Forecasting turning points of the business cycle: dynamic logit models for panel data

Transcription

1 The 9th Biennial Conference of the Czech Economic Society Forecasting turning points of the business cycle: dynamic logit models for panel data Anna Pestova Senior expert, CMASF Research fellow, National Research University Higher School of Economics, Institute for Economic Forecasting Prague, Czech Republic

2 The research question How to classify the states of the business cycle into the recessionary and expansionary episodes in advance? More importantly, how to predict recessions? Useful for macroeconomic forecasting Linear forecasting models tend to perform poorly at the outbreaks of recessions - Chauvet, Potter (203, Handbook of Ec.Forec.) Nonlinear models may improve the forecasting accuracy at the turning points of the cycle Panel data approach is helpful for the countries with short comparable macroeconomic data series European economies due to introduction of the Euro, post-communist economies,.. 2

3 Literature review 3 Recession prediction with discrete dependent variable models Estrella and Mishkin (998, REStat); Kauppi and Saikkonen (2008, REStat); Nyberg (200, JoF); Ng (202, JMacro), Christiansen at al. (204, JBF) Dynamic models with «classical recession predictors» (government bond term spread, short-term interest rate, stock market index), macroeconomic leading indicators, Ng (202) and sentiment variables, Christiansen at al. (204) Single country approach, mainly for the US Ex-ante fixed threshold (value that splits the predicted values of the binary dependent variable into the recessionary states and normal states) The determinants of duration of expansions and recessions Castro (200, JMacro) Panel data of 3 developed countries Additional explanatory variables: OECD leading indicators (calculated for the member countries in the unified methodology), the dynamics of private investment and the US business cycle phase

4 Literature review Evaluating the classification of business cycle phases Berge and Jorda (20, AEJ: Macro) Single-indicator analysis. Different indicators of economic activity, different recession definitions Receiver operating characteristic (ROC) analysis. Rank models on the entire space of classification «trade-offs» Dynamic discrete dependent variable models in a panel framework Candelon et al. (204, IJF) Currency crisis early warning systems (EWS) 6 emerging countries, dynamic fixed effects panel model 4

5 Contribution We generalize single-country time-series approach for recession forecasting for the panel data of OECD countries We test classical recession predictors and recently introduced sentiment indicators at the panel of countries We add to the literature by introducing additional countercyclical predictors Proxies for macroeconomic imbalances, signals of economy overheating, etc. We test whether one model with large set of predictors for a panel of countries could beat its competitor with country-specific OECD leading indicator in recession forecasting. If yes evidence in favor of panel data model (it fits all countries) 5

6 Methodology and data Panel quarterly dataset on OECD countries (initially 32, finally ~ 20 due to data availability) over the period *. Dependent variable state of the economy / business cycle phase (binary) y it =, if the economy i is in a recessionary state at time t 0, if the economy i is in a expansionary state at time t Methodology: dynamic panel fixed effects logit model Pr y it = α i, x it k, y it k = F y it k γ + x it k β + α i Pr(.) is a conditional probability of recession F(.) is a logistic distribution function x it-k is a set of explanatory variables for country i at quarter (t-k) β is a vector of parameters at x γ is a state dependence parameter (inertia) α i is a country-specific unobservable heterogeneity component 6 k is the quarter lag:, 2 and 4 quarters ahead models * GDP contraction during transitional period in post socialist countries was removed from the sample

7 Dating business cycle phases Classical (business cycle) Neither international organisation provides business cycle dates for the big set of countries in the same methodology NBER - USA Economic Cycle Research Institute 9 European countries+some Asia and America CEPR Euro area 7 Harding, Pagan (2002, JME) modified Bry Boschan (97) dating algorithm for the quarterly data This rule reproduces NBER dates quite closely (Harding, Pagan, 2005, JAE) I use BBQForExcel Program written by Adrian Pagan (downloaded from his website) to find turning points in GDP volume index (peaks / troughs) Recession = period between peak and trough Expansions = period between trough and peak

8 Estimation strategy 8 Different explanatory variables. Only country OECD leading indicator (OECD CLI) 2. Classical recession predictors (CLASSICAL): spread between long and short term interest rate (term spread), short-term interest rate, stock market growth 3. Sentiment indicators (SENTIMENT): OECD сonsumer and business confidence indicators 4. State of the world economy (WORLD): OECD US GDP leading indicator 5. New countercyclical indicators (NEW_COUNTERCYC): country GDP growth, CPI inflation, current account to GDP ratio, REER index, bank credit to GDP ratio Dynamic models (inertia in business cycle phases is accounted for) Compared with static Different lags One, two, four quarter ahead models

9 How to interpret countercyclical indicators Procyclical leading indicators - reach turning points before the economic activity, move in the same direction Sentiment indicators, interest rates, stock indices, state of the world economy Countercyclical leading indicators at some periods of time go above the «safe» thresholds, meaning that further development is impossible in this way (high inflation and high wages, trade deficit, currency appreciation, financial overheating) Missing link in the modern business cycle theory? What drives business cycles, shocks? But shocks are linear combinations of one step ahead forecasting errors (reduced-form innovations). 9 Are turning points driven by the exogenous «shocks» or they are (at least partially) explained by domestic economic development?

10 Evaluating the classification of business cycle phases Y t c Recession prediction Y t < c Expansion prediction True state of the economy S = Recession True positive rate TPR(с) S= 0 Expansion False positive rate FPR(с) TPR FPR ROC ROC r = TPR c r = FPR c TPR marker Dynamic OECD CLI 0 AUROC = ROC r dr FPR

11 Models with different set of predictors, lag=q () (2) (3) (4) (5) (6) (7) (8) (9) (0) () (2) (3) (4) Dependent variable: Reces Dynamic models Static model L.Reces 3.07*** 4.345*** 4.345*** 4.438*** 4.584*** 4.494*** 4.369*** 4.644*** 4.468*** 4.588*** 4.602*** 4.602*** 4.587*** (.70) (9.53) (9.53) (8.05) (7.63) (7.09) (5.58) (3.43) (3.03) (3.05) (2.92) (2.92) (3.05) L.oecd_cli *** (-7.05) L.term_spread -0.23** -0.28* (-2.20) (-.80) (-.50) (-0.98) (-0.67) (-0.58) (-0.47) (-.04) (-0.87) (-.0) (-.0) (-.0) (-.37) L.money_market_rate *** -0.42*** -0.56*** -0.56*** (-0.09) (-0.07) (-0.3) (-0.7) (0.02) (-0.7) (-2.64) (-2.67) (-2.89) (-2.89) (-0.90) (-.02) L.t_stock_q -0.00*** *** *** *** *** *** *** *** *** *** *** (-8.35) (-6.8) (-5.5) (-4.82) (-4.65) (-4.30) (-4.30) (-4.25) (-4.25) (-4.9) (-3.74) L.t_consum_confid_q *** *** -0.83*** *** *** *** *** *** -0.64*** *** (-4.5) (-3.37) (-3.47) (-3.40) (-3.0) (-2.87) (-2.8) (-2.8) (-2.68) (-2.58) L.t_bus_confid_q ** -0.45** * * * * * -0.44* *** (-2.43) (-2.07) (-.83) (-.87) (-.86) (-.76) (-.76) (-.85) (-4.73) L.oecd_cli_usa (-.2) (-.60) (-.32) (-0.96) (-.4) (-.4) (-0.64) (-.36) L.gdp_gr *** (.43) (0.86) (.20) (.39) (.39) (.48) (-8.40) L.cpi_inflat 0.32*** 0.344*** 0.360*** 0.360*** 0.36*** 0.493*** (4.08) (4.4) (4.56) (4.56) (4.28) (6.24) L.cab_to_gdp_4q -0.30** -0.9** -0.9** * (-2.43) (-2.22) (-2.22) (-.72) (-0.23) L.reer (.63) (.63) (0.57) (0.8) L.loans_to_gdp 0.06*** 0.07*** (3.20) (3.59) _cons ** ** * * -.987* * ** -3.23*** *** *** *** ** (-.22) (-2.35) (-2.24) (-.54) (-.72) (-.76) (-.66) (-.69) (-2.2) (-2.79) (-2.80) (-2.80) (-2.6) (-2.50) N pseudo R-sq ll chi

12 True positive rate Classification results, lag= quarter AUROC OECD CLI CLASSIC+SENTIMENT+ WORLD CLASSIC+SENTIMENT+ WORLD+NEW COUNTERCYCL Prob (Areas are equal) = Differencies are minor. All AUROCs are statistically equal 0 ROC Model Dynamic OECD CLI Lag= Dynamic Classical+Sentiment+International Lag= Dynamic Clas+Sent+Internat+New_countercycl Lag= 0 False positive rate 2

13 Models with different set of predictors, lag=2q () (2) (3) (4) (5) (6) (7) (8) (9) (0) () (2) (3) (4) Dependent variable: Reces Dynamic models Static model L2.Reces.080*** 2.480*** 2.479*** 2.27*** 2.272*** 2.66*** 2.68*** 2.507*** 2.309*** 2.446*** 2.456*** 2.456*** 2.45*** (4.50) (4.06) (4.03) (2.4) (.74) (0.75) (0.5) (9.35) (8.63) (8.97) (8.93) (8.93) (8.6) L2.oecd_cli -0.48*** (-7.30) L2.term_spread -0.22** -0.3* -0.8* * (-2.23) (-.8) (-.73) (-.20) (-0.97) (-0.98) (-0.89) (-.60) (-.4) (-.64) (-.64) (-.87) (-.53) L2.money_market_rate *** -0.76*** -0.99*** -0.99*** (-0.22) (-0.25) (-0.5) (-0.26) (-0.26) (-0.68) (-4.02) (-4.5) (-4.5) (-4.5) (-.49) (-.08) L2.t_stock_q *** *** *** *** *** *** *** *** *** *** *** (-7.74) (-5.63) (-4.58) (-4.55) (-4.44) (-4.25) (-4.28) (-4.25) (-4.25) (-4.5) (-3.88) L2.t_consum_confid_q *** ** ** ** ** -0.58** -0.52** -0.52** * * (-3.45) (-2.49) (-2.46) (-2.53) (-2.37) (-2.23) (-2.27) (-2.27) (-.90) (-.70) L2.t_bus_confid_q -0.38** ** ** -0.35** * * * ** *** (-2.23) (-2.2) (-.96) (-.96) (-.92) (-.78) (-.78) (-2.00) (-3.30) L2.oecd_cli_usa (0.03) (-0.64) (0.22) (0.77) (0.47) (0.47) (.30) (0.45) L2.gdp_gr 0.2** * 0.9** 0.9** 0.2** -0.50*** (2.0) (.3) (.83) (2.7) (2.7) (2.0) (-3.2) L2.cpi_inflat 0.348*** 0.394*** 0.420*** 0.420*** 0.43*** 0.465*** (5.52) (6.39) (6.63) (6.63) (6.7) (6.7) L2.cab_to_gdp_4q -0.57*** -0.40*** -0.40*** -0.09*** -0.06* (-3.82) (-3.38) (-3.38) (-2.63) (-.65) L2.reer 0.029*** 0.029*** (2.97) (2.97) (.20) (0.88) L2.loans_to_gdp 0.02*** 0.023*** (5.09) (5.20) _cons -2.04*** *** -2.82*** *** *** *** -3.0*** *** *** *** -6.96*** -6.96*** *** -5.26*** (-4.60) (-7.57) (-6.44) (-6.25) (-6.75) (-6.89) (-6.70) (-6.89) (-7.42) (-8.0) (-5.82) (-5.82) (-5.3) (-4.34) N pseudo R-sq ll chi

14 True positive rate Classification results, lag=2 quarters AUROC ROC OECD CLI CLASSIC+SENTIMENT+ WORLD CLASSIC+SENTIMENT+ WORLD+NEW COUNTERCYCL Prob (Areas are equal) = 0.09 Now differencies in AUROCs are more pronounced and statistically significant 0 Model Dynamic OECD CLI Lag=2 Dynamic Classical+Sentiment+International Lag=2 Dynamic Clas+Sent+Internat+New_countercycl Lag=2 0 False positive rate 4

15 Models with different set of predictors, lag=4q () (2) (3) (4) (5) (6) (7) (8) (9) (0) () (2) (3) (4) Static Dependent variable: Reces Dynamic models model L4.Reces *** 0.433** 0.47** (-3.22) (2.25) (2.3) (0.55) (0.2) (0.30) (0.44) (.8) (0.56) (0.87) (0.80) (0.80) (0.34) L4.oecd_cli -0.36*** (-7.46) L4.term_spread -0.04* -0.86** -0.77** -0.33* -0.39* -0.45* -0.47* -0.2** -0.83** -0.94** -0.94** -0.88** -0.86** (-.76) (-2.7) (-2.3) (-.65) (-.68) (-.74) (-.76) (-2.25) (-2.03) (-2.9) (-2.9) (-2.42) (-2.38) L4.money_market_rate * * * -0.87*** -0.80*** *** *** (-.66) (-.7) (-.53) (-.50) (-.57) (-.82) (-4.58) (-4.49) (-4.96) (-4.96) (-.09) (-.08) L4.t_stock_q *** -0.03*** *** *** *** *** *** *** *** *** *** (-5.92) (-3.90) (-4.09) (-4.0) (-4.20) (-4.08) (-4.35) (-4.26) (-4.26) (-3.97) (-3.97) L4.t_consum_confid_q *** *** *** *** *** *** -0.54*** -0.54*** ** ** (-2.79) (-2.72) (-2.70) (-2.94) (-2.83) (-2.78) (-2.84) (-2.84) (-2.54) (-2.52) L4.t_bus_confid_q (0.82) (0.6) (0.85) (0.87) (.04) (.9) (.9) (0.78) (0.72) L4.oecd_cli_usa ** 0.09** (0.54) (0.08) (.05) (.54) (.29) (.29) (2.38) (2.32) L4.gdp_gr 0.077* ** 0.088** 0.095** 0.087** (.88) (0.82) (.62) (.97) (.97) (2.03) (2.20) L4.cpi_inflat 0.292*** 0.335*** 0.37*** 0.37*** 0.339*** 0.34*** (4.84) (5.83) (6.23) (6.23) (5.04) (5.00) L4.cab_to_gdp_4q -0.77*** -0.58*** -0.58*** -0.0*** -0.08*** (-4.88) (-4.30) (-4.30) (-2.98) (-2.98) L4.reer 0.033*** 0.033*** (3.58) (3.58) (.58) (.58) L4.loans_to_gdp 0.025*** 0.025*** (6.05) (6.07) _cons -.764*** *** *** -2.0*** -2.64*** -2.58*** -2.9*** *** *** *** *** *** -6.0*** *** (-4.78) (-6.89) (-4.70) (-4.55) (-4.96) (-4.93) (-4.90) (-5.4) (-5.68) (-6.60) (-5.62) (-5.62) (-4.84) (-4.84) N pseudo R-sq ll chi

16 True positive rate Classification results, lag=4 quarters AUROC ROC OECD CLI CLASSIC+SENTIMENT+ WORLD CLASSIC+SENTIMENT+ WORLD+NEW COUNTERCYCL Prob (Areas are equal) = Differencies in AUROCs are pronounced and statistically significant 0 Model Static OECD CLI Lag=4 Static Classical+Sentiment+International Lag=4 Static Clas+Sent+Internat+New_countercycl Lag=4 0 False positive rate 6 For long forecasting horizons we outperform the model with country-specific CLI! Panel data model fits all countries?

17 Different types of model Lag=q Lag=2q Lag=4q Static Dynamic (Dynamic Static) *** ** 0 Note: ***/** - difference is statistically significant at %/5% Area under ROC Pr y it = α i, x it k, y it k = F y it k γ + x it k β + α i Static model: γ=0 7 Dynamic models are more accurate that static at short horizons. Dynamics is unimportant at 4 quarter lag

18 True positive rate ROC-curves for models with different lags AUROC Lag= Q Lag=2 Q Lag=4 Q Dynamic models with large set ROC of predictors Prob (Areas are equal) = Quality of in-sample fit decreases significantly with lag increase. Model 8 0 Dynamic Clas+Sent+Internat+New_countercycl Lag= Dynamic Clas+Sent+Internat+New_countercycl Lag=2 Static Clas+Sent+Internat+New_countercycl Lag=4 0 False positive rate

19 2007 Q 2007 Q Q Q2 200 Q 200 Q4 20 Q3 202 Q2 203 Q 203 Q4 204 Q3 205 Q2 206 Q How this model works for Czech Republic In-sample (up to 203) and out-of-sample ( ) forecasts of recession probabilities, 4 quarters ahead BBQ Reces Prob(Reces) Threshold 9

20 2007 Q 2008 Q 2009 Q 200 Q 20 Q 202 Q 203 Q 204 Q 205 Q 206 Q 2007 Q 2007 Q Q Q2 200 Q 200 Q4 20 Q3 202 Q2 203 Q 203 Q4 204 Q3 205 Q2 206 Q 2007 Q 2007 Q Q Q2 200 Q 200 Q4 20 Q3 202 Q2 203 Q 203 Q4 204 Q3 205 Q2 206 Q 2007 Q 2007 Q Q Q2 200 Q 200 Q4 20 Q3 202 Q2 203 Q 203 Q4 204 Q3 205 Q2 206 Q Other countries In-sample (up to 203) and out-of-sample ( ) forecasts of recession probabilities, 4 quarters ahead United States BBQ Reces 0.4 Prob(Reces) 0.2 Threshold 0 UK BBQ Reces Prob(Reces) Threshold Germany Spain BBQ Reces Prob(Reces) Threshold BBQ Reces Prob(Reces) Threshold

21 2007 Q 2007 Q Q 2008 Q Q 2009 Q3 200 Q 200 Q3 20 Q 20 Q3 202 Q 202 Q3 203 Q 203 Q3 204 Q 204 Q3 205 Q 205 Q3 206 Q Russia In-sample (up to 203) and out-of-sample ( ) forecasts of recession probabilities, 4 quarters ahead BBQ Reces Prob(Reces) Threshold 2 9

22 Main results Business cycle turning points can be predicted on the country panel data set in a uniform way, and the quality of these predictions is comparable to the analogues for country-specific leading indicator models for short horizons and even outperform them at medium-term horizons Lags of the dependent variable do matter at short horizons (dynamic mechanism) We test classical recession predictors, sentiment indicators and introduced here countercyclical (macro imbalances) indicators. We conclude that at different forecasting horizons different list of predictors work Term spread appears to be significant only at medium-term horizon. Stock market is useful at all horizons tested With lag increase the predictive power of business confidence indicators vanishes, while consumer sentiment is significant for up to one year lag (business expectations seem to be less accurate at long horizons). The role of domestic macro and financial imbalances and external trade misalignments, goes up at longer horizons. With these indicators we are able to build acceptable models with a year lag Quality of in-sample fit decreases significantly with lag increase. Trade-off between forecasting accuracy and the earliness of the recession signal