40 0.36 process, which has the form − VI 8.94 −0.77 −1.89 TF 6.81 6.87 9.65 X (t) = X (t − 1) + ρ(µ − X (t − 1)) + γ. [8] −

4 of 9 | PNAS Scholl et al. https://doi.org/10.1073/pnas.2015574118 How market ecology explains market malfunction Downloaded by guest on September 26, 2021 Downloaded by guest on September 26, 2021 IID)nral itiue eun,tetm eurdt detect to of required performance case time excess ideal the distributed the returns, identically distributed In and normally 3. independent (I.I.D.) ref. and in market Farmer stationary by a made estimate the century. a with than longer are wealth that relative scales time the over in place changes taking substantial are 3B There Fig. time. in ulated trajectory Each is slow: distribution asymptotic the exceedingly the from 20%. toward deviations than evolution more The the often 3C. Furthermore, substantial, Fig. are in point shown equilibrium as efficient y, 200 after wealth strategies the of volatilities are the the because as “loosely” point say fixed We this librium. to refer loosely will We stock. the from guishable to equal all of boundary the toward are evolves simplex. lower dynamics the ecosystem the the in the where region where a diagram corner is and left the There instabilities) of to change. (due top longer complicated point the more no fixed at strategies region a the a toward the also of of evolve wealths Most diagram the 3A. the Fig. where in in shown trajectories is result wealth The runs. different vector many wealth average allow each the we to at time, corresponding return through trajectories evolves plot strategies and the reinvestment of wealth understand the To Efficiency. ecosystem. how market Market the to of dynamics Approach the investigate Slow the and Adaptation is followers trend the of it. return to average insensitive noise the the contrast, of of wealth trend in the return traders; with average the strongly the higher increases here, of investors shown implies value in the wealth not wealth is strategies, the this follower Although both of volatility. trend of function alone—higher wealth returns monotonic followers their a the as is of contrast, monotonically volatility investors decreases The followers value trend increases. investors both the to value of return average and wealth The the followers. noise trend vary the and of and size point, the hold intersection we at where constant 2B, Fig. traders in given is volatility o akteooyepan aktmalfunction market explains ecology market How al. et Scholl ecosystem with the reinvestment simulate uniformly, under wealth dynamics initial of space the ple to convergence weak. the is and equilibrium 3A, efficient the to Fig. of correspondence trajectories little deterministic bear the the of trajectories performance typical the by The we in dominated uncertainty strategies. are as statistical dynamics the trajectories, the to individual that due noise, demonstrates few 3B, a Fig. Tracking in do 3A. dynamics Fig. deterministic the in from a the far shown fact, stray of In and efficiency. impression noisy market are misleading of dynamics state a a toward gives evolution equilibrium smooth efficient the ratios to Sharpe the anything, (if low). market bit real a are a efficient as as roughly inefficient) is Sharpe ecosystem (or model These investment our hold. for that numbers and indicating reasonable funds, buy are a variation their for and 0.22 ratios and for 0.44 strategies, and three 0.31, 0.50, the are which rate) returns, risk-free the ratio, its subtracting of (without Sharpe deviation the the standard of called the terms to is in mean is the strategy of trading to ratio a way common of A might performance risk. less the equilibrium with measure strategies the of flows, favor in fund slightly shift of model sophisticated more and followers, trend h ogtm cl orahefiinyosre eematches here observed efficiency reach to scale time long The are strategies three the to returns annual the point, fixed the At and returns average the of snapshot quantitative A ogtafeigfrteaypoi elhdsrbto,w sam- we distribution, wealth asymptotic the for feeling a get To attracted are that conditions initial of region central large The 4.07% o os traders, noise for hc,t nivso,i ttsial indistin- statistically is investor, an to which, 2.05%, 2% orsodn oterwat tthe at wealth their to corresponding 42%, ∆S 9.09% eunfo ipybyn n odn the holding and buying simply from return ihasaitclsgicneof significance statistical a with S 6.76% o u n od ota,wt a with that, so hold, and buy a for h orsodn hreratios Sharpe corresponding The . w hsi oeb vrgn over averaging by done is This . o au investors, value for f 1 = n eodtefinal the record and , pn 0 fsim- of y 200 spans fcetequi- efficient 4.62% enow We s stan- for atrdtriigteapoc oefiiny eddaseries in a detail did in 4 described we tion are efficiency, that to experiments approach of the determining factor noise. equilibrium, the the efficient by as dominated the are This down Near dynamics exponential. the slows efficient. an efficiency more than to becomes slower market approach much the is because market this happens to times, form approach 2 the large the of with For law 23, performance power ref. a superior in follows its efficiency shown confirm as to Furthermore, required SD. are y 400 g odsrb h arieefcso h ouaino species of population the of effects pairwise the describe to ogy of Matrix. Community degree the well. consequently, as and, fluctuates returns times, price long in fluctu- autocorrelation very followers trend at of of even wealth 40% ates, The process. about coming dividend autocorrelation invest the the autocorrela- from eliminating followers zero thereby trend wealth, to total when the corresponding happens simplex, This the tion. across of band white center returns striking a of the is autocorrelation There profits). landscape. one-time-step excess wealth the the that significant plot across statistically we enough 3D, make (close Fig. to zero In possible to not close is reasonably it auto- an is have that should returns correlation price Efficient efficiency. market of the effect. substantial in a has uncertainty strategies, statistical three noise all the the of affects performance varying which contrast, In volatility, same. trader the roughly remains bution by 0.1 roughly of ratio Sharpe a has index S&P the of S hold and buy a approximately is deviations dard h ryadrdzones. red 3C wealth and Fig. the gray in of the shown samples as lev- to y, trophic 200 correspond the after dots where vector cycles black are The there undefined. (1, region, gray is become the ordering els In the rather red. where colored ordering zone is the dominant 3) The 2, indicate level). now trophic follower), numbers precise trend (The the investor, than legend. value the trader, in (noise indicated the order as of levels the trophic in the of strategies, ordering three increasing the to according diagram the 4. Fig. 0 = odmntaeta ttsia netit stedominant the is uncertainty statistical that demonstrate To h bec fatcreaini rc eun sa indicator an is returns price in autocorrelation of absence The ∆ ≤ S o taeywoeanaie hrertoi superior is ratio Sharpe annualized whose strategy a For .5 f sterivsmn rate reinvestment the As . uvyo h rpi eesars h elhlnsae ecolor We landscape. wealth the across levels trophic the of survey A 0 = ≤ h aeo praht h smttcwat distri- wealth asymptotic the to approach of rate the 3, a .1, 20% h omnt arxi olue necol- in used tool a is matrix community The mrvmn vrabyadhl,roughly hold, and buy a over improvement h ytmsed oto t iein time its of most spends system The . https://doi.org/10.1073/pnas.2015574118 (s f /∆ nEq. in S ) 2 otk nexample, an take To . t 2 −α aisi h range the in varies where , sec- Appendix, SI PNAS 0 ≤ | α f9 of 5 ≤ 1.

ECONOMIC SCIENCES SPECIAL FEATURE A

B

C

D

Fig. 5. How ecosystem dynamics cause market malfunctions. A gives a color map of the price volatility over the wealth landscape; the volatility is low and constant throughout the lower right part of the diagram, where the system spends most of its time, but there is a high-volatility region running across the upper left. A sample trajectory spanning 200 y, beginning at the efficient equilibrium, is shown in black. The noise caused by statistical fluctuations in performance causes large deviations from equilibrium and excursions into the high volatility region. B shows the volatility of this trajectory as a function of time, plotted against the predicted volatility (see Eq. 12). C shows the actual mispricing plotted against the predicted mispricing. D shows the wealth of the value investors and trend followers.

6 of 9 | PNAS Scholl et al. https://doi.org/10.1073/pnas.2015574118 How market ecology explains market malfunction Downloaded by guest on September 26, 2021 Downloaded by guest on September 26, 2021 g.Tersrrsn euti ht tsm aaee values, parameter assump- the some Using at indefinitely. oscillate that, that solutions is have result biol- they through population surprising in evolve Their equations famous system ogy. most the predator–prey perhaps idealized are time, an in ulations wealth the in points commu- dependence. different density at in vice illustrating found landscape, variations not be predator–prey Other can but a relationships investors. increases, have nity value now followers with followers trend trend relationship of that implying wealth investors versa, value the to is if returns there words, drop other However, In positive. equilibrium). remains the term traders noise at negative of still a than wealth now case, so the this in less in increase wealth (although an Value the strategies. from of two benefit majority other strongly the the have than who prices and investors, by random, influenced mostly not less is should strategy is This traders’ themselves. noise with surprising—the weakly be only and strategies, compete other they by that affected that indicating strongly the the not small, are in are in traders terms traders noise 2, shift the noise the a the Table of to is all in corresponding before, there row given As dominant, relations. vector are community wealth investors pairwise value the the at where matrix community the opportunity strategies. two profit other a the provides for that equi- inefficiency from away an strategies creates the librium of at reflec- any strategies efficient driving definition, on and the by equilibrium, other, is, strategies. the of ecosystem each The other all sense: with makes the that this interactions tion, of surprising mutualistic any it have of found could wealth initially benefits the we strategy in While every increase that an implies from This mutualism. indicating traders. noise more much than crowding strongly experience to followers from trend expect considerably, the should varies however, to terms Interestingly, traders diagonal 2. noise the Fig. of in strategy this size the observed as already called is diminish We what returns themselves. causing average larger, with their gets competitive that are means strategies This the that indicating 1. Table in shown the result of the center the get the compute we in simplex, point we equilibrium If the dependent. near matrix density community strongly is numerically matrix matrix (see munity community differences the finite compute using we ecosystem, ket (26). interaction mutualistic a is there then with interaction, and predator–prey predator a is there then negative, of sign strategies the accord- to classified be ing can strategies between interactions pairwise strategy is ecology market for matrix lim o akteooyepan aktmalfunction market explains ecology market How al. et Scholl wealth The time. over one of units has This of behavior the an interpreting matrix, ecosystems. for gain market useful the it also called is who quantity (3), analogous Farmer by out pointed nally j ntepplto rwhrt fspecies of rate growth population the on h ok–otraeutos hc eciehwtepop- the how describe which equations, Lotka–Volterra The compute we If dependent. density is matrix community The positive, are matrix community the of entries other the of All negative, all are matrix community the of entries diagonal The mar- toy our for model differentiable a have not do we Because Let T →∞ π i i  elcstepplto ieo pce.Tepossible The species. a of size population the replaces i eteaeaertr fstrategy of return average the be Q and t T j =1 2 3) (2, stepe;ad fboth if and, prey; the as π −19.3 j i r opttv;if competitive; are (t G ) ij f-ignltr,wieteopposite the while term, off-diagonal − fboth If . o rn olwr.Ti en htwe that means This followers. trend for 1  G 1/T ij .Tecom- The Methods). and Materials = h nlgeo h community the of analogue The . G ∂ ∂π ij w crowding i j and . G G ij G ij spstv and positive is ji nfiaca markets. financial in and i r eaie then negative, are 2,2) sorigi- As 25). (24, w i G i htis, that , (t ji ) r positive, are netdin invested −0.89 i sthe as G 3 2) (3, π ji [9] i for = is yterelation the by let never species are we webs If food simple. real this that means detritivores, and three. omnivores level trophic have lions than and two, higher one, level level level trophic trophic one has have grass zebras system, definition, idealized this by in so, is, eats, it species what a of of The level grass. population with trophic interactions the direct no on have depends lions though then grass even grass, lions, of grass, of density density eat the the by zebras similarly, affected and, strongly and is zebras, lions of interactions eat population the lions the If understanding species. for between framework conceptual Level. tant Trophic and financial Webs Food in solutions oscillating question. open of simple an existence remains that ecosystems for The strong least so system. at approximation, is this poor system a are this that equations indicate in Lotka–Volterra here dependence results density Our (3). the Lotka–Volterra ecology derived market Farmer for dependence, equations density no of tion ahmtclterms, mathematical when is, strategy strategy strategy of of ence returns given the quantity of a analogous the that define role We ecosystem. the about think they ecosystem. to the but way in integers, plays useful species not a typically provide are still levels trophic resulting The needn aibeCoefficient variable Independent independent as wealth funds’ variables mispricing the and and volatility variables, with dependent at regressions as Multivariate levels trophic 3. Table compute under- we better to dependence, are value order strategies density the In three the from the changes. stand of this lot wealths unequal, a just the substantially quite profits where but follower equilibrium, trader the trend noise in the discussion the the and investor—see from trader, prof- little noise investor a value the the follower, from either trend from its the with profit or proximity strategies cannot investor three trader value The only noise the 2.99). The are levels: there = 1.00, wealth because = follower similar is trader trend values integer (noise 2.00, to are = levels investor trophic value the equilibrium, cient rn follower Trend investor Value 1/256, value, possible smallest size. the grid to simulation the it by set determined is but which zero to entries wealth diagonal the for order and os rdr2.4 trader Noise Mispricing, follower Trend investor Value trader Noise Volatility, A ij h xsec faiaswt oecmlctddes uhas such diets, more-complicated with animals of existence The ecnas opt rpi eesfrtesrtge nan in strategies the for levels trophic compute also can We Eqs. = A P ij 10 a [0, max i 0 = i hntetohclevel trophic the then , W k R twealth at max[0, 2 and R j = 2 0 = hntednmntri eo oeas ht in that, also Note zero. is denominator the when = 9 037observations 50,397 0.79; 11 3 037observations 50,397 0.33; π j u l fteohrwatsrmi h ae In same. the remain wealths other the of all but i ed hsb ipycmaigtertrsof returns the comparing simply by this do We . (W π lo st opt rpi ees tteeffi- the At levels. trophic compute to us allow i ( W W 1 A , T 1 otoewe strategy when those to ..., ij , i ..., + 1 = etesaeo species of share the be W i W j htcnb trbtdt h pres- the to attributed be can that , h odwbpoie nimpor- an provides web food The k ... A oee,aa from away However, Appendix. SI P , T https://doi.org/10.1073/pnas.2015574118 , ii ... W i j , ob end ecnntset not can we defined, be to fspecies of A W N ij )−π N T )−π j . i (W i −1.02 −0.15 ( W −68 107 i 1.5 A 1 j a ecomputed be can , 1 ij srmvd that removed, is 0, ..., , j ... stefraction the as ntede of diet the in 0, , PNAS ..., ..., t W statistic | −107 −249 W −18 169 69 10 N f9 of 7 [10] [11] N )] )] ,

ECONOMIC SCIENCES SPECIAL FEATURE Table 4. Details of the balance sheet items used by all funds value investor’s wealth and the trend follower’s wealth have large Assets Liabilities coefficients (in absolute value), and the fit is overwhelmingly sta- tistically significant. The noise trader is also highly statistically Equity significant, but the coefficients and the t statistics are more than Cash C Capital K an order of magnitude smaller. In Fig. 5 B and C, we compare Debt a time series of the predicted volatility and predicted mispricing M Loans L against the actual values. The predictions are very good. trading securities S+ borrowed securities S−

νˆ = −68wvi + 107wtf + 2.4wnt All securities use the most recent market value in valuation. . [12] mˆ = −1.02wvi + 1.5wtf − 0.15wnt each point in the wealth landscape. For three strategies, there Note that, in both cases, the coefficient for trend followers is are 3! = 6 possible orderings of the trophic levels. We display positive, indicating that they drive instabilities, and the coeffi- the ordering of the trophic levels across the wealth landscape cient for value investors is negative, indicating their stabilizing in Fig. 4. influence. The computation of trophic levels is complicated by the fact Nonetheless, due to their effect on the population of value that, for some wealth vectors, there are cycles in the food web. investors, the net effect of the trend followers on market mal- For example, for w = (0.05, 0.15, 0.80), value investors exploit functions is not obvious. In Fig. 5D, we plot the wealth of value noise traders (who cause reversion to fundamental value), trend investors and trend followers. The strong mutualism predicted followers exploit value investors (who induce autocorrelations), by the community matrix is clearly evident from the fact that the and noise traders exploit trend followers (who generate excess wealth of trend followers and value investors rise and fall together. volatility), to complete a cycle. When this happens, Eq. 9 does However, their dynamics are quite different—there are several not converge, and the trophic levels become undefined. Cycles in precipitous drops in the value investors’ wealth, whereas the the trophic web are not unique to markets—they can also occur trend followers tend to take more-gradual losses. As predicted, in biology, for example, due to cannibalism or detritovores. the highest-volatility episodes happen when the value investors’ A comparison of Figs. 4 and 3C makes it clear that, at long wealth drops sharply while the trend followers’ wealth is high. times (after transients have died out), the system divides its time between the region in which the trophic levels are ordered as Discussion (noise trader, value investor, trend follower), as they are at the equilibrium point, and the region where there are cycles, where Our analysis here demonstrates how understanding fluctuations the trophic levels are undefined. of the wealth of the strategies in the ecosystem can help us pre- dict market malfunctions such as mispricings and endogenously How Ecosystem Dynamics Cause Market Malfunction. The wealth generated clustered volatility. The toy model that we study here dynamics of the market ecosystem help explain why the mar- is simple and highly stylized, but it illustrates how one can import ket malfunctions and illuminate the origins of excess volatility ideas from ecology to better understand financial markets. Our and mispricing, that is, deviations of prices from of this model illustrates several properties of market values. Both in real markets (2) and in the agent-based models ecosystems that we hypothesize are likely to be true in more mentioned earlier (including our model here), volatility and mis- general settings. pricings change endogenously with time—there are eras where Concepts from ecology give important insights into how devi- they are large and eras where they are small. Volatility varies ations from market efficiency occur and how they affect prices. intermittently, with periods of low volatility punctuated by bursts While the market may be close to efficiency in the sense that the of high volatility, called clustered volatility. The standard expla- excess returns to any given strategy are small, nonetheless, there nations for clustered volatility are fluctuating agent populations can be substantial deviations in the wealth of different strategies, (9, 10, 27) and leverage (28). We focus here on the first explana- that can cause excess volatility and market instability. tion; while we also observe that clustered volatility increases with Market ecology is a complement rather than a substitute for increasing leverage, we have not investigated this, in detail, here. the theory of market efficiency. There are circumstances, such Fig. 5A presents the variation of the volatility across the as pricing options, where market efficiency is a useful hypothe- wealth landscape. The landscape can roughly be divided into two sis. Market ecology, in contrast, provides insight into how and regions. On the lower right, there is a flat “low-volatility plain” why markets deviate from efficiency, and what the consequences occupying most of the landscape. On the upper left, there is a of this are. It can be used to explain the time dependence high-volatility region, with a sharp boundary between the two. in the returns of trading strategies, and, in some cases, it can As we will now show, excursions into the high-volatility region be used to explain market malfunctions. One of our main cause clustered volatility. A similar story holds for mispricing. Fig. 5A shows a sample trajectory that begins at the efficient Table 5. List of the model parameters and their values equilibrium and spans 200 y. This sample is representative of trajectories that fluctuate around the equilibrium point indefi- Parameter Value Description nitely. We choose a 200-y time span to show the scale at which r 1% annual Risk-free rate spikes in volatility and mispricing occur. The statistical fluctu- g 1% annual Dividend growth rate ations in the performance of the three strategies act as noise, k 2% annual Cost of equity causing large excursions away from equilibrium. The trajectory σ 10% annual dividend growth Volatility mostly remains on the volatility plain, but there are several epochs ω 0.1 Dividend autocorrelation parameter where it ventures into the high-volatility region, causing bursts of τ 1 d Dividend autocorrelation lag high volatility. f 1 Reinvestment rate q The wealth dynamics have strong explanatory power for both ρ 1 − 6·252 1 Noise trader rate mispricing and volatility. This is illustrated in Table 3, where we 2 σNT 20% annual Noise trader volatility perform regressions of the strategies’ wealth against volatility λ¯ , λ¯ , λ¯ 1, 8, 1 Leverage limit using daily values for the time series shown in Fig. 5A. For volatil- NT VI TF c , c , c 5, 10, 4 Signal scale ity, R2 = 0.79, and, for mispricing, R2 = 0.33. In both cases, the NT VI TF

8 of 9 | PNAS Scholl et al. https://doi.org/10.1073/pnas.2015574118 How market ecology explains market malfunction Downloaded by guest on September 26, 2021 Downloaded by guest on September 26, 2021 5 .LBrn Aetbsdcmuainlfiac”in finance” computational “Agent-based LeBaron, B. 15. 4 .F aa .R rnh iiedyed n xetdsokreturns. stock expected and yields Dividend French, R. K. Fama, F. E. 14. investors short-memory and long- interacting from regularities a “Empirical LeBaron, behaviour. of B. speculative model 13. of multi-agent dynamics stochastic The a Chiarella, in C. criticality and 12. Scaling Marchesi, M. randomness. Lux, to T. route rational 11. A Hommes, H. C. Brock, A. W. 10. o akteooyepan aktmalfunction market explains ecology market How al. et Scholl htcudb sdt nepe h eairo elworld real of behavior the interpret to methods used and concepts be examples. provides could here stabi- that analysis to a through Our counterparty construct with tend or data to identifiers. way against is innovation validated strategies, empirically open-ended follow-up is strategy that Another of model an markets? of destabilize space in process or larger lize evolve the a strategies Does from explore time. new deviations to let is large to follow-up expect ous should even we that, results term, indicate Our they long appreciated. efficiency. widely and the not dramatically, in is sta- this straightforward it a demonstrate but from already obvious analysis, As be slow. tistical should exceedingly this and out, uncertain pointed highly mortgage-backed is efficiency as such assets of construct types the to securities. example, new possible for innovations, it of of make introduction effect mar- and ecological those failure, when the as for of into models such danger insight in Ideas detail, valuable are regulate. provide in kets they could track, markets here to the presented them of allow ecology can the which participants, market researchers most for without difficult Unfortu- are animal, involved. obtain. data another to animals such of markets, ate types for animal which nately, the in an about ecosystem information that biological any a observe study without can to ecosystem trying market one who like a is know study requires data to to study such Trying order a whom. in such with transactions because traded on is identifiers This 30). counterparty (29, ecology mutualistic market have insight strategies three provide efficient another. all one which the to that at relationships web, find Surprisingly, we strategies. com- trophic of equilibrium, the the interactions compute the to and into how matrix demonstrate to munity is here innovations .W .Atu,J .Hlad .D eao,R .Ple,P alr Astpiigunder pricing “Asset Tayler, P. Palmer, G. R. LeBaron, D. B. Holland, H. J. markets. Arthur, efficient B. informationally W. of impossibility 9. the On Stiglitz, E. J. value-tracking Grossman, of Dynamics J. MacKay, S. S. R. 8. Battey, S. H. Bashe, L. K. Gunton, trading. R. computer Beale, of N. future the 7. on perspective ecological An Skouras, S. Farmer, D. evolutionary J. an 6. from efficiency Schenk-Hopp Market R. K. Hens, hypothesis: T. market 5. adaptive The Lo, evolution. A. and 4. ecology force, prices? Market Farmer, stock D. moves J. What 3. Summers, Shiller, J. H. R. L. 2. Poterba, M. J. Cutler, M. D. 1. hr r aypsil xesost hswr.A obvi- An work. this to extensions possible many are There to approach the that is results striking most our of One all of sheets balance the to access have potentially Regulators of studies empirical of examples few a only far, so are, There .Tsaso,K .Jd,Es Esve,e.1 06,vl ,p.1187–1233. pp. 2, vol. 2006), 1, ed. (Elsevier, Eds. Judd, L. K. Tesfatsion, L. Economics, (1988). 3–25 22, IE,20) o.5 p 442–455. pp. 5, vol. in 2001), market” (IEEE, stock agent-based an in (1992). market. financial (1997). II System in market” Complex stock artificial an in expectations endogenous Rev. Econ. Am. markets. financial in Finance Quant. 2009). Holland, (North perspective. (2002). Manag. –2(1989). 4–12 15, aktVolatility Market .PrfloManag. Portfolio J. 9–0 (1980). 393–408 70, 2–4 (2013). 325–346 13, Nature .B rhr .N ulu,D ae d.(R,19) p 15–44. pp. 1997), (CRC, Eds. Lane, D. Durlauf, N. S. Arthur, B. W. , .Sc pnSci. Open Soc. R. 9–0 (1999). 498–500 397, e, ´ adoko iaca akt yaisadEvolution and Dynamics : Markets Financial of Handbook MTPes 1989). Press, (MIT 52 (2004). 15–29 30, EETascin nEouinr Computation Evolutionary on Transactions IEEE –2(2019). 1–22 11, n.Cr.Change Corp. Ind. n.Oe.Res. Oper. Ann. adoko Computational of Handbook h cnm sa Evolving an as Economy The Econometrica .Fnn.Econ. Financ. J. 101–123 37, 895–953 11, .Portfolio J. 1059 65, 9 .Krlno .S ye .Smd,T uu,Tefls rs:Hg-rqec trading High-frequency clustered crash: and flash The tails Tuzun, T. fat Samadi, M. causes Kyle, S. Leverage A. Kirilenko, Geanakoplos, A. J. 29. Farmer, dynam- D. pricing J. asset Thurner, S. in 28. volatility stochastic Structural Westerhoff, F. Franke, R. 27. interactions. mutualistic in outcomes Conditional Bronstein, L. J. 26. stable? be system Novak complex large M. a game. Will 25. May, reality M. The R. Lloyd, 24. S. Farmer, D. J. Cherkashin, D. 23. Bouchaud P. J.-P. 22. trading. Asness momentum S. toward C. bias a 21. and dynamics Wealth LeBaron, B. 20. 0 .Msito .Mrta .Pio .N atga ogtr clg fivsosi a in investors of ecology Long-term Mantegna, N. R. Piilo, J. Marotta, L. Musciotto, F. 30. strategies. trading common of dynamics price simple The a Joshi, in S. chaos to Farmer, routes D. and J. common beliefs 19. Heterogeneous on Hommes, policy H. dividend C. and Brock, A. W. dividend of 18. effects profit. The of Scholes, rate M. required Black, The F. analysis: 17. equipment Capital Shapiro, E. Gordon, J. M. 16. 00rsac n noainpormudrGatAreet952215. Agreement Horizon (TAILOR) Grant Union under Reasoning program European innovation and by and research Optimisation funded 2020 Learning, project Trust- Economic by a Integrating the supported (https://tailor-network.eu/), - partially We by was Baillie-Gifford, AI research funded work. This Awards, worthy program, Council. this Faculty Research Macroeconomics Social of AI and Rebuilding Morgan development the J.P. the and by during Svosve funding Cephas discussion acknowledge and Kleinnijenhuis, enlightening Alissa for Cont, Rama Wooldridge, Michael ACKNOWLEDGMENTS. (https://github.com/INET-Complexity/market-ecology). GitHub in deposited Availability. Data Library. default Software. the lists 5 Table insolvent. losses. are to funds due more condition parameters. or happen solvency one changes can the when assets meet This ends funds lation risky capital. all equity of that of require proportion We amount the the when than constraint faster leverage its late odns on,admri stesm sters-rert rmholding from rate risk-free the as same the bond. is the margin and 10. loans, and holdings, 1 cash common between borrowing indicate managers ratios by fund leverage institutional assets, use US risky of can filings additional regulatory funds Public purchase The to sets cash. fund funds of the borrowed form time, the later can in a fund rowings, amount at short-selling lender margin the the a that to aside guarantee securities to borrowed order the In return position. short a ate equity with endowed S are Funds 4. Table securities in presented is capital that sheet ance Sheets. Balance and Accounting Methods and Materials ‡ † h lcrncDt ahrn,Aayi,adRtivlsse EGR,publicly (EDGAR), system Retrieval and Commission. Analysis, Exchange and Securities Gathering, https://www.sec.gov/edgar.shtml , at Data available Electronic The noe-orelbayfraetbsdmdln hc spbil cesbeat accessible publicly is which modeling agent-based for https://github.com/INET-Complexity/ESL. library open-source An < elhi acltdas calculated is Wealth na lcrncmarket. electronic an in volatility. contest. model and (2012). Estimation ics: 1994). (6 214–217 matrix? community the (2016). is What bations: (2009). 1091–1105 2018). June 1 (Accessed https://arxiv.org/abs/1711.04717 (2017). [Preprint] (2013). Lett. Organ. Behav. nnilmarket. financial model. pricing asset returns. and prices stock Sci. Manag. ,i a orwdti mutfo te aktpriiat,t cre- to participants, market other from amount this borrowed has it 0, 12 (2012). 21–28 9, ‡ K = S h iuaini hspprbid nteEooi Simulation Economic the on builds paper this in simulation The When . un.Financ. Quant. W hrceiigseisitrcin oudrtn rs pertur- press understand to interactions species Characterizing al., et 0–1 (1956). 102–110 3, 0,i h omo cash of form the in (0), au n oetmeverywhere. momentum and Value al., et 4–7 (2002). 149–171 49, agaeCommun. Palgrave h oe n oet u h xeiet aebeen have experiments the run to code and model The lc a ih:Piei ihnafco fvle arXiv value. of 2 factor a within is Price right: was Black al., et S > .Eo.Dnm Contr. Dynam. Econ. J. ,tefn od hsaon fscrte,adwhen and securities, of amount this holds fund the 0, .Finance J. etakKasSchenk-Hopp Klaus thank We .Fnn.Econ. Financ. J. 9–0 (2012). 695–707 12, M W qa otecretmre au ftebor- the of value market current the to equal (t ) = h ud normdlueasyie bal- stylized a use model our in funds The 6–9 (2017). 967–998 72, C 2(2018). 92 4, (t † ) h neetrt htapist cash to applies that rate interest The –2(1974). 1–22 1, + https://doi.org/10.1073/pnas.2015574118 .Eo.Dnm Contr. Dynam. Econ. J. nu e.Eo.Eo.Syst. Evol. Ecol. Rev. Annu. C S 2517 (1998). 1235–1274 22, (t ndlasadsae ftrading of shares and dollars in )p(t Nature ) − L(t enn using meaning leverage, .Eo.Dnm Contr. Dynam. Econ. J. 1–1 (1972). 413–414 238, ). .Finance J. udcnol vio- only can fund A W ,Rbr MacKay, Robert e, ´ (t rnsEo.Evol. Ecol. Trends ) > PNAS 1193–1211 36, .Tesimu- The 0. 929–985 68, iac Res. Finance 409–432 47, | .Econ. J. f9 of 9 33, L. 9, J.

ECONOMIC SCIENCES SPECIAL FEATURE