Banque nationale de Belgique SA · EURONEXT:BNB
Banque nationale de Belgique SA — Investor Presentation (June 2013)
56 slides from Banque nationale de Belgique SA's investor presentation, Jun 5, 2013
Slide 1: A Search and Learning Model of Export Dynamics
A Search and Learning Model of Export Dynamics Jonathan Eaton,a,b Marcela Eslava,c David Jenkinsa , C.J. Krizan,d James Tybouta,b a Penn State, b NBER, c U. de los Andes, d Census Bureau (CES) June 5, 2013 Eaton et al. (a Penn State, b NBER, c U. de los Andes,Search d Census andBureau Export(CES)) D
Slide 2: 2 sets of relevant issues
2 sets of relevant issues Aggregate/industry level export dynamics What makes export responses to exchange rates vary across countries and time periods? Why are export responses to trade liberalization unpredicable? What are the underlying causes of export booms? Eaton et al. (a Penn State, b NBER,
Slide 3: 2 sets of relevant issues
2 sets of relevant issues Aggregate/industry level export dynamics What makes export responses to exchange rates vary across countries and time periods? Why are export responses to trade liberalization unpredicable? What are the underlying causes of export booms? Trade frictions at the …rm level Wha
Slide 4: 2 sets of relevant issues
2 sets of relevant issues Aggregate/industry level export dynamics What makes export responses to exchange rates vary across countries and time periods? Why are export responses to trade liberalization unpredicable? What are the underlying causes of export booms? Trade frictions at the …rm level Wha
Slide 5: The exercises
The exercises Chararcterize buyer-seller relationships in decade’s worth of data on individual merchandise shipments from Colombia to the United States Develop a (partial equilibrium) dynamic search and learning model that explains patterns found in shipments. Fit the model to the data, and quantify
Slide 6: Related literature
Related literature Heterogeneity and trade Melitz (2003), etc. Beachhead exporting costs: Theory : Dixit (1989), Baldwin and Krugman (1989), Impullitti, Irarrazabal, and Opromolla (2012) Quantitative: Roberts and Tybout (1997), Bernard and Jensen (2004) Das, Roberts, and Tybout (2008) Marketing cost
Slide 7: Stylized facts
Stylized facts Evidence from Colombian customs data Population of (legal) Colombian export transactions over the course of a decade (1996-2005). Each transaction has a date, value, product code, …rm ID, and destination country. See also: Besedes (2006); Bernard et al (2007); Blum et al (2009); Albor
Slide 8: Exporters by durability
Exporters by durability As a fraction of total exporters, …rms that enter a market and immediately exit are important. Eaton et al. (a Penn State, b NBER, c U. de los Andes,Search d Census andBureau Export(CES)) Dynamics 6/5/13 6 / 54
Slide 9: Exporters by durability
Exporters by durability But as a fraction of total export revenue, brand new exporters don’t account for much. Eaton et al. (a Penn State, b NBER, c U. de los Andes,Search d Census andBureau Export(CES)) Dynamics 6/5/13 7 / 54
Slide 10: Cohort maturation
Cohort maturation The …rms that survive their …rst year grow exceptionally rapidly (see also Ruhl and Willis, 2008). Eaton et al. (a Penn State, b NBER, c U. de los Andes,Search d Census andBureau Export(CES)) Dynamics 6/5/13 8 / 54
Slide 11: Cohort maturation
Cohort maturation Hence young cohorts typically gain market share despite rapid attrition. Post-1996 entrants account for about half of cumulative export expansion by 2005. Eaton et al. (a Penn State, b NBER, c U. de los Andes,Search d Census andBureau Export(CES)) Dynamics 6/5/13 9 / 54
Slide 12: Cohort maturation
Cohort maturation Most new matches fail within a year, but Chances of survival are higher for matches with large initial sales Survival rates improve and converge for all matches after the …rst year. To sustain or increase exports, …rms must continually replenish their foreign clientele. Eaton et al
Slide 13: Matches that start small tend to stay small
Matches that start small tend to stay small. After a match’s …rst year, there is no systematic tendency for its annual sales to grow. Eaton et al. (a Penn State, b NBER, c U. de los Andes,Search d Census andBureau Export(CES)) Dynamics 6/5/13 11 / 54
Slide 14: A seriously Pareto client distribution
A seriously Pareto client distribution Most …rms have a single buyer, but the distribution of client counts across exporters is fat-tailed. Eaton et al. (a Penn State, b NBER, c U. de los Andes,Search d Census andBureau Export(CES)) Dynamics 6/5/13 12 / 54
Slide 15: Year-to-year transitions in numbers of clients
Year-to-year transitions in numbers of clients Table 3: Transition Probabilities, Number of Clients tnt+1 exit texit 1 2 3 4 5 6-10 11+ enter 0.000 0.000 0.947 0.044 0.007 0.002 0.001 0.001 0.000 texit 0.000 . 0.896 0.086 0.014 0.004 . . 0.000 1 0.533 0.081 0.332 0.043 0.008 0.002 0.001 . . 2 0.180
Slide 16: Key model features
Key model features Firms engage in costly search to meet potential buyers at home and (possibly) abroad. Firms new to the foreign market don’t know what fraction of buyers there will be willing to do business with them. As they encounter potential buyers, …rms gradually learn the scope of the market
Slide 17: Three model components
Three model components 1 A Seller-Buyer Relationship 2 Learning About Product Appeal from Encounters with Potential Buyers 3 Searching for Potential Buyers Eaton et al. (a Penn State, b NBER, c U. de los Andes,Search d Census andBureau Export(CES)) Dynamics 6/5/13 15 / 54
Slide 18: Why continuous time?
Why continuous time? Two types of discrete events occur at random intervals, sometimes with high frequency Sellers meet buyers Once business relationships are established, orders are placed With a continuous time formulation, we can: allow for an arbitrarily large number of events during any discret
Slide 19: Relationship dynamics: profits from a shipment
1. Relationship dynamics pro…ts from a shipment De…ne exogenous state variables: ϕj productivity of seller j (time invariant) xtm size of market m 2 fh, f g (Ehrenfest jump process) Details m idiosyncratic shock to operating pro…ts from shipment to buyer i by yijt seller j in market m (Ehrenfest jum
Slide 20: Relationship dynamics: value of a business relationship
1. Relationship dynamics value of a business relationship In active business relationships, buyers place orders with exogenous hazard λb . Details After each order, sellers must pay …xed cost F to keep a business relationship active. Value to a type-ϕ seller of a relationship in state fx, y g: e ϕ (
Slide 21: Relationship dynamics: expected value of a new relationship
1. Relationship dynamics expected value of a new relationship Sellers don’t know what y value their next business relationship will begin from. Let Pr(y s ) be the probability of initial shock y s , determined by the ergodic distribution of y . Expected value of a successful new encounter: e ϕ (x )
Slide 22: Learning about product appeal
2. Learning about product appeal the "true" scope of the market Let θ m j 2 [0, 1] be the fraction of potential buyers in market m who are interested in seller j’s product. m Assume θ j 0 s are time-invariant, mutually independent draws from a beta distribution: Γ(α + β) r (θ jα, β) = ( θ ) α 1 (1 θ
Slide 23: Searching for buyers: the cost of search
3. Searching for buyers the cost of search Seller continuously chooses the hazard s with which she encounters a potential buyer at a ‡ow cost c (s, a) Maintain web site Pay to be near top of web search listings Attend trade fairs Research foreign buyers Send sales reps. to foreign markets Maintain f
Slide 24: Searching for buyers: the value of search abroad
3. Searching for buyers the value of search abroad De…ne the value of continued search for a type-ϕ …rm with a successes in n meetings, market state x: V ϕ (a, n, x ) The …rst-order for optimal search abroad is: Details e ϕ (x ) + V ϕ (a + 1, n + 1, x )) cs (s , a) = θ a,n (π +(1 θ a,n )V ϕ (a, n +
Slide 25: Searching for buyers: search in the domestic market
3. Searching for buyers the value of search in the domestic market As n increases, θ a,n converges to the true θ. There is no more learning, and the reward to search depends on a and n only through network e¤ects. We assume this characterizes the domestic market. If network e¤ects are ignored, the …
Slide 26: Estimation: the exogenous state variables
Estimation The exogenous state variables Notation refresher: if z follows an Ehrenfest di¤usion process: e 2 I + and ∆ 2 R + determine support: z 2 f e∆, (e 1)∆, .., 0, .., (e 1)∆, e∆g The process jumps with hazard λz : λz t F [t ] = 1 e As the grid becomes …ner, this type of random variable asympto
Slide 27: Estimation: the exogenous state variables
Estimation The exogenous state variables If z observed at regular intervals, can estimate µ and σ by regressing z on lagged z For x f , x h , obtain maximum likelihood estimates of µ and σ using logged and de-meaned time series on total real consumption of manufactured goods in each country. Recover
Slide 28: Estimation: the exogenous state variables
Estimation The exogenous state variables Market-wide Shock Processes (x f , x h ) Orstein-Uhlenbeck Parameters Colombia United States µ Mean Reversion 0.171 0.174 σ Dispersion 0.003 0.058 Ehrenfest Process Parameters λ Jump Hazard 1.200 1.215 ∆ Jump Size 0.003 0.053 grid points 15 15 Eaton et al. (a
Slide 29: Estimation: remaining parameters
Estimation remaining parameters Unidenti…ed preference parameters taken from literature: ρ = 0.05, σ=5 Remaining parameters identi…ed using indirect inference Λ = Πh , Πf . , δ,F , α, β, σ ϕ , λy , λb , γ, κ 0 , κ 1 Eaton et al. (a Penn State, b NBER, c U. de los Andes,Search d Census andBureau Expo
Slide 30: Indirect inference: basic idea
Indirect inference (Gouriéroux and Monfort, 1996) basic idea Using reduced-form auxillary regressions and/or moments, summarize b key relationships in the data using a vector of statistics (M) For a candidate set of parameter values (Λ), simulate same statistics using the model Mb s (Λ). Construct t
Slide 31: Indirect inference: identification
Indirect inference identi…cation Pro…t scaling constants, (Πh , Πf ) means of log home and foreign sales Shipment hazards (λb ) average annual shipment rates per match Product appeal parameters (α, β) distribution of home and foreign sales Firm productivity dispersion (σ ϕ ) distribution of home and
Slide 32: Indirect inference: identification
Indirect inference denti…cation i Idioysncratic shocks to importers (λy ) cross-plant variances in home and foreign sales covariation of home and foreign sales autocorrelation, match-speci…c sales client frequency distribution, client transition probabilites Match maintenance costs (F ) client frequ
Slide 33: Data versus simulated statistics
Data versus simulated statistics Transition probs., Share of …rms no. clients (nc ) Data Model exporting Data Model b [n c c b (1 f ) P jt +1 = 0jnjt = 1] 0.618 0.534 E X >0jt 0.299 0.351 b [n c c P jt +1 = 1jnjt = 1] 0.321 0.358 b [n c c P jt +1 = 2jnjt = 1] 0.048 0.082 Log foreign sales on b [n P
Slide 34: Data versus simulated statistics
Data versus simulated statistics Match death hazards Data Model Exporter exit rate Data Model Death rate, Am ijt 1 = 0 0.694 0.857 Exit rate, Am ijt 1 = 0 0.709 0.748 Death rate, Am ijt 1 = 1 0.515 0.329 Exit rate, Am ijt 1 = 1 0.383 0.099 Death rate, Am ijt 1 = 2 0.450 0.304 Exit rate, Am ijt 1 = 2
Slide 35: Data versus simulated statistics
Data versus simulated statistics Log sales per client Ave. log sales vs. no. clients Data Model by cohort age Data Model m b β1 2.677 0.842 b E (ln Xjtf jAcjt = 0) 8.960 9.306 m b β -0.143 0.042 b (ln X f jAc = 1) E 10.018 10.806 2 jt jt e ( em ) sb 2.180 1.622 b (ln X f jAc = 2) E 10.231 10.755 jt
Slide 36: Data versus simulated statistics
Data versus simulated statistics Match death Log match prob regression Data Model sale autoreg. Data Model d f b β 1.174 1.640 b β 0.811 0.613 0 1 d b β1st year 0.166 0.203 f β1st year 0.233 0.370 d b βlsales -0.070 -0.100 e ( ef ) sb 1.124 0.503 e ( ed ) sb 0.453 0.395 Log dom. sales Match shipment
Slide 37: Parameters
Parameters Parameters Estimated using indirect inference (Λ) Parameter value std. error rate of exogenous separation δ 0.267 0.001 domestic market size Πh 11.344 0.017 foreign market size Πf 10.675 0.017 log …xed cost ln F 7.957 0.018 First θ distribution parameter α 0.716 0.007 Second θ distributio
Slide 38: Parameters
Parameters Parameters Estimated using indirect inference (Λ) Parameter value std. error demand shock jump hazard λy 0.532 0.001 demand shock jump size ∆y 0.087 0.001 shipment order arrival hazard λb 8.836 0.006 std. deviation, log …rm type σϕ 0.650 0.002 network e¤ect parameter γ 0.298 0.001 search
Slide 39: The policy function
The policy function Search intensity over trials and productivity, holding the number of successes constant at 0. original no successes 0.2 0.15 se a r ch in ten sity 0.1 0.05 0 0 5 0 10 5 15 20 10 25 15 productivity trials Eaton et al. (a Penn State, b NBER, c U. de los Andes,Search d Census andBur
Slide 40: History and the policy function
History and the policy function Search intensity as a function of past successes and failures, allowing for reputation e¤ects o r ig in a l 4 3 .5 3 se a rch in te n sity 2 .5 2 1 .5 1 0 .5 0 25 20 25 15 20 10 15 10 5 5 0 0 fa ilu r e s s uc c es es Eaton et al. (a Penn State, b NBER, c U. de los An
Slide 41: History and the policy function
History and the policy function Search intensity as a function of past successes and failures, shutting down reputation e¤ects networkp olicy no network 0.35 0.3 0.25 search intensity 0.2 0.15 0.1 25 20 25 15 20 10 15 10 5 5 0 0 failures 1 0.pdf succeses Eaton et al. (a Penn State, b NBER, c U. de l
Slide 42: A 20% reduction in search costs
A 20% reduction in search costs 18.7 6.35 6.3 Log Active Exporters 18.6 Log Total Exports 6.25 18.5 6.2 18.4 6.15 18.3 6.1 1992 1997 2002 2007 2012 1992 1997 2002 2007 2012 Searc h C os t Dec reas e 11.6 0.6 Log Mean Number of C lients Log Mean Sales per C lient 0.58 11.5 0.56 11.4 0.54 11.3 0.52 0.
Slide 43: A 20% reduction in fixed costs
A 20% reduction in …xed costs 19 6.6 18.9 6.5 Log Active Exporters Log Total Exports 18.8 18.7 6.4 18.6 6.3 18.5 6.2 18.4 18.3 6.1 1992 1997 2002 2007 2012 1992 1997 2002 2007 2012 Fix ed C os t D ec reas e 11.7 0.7 Log Mean N umber of C lients Log Mean Sales per Client 11.6 0.65 11.5 0.6 11.4 0.55
Slide 44: A 20% increase in foreign market size
A 20% increase in foreign market size 19.2 6.6 19 6.5 Log Active Exporters Log Total Exports 18.8 6.4 18.6 6.3 18.4 6.2 18.2 6.1 1992 1997 2002 2007 2012 1992 1997 2002 2007 2012 Mac ro Shoc k Lev el Inc r eas e 11.8 0.7 Log Mean Number of Clients Log Mean Sales per Client 11.7 0.65 11.6 0.6 11.5 0.
Slide 45: Eliminating reputation effects
Eliminating reputation e¤ects 19 6.5 18.5 6 Log Active Exporters Log Total Expor ts 18 5.5 17.5 5 17 4.5 16.5 16 4 1992 1997 2002 2007 2012 1992 1997 2002 2007 2012 N o N etw ork Effec ts 12.2 0.7 Log Mean Number of Clients Log Mean Sales per Client 0.6 12 0.5 11.8 0.4 11.6 0.3 0.2 11.4 0.1 11.2 0 1
Slide 46: Summary
Summary Micro patterns of transactions and buyer-seller relationships through the lens of the model: Large volume of small scale exporters explained by large volume of inexperienced …rms, searching at a low level. High exit rate re‡ects short lifespan of typical match, combined with low-level search
Slide 47: Hazards
Hazards From the perspective of time 0, let the probability that an event will occur before time t be described by the exponential distribution: F [t ] = 1 e qt The likelihood of the event happening exactly at t (the "hazard rate" at t) is then: f (t ) qe qt = =q 1 F (t ) e qt This hazard rate doesn
Slide 48: Hazards
Hazards Suppose k independent events occur with hazard q1 , q2 , ...qk . The probability that none occur before t is: k ∏ (1 Fj (t )) = e t Σj qj j =1 So by time t, at least one event occurs with probability 1 e t Σj q j , and the likelihood that this happens exactly at t is Σ j qj e t Σ j q j = Σ j
Slide 49: Relationship dynamics: Ehrenfest jump processes
Relationship dynamics Ehrenfest jump processes Any variable z that obeys Ehrenfest process: changes value with hazard λz . Next jumps occur within interval t with probability F [t ] = 1 e λz t has discrete support, equally-spaced values: e 2 I + : z 2 f e∆, (e 1)∆, .., 0, .., (e 1)∆, e∆g jumps only
Slide 50: Relationship dynamics: match continuation value
Relationship dynamics match continuation value X be the hazard of transiting from market state x to state x 0 . let qxx 0 Y be the hazard of transiting from match-speci…c state y to let qyy 0 state y 0 . λXx = ∑x 0 6=x qxx X is hazard of any change in market-wide state x 0 λYy = ∑y 0 6=y qyy Y is ha
Slide 51: Relationship dynamics: match continuation value
Relationship dynamics match continuation value Continuation value of a business relationship in state (x, y ) for a type-ϕ exporter : " 1 b ϕ (x, y ) = Eτb e (ρ+δ)τb b π λ + λXx + λYy !# ∑ 0 X qxx b ϕ (x 0 , y ) + ∑ qyy 0π 0 Y b ϕ (x, y 0 ) + λb π 0π e ϕ (x, y ) x 6 =x y 6 =y ! 1 = h ∑ 0 X qxx b ϕ (
Slide 52: Learning about product appeal
Learning about product appeal experience and expected success rates Suppress market superscripts to reduce clutter. The prior distribution is: Γ(α + β) r (θ jα, β) = ( θ ) α 1 (1 θ )β 1 , Γ(α)Γ( β) The likelihood: Given θ, and given that a seller has met n potential buyers, the probability that a of
Slide 53: Learning about product appeal
Learning about product appeal experience and expected success rates The posterior distribution for θ: p (θ ja, n ) ∝ q [ajn, θ ] r (θ jα, β) The expected success rate after a successes in n trials is thus: a+α θ (a, n ) = E [θ ja, n ] = n+α+β Sellers base their search intensity on this posterior mea
Slide 54: Searching for buyers: the value of search
Searching for buyers the value of search The value of continued search for a type-ϕ …rm with a successes in n meetings is: V ϕ (a, n, x ) = " Z τs e ρτs max Eτs s c (s, a) 0 e ρt dt + s + λXx ∑ 0 X qxx 0 0 V ϕ, (a, n, x ) x 6 =x e ϕ (x ) + V ϕ (a + 1, n + 1, x ) + (1 + s θ a,n (π θ a,n )V ϕ (a, n +
Slide 55: Searching for buyers: the value of search
Searching for buyers the value of search Taking expectations over τ s yields: V ϕ (a, n, x ) " 1 = max c (s, a) + ∑ qxx X 0 0 V ϕ, (a, n, x ) s ρ + s + λXx x 0 6 =x e ϕ (x ) + V ϕ (a + 1, n + 1, x ) + (1 + s θ a,n π θ a,n )V ϕ (a, n + 1, x ) The …rst-order condition is thus: e ϕ (x ) + V ϕ (a + 1, n
Slide 56: Searching for buyers: the domestic market
Searching for buyers when the truth is known: the domestic market In the domestic market the reward to search depends on a and n only through network e¤ects. The value of search at home is thus simply: " # 1 V ϕ (x ) = max c (s, a) + ∑ qxx X 0 e ϕ (x ) 0 V ϕ (x ) + sθ j π s ρ + λX 0 x x 6 =x The ass
Banque Nationale De Belgique — Investor Presentation (June 2013) on 6ix