Advertising and Market Share Dynamics. Minjung Park University of Minnesota

Transcription

1 Advertising and Market Share Dynamics Minjung Park University of Minnesota

2 Introduction Sutton (1991) s Model of Endogenous Sunk Costs - As market size increases, (1) Market structure in exogenous sunk cost industries gets fragmented (2) Market structure in endogenous sunk cost (ESC) industries remains unfragmented. - Crucial elements for ESC (1) Quality improvement falls on fixed costs (2) A significant fraction of responsive consumers 2

3 Introduction US mutual fund industry (1) Load funds: sold through brokers; consumers not very responsive to mass advertising (2) No-load funds: sold directly to investors; consumers responsive to mass advertising Significant market growth from 1985 to 2004 Research Question: different consumer responsiveness to ads in two segments how evolution of advertising choices and market structure differ between the two?

4 Introduction Roadmap (1) Show data patterns : evolution of ad spending and concentration ratio in load vs. no-load segments (2) Empirical model of advertising dynamics : estimation using BBL (2007) () Results & Discussion (4) Conclusion 4

5 Year Market Structure Load Segment No-Load Segment AUM ($b) C C 5 AUM ($b) C C

6 Load Segment (No. of firms) No-Load Segment (No. of firms) Year 50% of market: Dominant firms 50% of market: Dominant firms

7 Ad-sales ratio (sales = assets expense ratio) Top 5 Others Load No-load

8 Two Tired Market Load Segment (in $000) Av. Spending Av. Spending by Big 5 by (Big 5) c No-Load Segment (in $000) Av. Spending Av. Spending by Big 5 by (Big 5) c

9 Are Big Firms Also Big Ad Spenders? Load Segment No-Load Segment

10 Model of Advertising Competition Firms j = 1,2, N, each producing one good Discrete time with infinite horizon t = 0,1, Firms compete in advertising and price Focus on large players; Persuasive advertising Model Components (1) State space and timing (2) State transition and goodwill accumulation () Demand and profits (4) Markov Perfect Equilibrium 10

11 State space s t S -( N + 2) 1 vector - for each firm j, τ j (type: load (1) or no-load (0)), GW jt-1 (goodwill stock), μ jt (quality) - quality of fringe (μ Ft ) common to all firms - market size M t common to all firms Each period consists of two stages - stage 1: after observing state vector s t and an iid private shock ν A jt, each firm chooses A jt - stage 2: after observing updated GW and an iid private shock ν P jt, each firm sets P jt 11

12 State transition - τ is fixed over time - μ jt and M t stochastically evolve over time indep. of actions; F μ (μ jt+1 μ jt ), F M (M t+1 M t ) - μ Ft determined by fringe firms actions, but modeled ad hoc for now; F μf (μ Ft+1 μ Ft, μ t,m t ) -GW jt deterministically evolves as a finite distributed lag of advertising q GW jt = λ k A : carry-over effects of advertising jt k k= 0 12

13 Demand: discrete choice U ijt = X jt β-αp jt + ξ jt + ε ijt -X jt β= (β 1 + β 2 τ j ) ln(gw jt ) + Past Perf jt + stuff jt - P jt : P 1jt (load), P 2jt (expense ratio) - ε ijt ~ type 1 extreme value distribution Per-period profits π jt = (P 2jt -MC jt )Q jt -FC jt - (1+ν A jt)a jt -c 2 (τ j )1(A jt >0) -Q jt = M t market share jt -MC jt = c 1 (τ j ) + ν P jt, ν P ~ N(0,σ 2 νp) -ν A ~ N(μ ν A,σ 2 νa) 1

14 Markov strategy for firm j σ j : S ν j (A j, P j ), ν j = (ν A j, ν P j ) Strategy profile σ = (σ 1,..., σ N ) σ: S ν 1 ν N (A, P) F(s' σ,s): transition probability of s given σ The ex-ante value function of firm j in state s when firms follow strategy profile σ Vj(s σ)=eν[πj(σ(s,ν),s,νj )+δ Vj(s' σ)df(s' σ(s,ν),s) s] Strategy profile σ is an MPE if for all j, s, and σ j ' V j (s σ) V j (s σ j ',σ -j ) 14

15 CRSP Data - Info on asset size, management company, investment objective, return, etc. of each fund - Categorize firms into either load or no-load Ad $ Summary - Covers major media (TV, magazines, ) - Reports annual ad spending for each fund company Fed Flow of Funds - Market Size = Financial Assets held by domestic financial sector 15

16 Estimation - BBL (2007) Computationally light two-step estimator If we use Nested Fixed Point Algorithm (Rust, 1987), need to solve the dynamic programming repeatedly for each trial value of the parameter vector. But solving the dynamic game even once is computationally burdensome! Two-step estimator (BBL, 2007; AM, 2007): - Step 1: Recover what agents do - Step 2: Find parameter values that rationalize the observed behavior using forward simulation 16

17 Estimation - BBL (2007) Model 0 largest firms choices The rest is aggregated into fringe ; fringe not an active player First stage (1) Demand - logit; do Berry inversion; use BLP-type IV - following A&R (2005), use ln(n) as a regressor - back out μ jt and μ Ft 17

18 OLS Logit IV Logit (IV for P) β (22.974) (18.50) β (-11.2) (-8.48) λ (7.040) (5.696) P (-0.587) 0.11 (0.874) P ( ) (-8.490) ln(n) (-1.415) (1.152) Age (5.920) (2.92) 1(Age<2) (-7.55) (-6.217) Perf t (1.445) (2.95) 2 Perf t (-6.576) (0.121) Perf t (10.61) (0.167) 2 Perf t (-4.92) (1.497) No. obs Adjusted R Implied own price elasticity OLS Logit P1 P2 P1 P IV Logit

19 : Consumers more responsive to ads by no-load firms than to ads by load firms; Other coefficients plausible (2) State transition - Estimate F μ (μ t+1 μ t ) as AR(1) process - Estimate F μf (μ Ft+1 μ Ft, μ t, M t ) - Estimate first-differenced market size as AR(1) for stationarity - GW depreciation parameter λ from demand estimation 19

20 μ jt+1 = μ jt + N(0, 0.47) (0.090) (0.015) R 2 = 0.90 μ Ft+1 = μ Ft Σμ t (M t /1000) + N(0, 0.098) R 2 = 0.96 ΔM t+1 = ΔM t + N(0, 281.) (70.156) (0.097) R 2 = 0.64 () Product market competition - static Bertrand-Nash - back out MC and σ 2 νp MC j = τ j + N(0, 0.00) - load firms hire wholesalers for sales pitch and give incentive pay, hence higher MC 20

21 (4) Policy function - π j has Inc. Diff. in (A j, -ν A j) σ A (s,ν A j) = F -1 (1-Φ(ν A j /σ ν A ) s) where F(A j s) = Pr(σ A (s,ν A j) A j s) is estimated as truncated normal (tobit) A jt = A jt* if A jt* > 0 = 0 if A jt* 0 where A jt* = f(gw t-1, M t, μ t, μ Ft ) + N(0, σ 2 ) - Separately estimate for L and NL firms 21

22 Second Stage - Remaining parameters (μ ν A,σ 2 νa, c 2 (τ j )are estimated using forward simulation - Estimates: μ νa = 0.7; σ 2 νa = 2.96; c 2 (τ j ) = τ j Crucial to get policy functions right (extrapolation to states unobserved in data) Compare actual market share dynamics to simulated paths to check the performance of estimated policy function 22

23 2 Market Share Dynamics - Data No-Load Firm1 Firm2 Firm Firm4 Firm5 Firm6 Firm7 Firm8 Firm9 Firm10 Firm11 Load Firm12 Firm1 Firm14 Firm15 Firm16 Firm17 Firm18 Firm19 Firm20 Firm21 Firm22 Firm2 Firm24 Firm25 Firm26

24 Market Share Dynamics Prediction No-Load Firm1 Firm2 Firm Firm4 Firm5 Firm6 Firm7 Firm8 Firm9 Firm10 Firm Load Firm12 Firm1 Firm14 Firm15 Firm16 Firm17 Firm18 Firm19 Firm20 Firm21 Firm22 Firm2 24

25 Counterfactual: If no firm advertises No-Load Firm1 Firm2 Firm Firm4 Firm5 Firm6 Firm7 Firm8 Firm9 Firm10 Firm Load Firm12 Firm1 Firm14 Firm15 Firm16 Firm17 Firm18 Firm19 Firm20 Firm21 Firm22 Firm2 Firm24 25

26 Model does a reasonable job in explaining difference between L and NL segments But underestimates differences among firms within segment need richer firm heterogeneity Advertising is an important strategic tool for keeping a concentrated market structure in a growing market 26

27 Conclusion Empirically show the role of advertising in keeping market structure concentrated Fill in the gap in empirical work on dynamic advertising Future Work - Deal with endogeneity of advertising in demand estimation - Add random coefficients in demand estimation - Add demand dynamics (due to switching costs) 27