Complex evolutionary systems in behavioural finance Application: financial instruments and market stability Florian Wagener CeNDEF University of Amsterdam Dynamics in Games and Economics 1 / 16 Overview Brock, Hommes & Wagener (2009) Extension of the Brock & Hommes (1998) heterogeneous agents asset pricing model by adding hedging instruments (Arrow securities) Main Question Is it true that giving traders more trading options will improve the market outcome? 2 / 16 BH asset pricing model with Arrow securities Stochastics In the next period (“tomorrow”) one of S possible states of the world occurs with probability αs Investment possibilities • Bonds, totally elastically supplied, price 1, return R 0 s • Risky asset, total supply ζ 0 , price pt0 , return pt+1 + yt+1 • n Arrow securities, total supply 0, price ptj , j = 1, .., n, return 1 δjs = 0 if s = j, otherwise 3 / 16 Profits 0 Trader of “type h” expects price pht+1 of the risky asset; 1 0 excess return of portfolio (zt , zt , · · · , ztn ) is s 0 s πht+1 = (pht+1 + yt+1 − Rpt0 )zt0 + X (δjs − Rptj )ztj j Demands and prices written as vectors zt = (zt0 , z̃t ) = (zt0 , zt1 , · · · , ztn ) pt = (pt0 , p̃t ) = (pt0 , pt1 , · · · , ptn ) Expected excess returns in state s Bsht = 0 s pht+1 + yt+1 − Rpt0 δ s − R p̃t ! , Bht = X Bsht αs s Excess return of portfolio zt expected by trader type h πht+1 = hBht , zt i 4 / 16 Demands Traders maximise expected risk-corrected profits Uht (z) = Eht πht+1 − a 1 Var πht+1 = hBht , zi − hz, Vn zi, 2 2 where • a: coefficient of risk aversion • Vn : (dividend) risk structure s s Vn = a Cov(yt+1 , δ s ) = a Cov(yt+1 , δ1s , · · · , δns ) Key assumption: risk structure is common knowledge Demands zht = Vn−1 Bht 5 / 16 Benchmark: homogeneous market Single rational trader • Outside supply of risky assets: ζ 0 • No outside supply of Arrow securities • Total outside supply of assets ζ = (ζ 0 , 0, · · · , 0) Market equilibrium at steady state prices p∗ ζ = Vn−1 B implies B= s (1 − R)p∗0 + Eyt+1 α − R p̃∗ ! = Vn ζ, determines homogeneous benchmark prices p∗ = (p∗0 , p ~∗ ) 6 / 16 Heterogeneous agents market equilibrium • nht : market fraction of traders of type h • xt = (xt0 , x̃t ) = pt − p∗ : deviation from benchmark price • Expectations of type h on the price deviation of the risky asset 0 0 − p∗0 = fht (xt−1 , · · · ) = pht+1 xht+1 Demands in deviation from the homogeneous demands ζ zht (xt ) = ζ + Vn−1 Market clearing ζ = xt0 = 1X nht fht , R fht − Rxt0 − R x̃t P h ! nht zht (xt ) implies x̃t = 0. 7 / 16 Reinforcement learning Agents flock to strategies having good fitness scores • Uht−1 is the fitness of type h after trading round t − 1 • The intensity of choice β is inversely related to noise when observing Uht−1 • Discrete choice model nht = X eβUht−1 Zt , X nht = 1 h We take as fitness measure the average realised risk-adjusted profits 1 hzht , Vn zht i 2 0 1 = − e0 , Vn−1 e0 (xt+1 − fht )2 + Ct 2 Uht+1 = hBt , zht i − 8 / 16 Risk measure σn2 : measure of risk in presence of n Arrow securities 1 = a e0 , Vn−1 e0 2 σn 2 2 σ02 > σ12 > · · · > σS−2 > σS−1 =0 Theorem: 0 zht = ζ 0 ! 1 2 + (fht − Rxt0 ) σn w̃n Less (percieved) risk means more demand ⇒ leveraging Price dynamics X Rxt0 = − e X β 2 σn − e β 2 σn 0 (xt−1 −fht−2 )2 fht 0 (xt−1 −fht−2 )2 9 / 16 Example Two-type model: near-fundamentalists versus trend chasers f1t = 1, f2t = xt−1 + g(xt−1 − xt−2 ) 1 x SN2 12 SN1 Hopf 0 0 5 10 15 10 / 16 Welfare Welfare: population average of realised utility X Wt = nht−1 Uht With prediction errors εht = xt0 − fht−1 of type h, welfare can be written as Wt = 1 2 2 0 2 0 a σ0 (ζ ) + (xt0 − Rxt−1 )aζ 0 2 | {z } | {z } risk premium − n0,t−1 C | {z } irrationality bias perfect foresight costs − 1 Var εht 2a σ 2 | {z n } variance of prediction errors 11 / 16 Example II Average welfare W̄ = T −1 X Wt Hopf 1 SN1 SN2 welfare x -0.1 12 SN1 -0.2 Hopf 0 0 5 10 15 ΒΣn 2 SN2 -0.3 -0.4 -0.5 0 5 10 15 20 25 30 ΒΣn 2 Theorem: limσn →0 W̄ = −∞ 12 / 16 Adding a perfect foresight type • Demand of perfect foresight type has to clear the market at the previously predicted price • First round prediction should be such that local rational bubbles are avoided • Perfect foresight induces an additional cost C Market clearing equation: X Rxt = n0t xt+1 + nht fh (xt−1 , · · · ) h Fitness (no prediction error) U0t = −C Typically C 1 13 / 16 Singular perturbation theory Introduce ε = e−βC , rewrite dynamics X X εxt+1 = Rxt e−βUht − e−βUht fh (xt−1 , · · · , xt−L ) h h Here 0 < ε 1: singular perturbation (increases order) Reduction to centre-stable manifold at stability loss • rules out rational bubbles • “correct” limit behaviour as C → ∞ 14 / 16 Effects of perfect foresight traders If all boundedly rational types in the market are biased in the fundamental equilibrium, then, if Arrow securities are added to the market • all boundedly rational types are driven out • the price is driven to the fundamental • welfare stabilises at forecasting costs If some types are nonbiased in the fundamental equilibrium (trend chasers, naive expectations), then • nonbiased types are not driven out • prices may remain volatile • welfare may still decrease towards minus infinity 15 / 16 Summary In markets with heterogeneous, adaptively learning traders, adding hedging instruments • decreases risk • amplifies effect of forecast errors • destabilises rather than stabilises the market • increases volatility • decreases welfare Adding perfect foresight traders can counteract the last three conclusions in some cases But: No, adding trading options is not automatically a good idea 16 / 16