Surprise indexes and nowcasting: why do markets react to macroeconomic news?

Transcription

1 Introduction: surprise indexes Surprise indexes and nowcasting: why do markets react to macroeconomic news? Alberto Caruso Confindustria and Université libre de Bruxelles Conference on Real-Time Data Analysis, Methods and Applications Banco de España, Madrid, October / 35

2 Outline Introduction: surprise indexes 1 Introduction: surprise indexes / 35

3 Outline Introduction: surprise indexes 1 Introduction: surprise indexes / 35

4 Introduction Introduction: surprise indexes Macroeconomic news market reactions Established result! Among the others see Andersen et al. (23, 27); Ehrmann and Fratzscher (25); Gürkaynak et al. (25); Faust et al. (27); Gürkaynak and Wright (213). The effect is persistent and amplified at lower frequencies, explaining more than one third of the quarterly fluctuations of bond yields (Altavilla et al., 217). 4 / 35

5 Introduction Introduction: surprise indexes How to aggregate macro news? Different concepts, units, importance. Solutions so far: 5 / 35

6 Introduction Introduction: surprise indexes How to aggregate macro news? Different concepts, units, importance. Solutions so far: Surprise indexes (e.g. Citi, Altavilla et al., 217) use market forecast to construct the news, weighting them on the basis of their effect on assets. 5 / 35

7 Introduction Introduction: surprise indexes How to aggregate macro news? Different concepts, units, importance. Solutions so far: Surprise indexes (e.g. Citi, Altavilla et al., 217) use market forecast to construct the news, weighting them on the basis of their effect on assets. Nowcasting (Giannone et al., 28; Banbura et al., 213): look at the impact of news on the update of the assessment of the current state of the economy. 5 / 35

8 Introduction Introduction: surprise indexes How to aggregate macro news? Different concepts, units, importance. Solutions so far: Surprise indexes (e.g. Citi, Altavilla et al., 217) use market forecast to construct the news, weighting them on the basis of their effect on assets. Nowcasting (Giannone et al., 28; Banbura et al., 213): look at the impact of news on the update of the assessment of the current state of the economy. This work I merge the two approaches and study the relationship between surprise indexes and nowcasting 5 / 35

9 Introduction Introduction: surprise indexes How to aggregate macro news? Different concepts, units, importance. Solutions so far: Surprise indexes (e.g. Citi, Altavilla et al., 217) use market forecast to construct the news, weighting them on the basis of their effect on assets. Nowcasting (Giannone et al., 28; Banbura et al., 213): look at the impact of news on the update of the assessment of the current state of the economy. This work I merge the two approaches and study the relationship between surprise indexes and nowcasting Other "hybrid" examples: Scotti (216); Grover et al. (216); Gilbert et al. (217) 5 / 35

10 Introduction: surprise indexes Different types of surprise indexes To construct surprise indexes, we can combine: 6 / 35

11 Introduction: surprise indexes Different types of surprise indexes To construct surprise indexes, we can combine: Market news: (actual - survey forecast) 6 / 35

12 Introduction: surprise indexes Different types of surprise indexes To construct surprise indexes, we can combine: Market news: (actual - survey forecast) Model news: (actual - model forecast) 6 / 35

13 Introduction: surprise indexes Different types of surprise indexes To construct surprise indexes, we can combine: Market news: (actual - survey forecast) Model news: (actual - model forecast) Market weights: impact on asset prices (FX or 1y yields) 6 / 35

14 Introduction: surprise indexes Different types of surprise indexes To construct surprise indexes, we can combine: Market news: (actual - survey forecast) Model news: (actual - model forecast) Market weights: impact on asset prices (FX or 1y yields) Model weights: importance assigned by a model 6 / 35

15 Introduction: surprise indexes Different types of surprise indexes To construct surprise indexes, we can combine: Market news: (actual - survey forecast) Model news: (actual - model forecast) Market weights: impact on asset prices (FX or 1y yields) Model weights: importance assigned by a model News i,t x i,t E[(x i,t Info ν )] t SI t W i,s News i,s s=t win i I 6 / 35

16 Introduction: surprise indexes Different types of surprise indexes Weights News Market Model Market Citi; Altavilla et al. THIS PAPER Model Grover et al.; Scotti; Gilbert et al. THIS PAPER 7 / 35

17 Introduction: surprise indexes Different types of surprise indexes Weights Model weights: News Market Model Market Citi; Altavilla et al. THIS PAPER Model Grover et al.; Scotti; Gilbert et al. THIS PAPER Grover et al. (216): weights from regressions of market news on GDP (predictive content) Scotti (216): contributions of the variables in forecasting the factor Gilbert et al. (217): regression coefficients of announcements multiplied by Kalman gain matrix 7 / 35

18 Outline Introduction: surprise indexes 1 Introduction: surprise indexes / 35

19 Outline Introduction: surprise indexes 1 Introduction: surprise indexes / 35

20 Introduction: surprise indexes Nowcasting: model-based news and weights Let yt Q be the GDP at time t, and Ω ν the information set at time ν, where ν is a vintage of data. The nowcast is the projection of yt Q using the available data, E[yt Q Ω ν ]. At any release, the information set expands : Ω ν Ω ν+1, and it is possible to decompose the new forecast in: E[y Q t Ω ν+1 ] }{{} new forecast = E[yt Q Ω ν ] }{{} old forecast + E[y Q t I ν+1 ] }{{} revision Where I ν+1 is the information in Ω ν+1 orthogonal to Ω ν. Express the revision as: E[yt Q Ω ν+1 ] E[yt Q Ω ν ] = w j,t,ν+1 (x ij,t }{{} j E[x ij,t j Ω ν ]) }{{} j J revision ν+1 news 1 / 35

21 Introduction: surprise indexes Nowcasting: model-based news and weights Let yt Q be the GDP at time t, and Ω ν the information set at time ν, where ν is a vintage of data. The nowcast is the projection of yt Q using the available data, E[yt Q Ω ν ]. At any release, the information set expands : Ω ν Ω ν+1, and it is possible to decompose the new forecast in: E[y Q t Ω ν+1 ] }{{} new forecast = E[yt Q Ω ν ] }{{} old forecast + E[y Q t I ν+1 ] }{{} revision Where I ν+1 is the information in Ω ν+1 orthogonal to Ω ν. Express the revision as: E[yt Q Ω ν+1 ] E[yt Q Ω ν ] = w j,t,ν+1 (x ij,t }{{} j E[x ij,t j Ω ν ]) }{{} j J revision ν+1 news 11 / 35

22 Introduction: surprise indexes The weights represent the importance given by the model to a macro news in the update of the backcast-nowcast-forecast. Nowcasting is a fixed event forecast and refers to GDP in a specific quarter: we should not use the nowcast directly! I use a rolling window with a consistent weighting scheme. 12 / 35

23 Introduction: surprise indexes News Index construction: Weighting scheme 13 / 35

24 Introduction: surprise indexes News Index construction: Weighting scheme 13 / 35

25 Introduction: surprise indexes News Index construction: Weighting scheme 13 / 35

26 Introduction: surprise indexes News Index construction: Weighting scheme 13 / 35

27 Introduction: surprise indexes News Index construction: Weighting scheme 13 / 35

28 Introduction: surprise indexes News Index construction: Weighting scheme 13 / 35

29 Introduction: surprise indexes News Index construction: Weighting scheme 13 / 35

30 Introduction: surprise indexes News Index construction: Weighting scheme 13 / 35

31 Introduction: surprise indexes News Index construction: Weighting the weights Let wi,t BC, w i,t NC, w i,t FC be the weights corresponding to the updates in the Backcast, Nowcast and Forecast. Temporally weight them coherent weights. Call d the distance from the beginning of the reference quarter. If d 33, then W i,t = 33+d 66 wi,t NC + 33 d 66 wi,t BC If 33 d 66, then W i,t = 99 d 66 wi,t NC + d wi,t FC This weighting scheme gives us the right "rolling assessment" of macro surprises. 14 / 35

32 Introduction: surprise indexes Why model-based? Advantages of a model-based index: it weights surprises in a coherent way 15 / 35

33 Introduction: surprise indexes Why model-based? Advantages of a model-based index: it weights surprises in a coherent way model expectations are judgement-free, transparent, not prone to mood, herding and strategic behaviour 15 / 35

34 Introduction: surprise indexes Why model-based? Advantages of a model-based index: it weights surprises in a coherent way model expectations are judgement-free, transparent, not prone to mood, herding and strategic behaviour it can be used for every country, also when market expectations are not available 15 / 35

35 Introduction: surprise indexes Why model-based? Advantages of a model-based index: it weights surprises in a coherent way model expectations are judgement-free, transparent, not prone to mood, herding and strategic behaviour it can be used for every country, also when market expectations are not available In this sample model expectations are more efficient. News analysis 15 / 35

36 Outline Introduction: surprise indexes 1 Introduction: surprise indexes / 35

37 Introduction: surprise indexes Nowcasting model Dynamic factor model, REAL TIME out-of-sample exercise (mixed frequency, exact calendar of macro releases) for US GDP QoQ growth rate. I estimate with Maximum Likelihood within a Expectation-Maximization algorithm: it allows for autocorrelation of idio. and restrictions, consistent and feasible even in case of an approximate factor model. Model and estimation details See Giannone et al. (28); Doz et al. (211, 212); Banbura et al. (213). Estimation starts in 1991; evaluation period: ; 1 factor, 2 lags (results are robust to changes in the specification). 17 / 35

38 Introduction: surprise indexes Data: 13 variables compatible with Bloomberg Name Bloomberg Transformation Building Permits MoM Capacity Utilization Diff Civilian Unemployment Rate Diff Conference Board: Consumer Confidence Level Consumer Price Index MoM Housing Starts MoM Industrial Production MoM ISM Mfg: PMI Composite Index Level Producer Price Index MoM Real Gross Domestic Product MoM Total Nonfarm Employment Diff Trade balance MoM University of Michigan: Consumer Sentiment Level 18 / 35

39 Introduction: surprise indexes Data: additional variables for robustness Name Bloomberg Transformation Building Permits MoM Capacity Utilization Diff Civilian Unemployment Rate Diff Conference Board: Consumer Confidence Level Consumer Price Index MoM Housing Starts MoM Industrial Production MoM ISM Mfg: PMI Composite Index Level Producer Price Index MoM Real Gross Domestic Product MoM Total Nonfarm Employment Diff Trade balance MoM University of Michigan: Consumer Sentiment Level 3-Month Treasury Bill Diff 1-Year Treasury Constant Maturity Rate Diff All Employees: Total Private Industries MoM Average Weekly Hours Mfg MoM Commercial and Industrial Loans MoM Disposable Personal Income MoM Inventories to Sales Ratio Diff M2 Money Stock MoM Mfg New Orders: Durable Goods MoM Personal Consumption Expenditures MoM Retail Sales MoM Total Business Inventories MoM 19 / 35

40 Outline Introduction: surprise indexes 1 Introduction: surprise indexes / 35

41 Introduction: surprise indexes Nowcasting Surprise Index 21 / 35

42 Introduction: surprise indexes Surprise Indexes: market and model-based 22 / 35

43 Introduction: surprise indexes Surprise Indexes: market and model-based The indexes are very similar! Similar news: the model replicates market forecasts Similar weights: market reaction is related to the impact of news in changing the assessing the state of the economy (and to its consequences, e.g. monetary policy?) Detail 23 / 35

44 Introduction: surprise indexes Nowcasting Surprise Index 24 / 35

45 Introduction: surprise indexes Nowcasting Surprise Index and S&P 5 excess return 25 / 35

46 Introduction: surprise indexes Nowcasting Surprise Index, S&P 5 and Scotti Index 26 / 35

47 Introduction: surprise indexes Correlations - different frequencies Nowcasting Surprise Index Correlations 1-month 2-months Quarterly Change in 1y yields.23 /.19*.33 /.3*.36 /.41* S&P 5 excess returns.23 /.23*.37 /.36*.42 /.45* Market Surprise Index Correlations 1-month 2-months Quarterly Change in 1y yields S&P 5 excess returns * Larger model with 26 variables Other benchmarks 27 / 35

48 Regressions Introduction: surprise indexes I regress asset prices (change in 1y and S&P5 excess returns) on Surprise Indexes at different frequencies: w AssetPrices i,t = α + β i (SurpriseIndex w t ) + ɛ i,t Where the w can be 22, 44 or 66 working days. For example, if w = 22, w AssetPrices i,t is the monthly return of asset i and SurpriseIndext w is the index aggregated over a month. 28 / 35

49 Introduction: surprise indexes Regression - different frequencies Nowcasting Surprise Index OLS - R 2 1-month 2-months Quarterly Change in 1y yields.5 /.4*.11 /.9*.17 /.17* S&P 5 excess return.4 /.5*.8 /.13*.18 /.21* Market Surprise Index OLS - R 2 1-month 2-months Quarterly Change in 1y yields S&P 5 excess return * Larger model with 26 variables Other benchmarks 29 / 35

50 Introduction: surprise indexes Different kind of indexes and correlations Recall, we started from here: Weights News Market Model Market Citi; Altavilla et al. THIS PAPER Model Grover et al.; Scotti; Gilbert et al. THIS PAPER I compare the differences in correlation with S&P5 of indexes with model/market news and weights. Market weights: coefficients of the regression of news on asset prices (as in Altavilla et al., 217). See market weights 3 / 35

51 Introduction: surprise indexes Correlations with S&P5: market and model Quarterly correlation of model/market based indexes with S&P5. Weights News Market Model Market Model.4.42 /.25* * Nowcast (not rolling; no rolling weights) 31 / 35

52 Introduction: surprise indexes Markets and model: extracting the judgemental component Forecasts differences to be explored: How much and when? (work in progress) 32 / 35

53 Introduction: surprise indexes Markets and model: extracting the judgemental component Forecasts differences to be explored: How much and when? (work in progress) I define "judgemental component": the part of a survey forecast not implied by a nowcasting model: BB ij,t j JC i,t = (BB i,t E[x i,t Ω ν ]) 2 = median of Bloomberg surveys; E[(x i,t Ω ν )] = model expectation I aggregate the deviations between market/model forecast across the variables and smooth it: JC follows disagreement and in specific episodes is more pronounced (non linearities, sentiment). 32 / 35

54 Introduction: surprise indexes News analysis - "judgemental" component and disagreement 33 / 35

55 Outline Introduction: surprise indexes 1 Introduction: surprise indexes / 35

56 Introduction: surprise indexes I study the relationship between surprise indexes and nowcasting constructing a real-time Nowcasting Surprise Index, weighting in a coherent way news about macro indicators. 35 / 35

57 Introduction: surprise indexes I study the relationship between surprise indexes and nowcasting constructing a real-time Nowcasting Surprise Index, weighting in a coherent way news about macro indicators. Market-based and model-based indexes give similar results! Robust to change in # of variables and specification. 35 / 35

58 Introduction: surprise indexes I study the relationship between surprise indexes and nowcasting constructing a real-time Nowcasting Surprise Index, weighting in a coherent way news about macro indicators. Market-based and model-based indexes give similar results! Robust to change in # of variables and specification. Market reacts in correspondence of news that change the assessment of macro conditions. 35 / 35

59 Introduction: surprise indexes I study the relationship between surprise indexes and nowcasting constructing a real-time Nowcasting Surprise Index, weighting in a coherent way news about macro indicators. Market-based and model-based indexes give similar results! Robust to change in # of variables and specification. Market reacts in correspondence of news that change the assessment of macro conditions. Both indexes show good correlation with asset prices: to be explored. 35 / 35

60 Introduction: surprise indexes I study the relationship between surprise indexes and nowcasting constructing a real-time Nowcasting Surprise Index, weighting in a coherent way news about macro indicators. Market-based and model-based indexes give similar results! Robust to change in # of variables and specification. Market reacts in correspondence of news that change the assessment of macro conditions. Both indexes show good correlation with asset prices: to be explored. Further research, in progress: predictive regressions; deviations between market/model forecasts in case of particular events. 35 / 35

61 References Appendix Altavilla, C., Giannone, D., and Modugno, M. (217). Low frequency effects of macroeconomic news on government bond yields. Journal of Monetary Economics. Andersen, T. G., Bollerslev, T., Diebold, F. X., and Vega, C. (23). Micro effects of macro announcements: Real-time price discovery in foreign exchange. The American Economic Review, 93(1): Andersen, T. G., Bollerslev, T., Diebold, F. X., and Vega, C. (27). Realtime price discovery in global stock, bond and foreign exchange markets. Journal of International Economics, 73(2): Banbura, M., Giannone, D., Modugno, M., and Reichlin, L. (213). Nowcasting and the real-time data flow. Handbook of Economic Forecasting, Volume 2, ed. by G. Elliott, and A. Timmermann, NBER Chapters. Elsevier-North Holland. Doz, C., Giannone, D., and Reichlin, L. (211). A two-step estimator for large approximate dynamic factor models based on kalman filtering. Journal of Econometrics, 164(1): Doz, C., Giannone, D., and Reichlin, L. (212). A quasi maximum likelihood approach for large, approximate dynamic factor models. Review of economics and statistics, 94(4): / 19

62 References Appendix Ehrmann, M. and Fratzscher, M. (25). Exchange rates and fundamentals: new evidence from real-time data. Journal of International Money and Finance, 24(2): Faust, J., Rogers, J. H., Wang, S.-Y. B., and Wright, J. H. (27). The high-frequency response of exchange rates and interest rates to macroeconomic announcements. Journal of Monetary Economics, 54(4): Giannone, D., Reichlin, L., and Small, D. (28). Nowcasting: The realtime informational content of macroeconomic data. Journal of Monetary Economics, 55(4): Gilbert, T., Scotti, C., Strasser, G., and Vega, C. (217). Is the intrinsic value of a macroeconomic news announcement related to its asset price impact? Journal of Monetary Economics, 92: Grover, S. P., Kliesen, K. L., and McCracken, M. W. (216). A Macroeconomic News Index for Constructing Nowcasts of U.S. Real Gross Domestic Product Growth. Review, 98(4): Gürkaynak, R. S., Sack, B., and Swanson, E. (25). The sensitivity of long-term interest rates to economic news: Evidence and implications for macroeconomic models. American economic review, pages / 19

63 References Appendix Gürkaynak, R. S. and Wright, J. H. (213). Identification and inference using event studies. The Manchester School, 81(S1): Mariano, R. S. and Murasawa, Y. (23). A new coincident index of business cycles based on monthly and quarterly series. Journal of Applied Econometrics, 18(4): Scotti, C. (216). Surprise and uncertainty indexes: Real-time aggregation of real-activity macro-surprises. Journal of Monetary Economics, 82:1 19. / 19

64 References Appendix Appendix: The model x t = Λf t + ɛ t f t = A 1 f t A p f t p + u t ; u t i.i.d. N (, Q) x t : vector of standardized stationary monthly variables f t : unobserved common factors following a VAR(p) Λ: factor loadings ɛ t : vector of idiosyncratic components following an AR(1) Quarterly variables are modelled as monthly variables with periodically missing values; see the approximation of Mariano and Murasawa (23). Go back 1 / 19

65 References Appendix Real time out-of-sample - US GDP nowcast 2 / 19

66 References Appendix News analysis Studies show that market-based forecast are not always efficient (Pierce and Roley, 1985; Balduzzi et al., 21; Andersen et al., 21; Scotti, 216). I test the efficiency of forecasts F (Bloomberg survey or model-based forecast), testing for α i = β i = in the regression (coefficients jointly significant=not efficient): News i,t = α i + β i F i,t + ɛ i,t (1) 3 / 19

67 References Appendix News analysis Bloomberg surveys α β F F-pvalue Industrial Production -.3 ***.781 *** ***. Capacity Utilization **.846 *** ***. Housing Starts *** 8.47 ***.5 Building Permits Trade Balance Change in Nonfarm Payrolls U. of Mich. Sentiment *** -.24 *** ***.5 Unemployment Rate -.27 ** *** ***.6 CPI *** *** ***. PPI *** ***. Consumer Confidence Index ISM Manufacturing GDP Annualized / 19

68 References Appendix News analysis Nowcasting forecasts α β F F-pvalue Industrial Production *** ***. Capacity Utilization *** ***. Housing Starts Building Permits * *.75 Trade Balance Change in Nonfarm Payrolls U. of Mich. Sentiment Unemployment Rate CPI PPI ** **.48 Consumer Confidence Index ISM Manufacturing GDP Annualized Go back 5 / 19

69 References Appendix 2 IP 2 CapUt 5 HousSt 2 NewPrivateH Feb4 Aug9 Feb15 5 TradeBal -2 Feb4 Aug9 Feb15 5 NonfarmEmpl -5 Feb4 Aug9 Feb15 1 UoM -4 Feb4 Aug9 Feb15 1 Unempl -1-5 Feb4 Aug9 Feb15 1 CPI -5 Feb4 Aug9 Feb15 5 PPIFinGoods -2 Feb4 Aug9 Feb15 2 PMIMfg -1 Feb4 Aug9 Feb15 1 ConsConf -2-1 Feb4 Aug9 Feb15 1 GDP -5 Feb4 Aug9 Feb15-4 Feb4 Aug9 Feb15-1 Feb4 Aug9 Feb Feb4 Aug9 Feb15 Bloomberg News Nowcasting News 6 / 19

70 References Appendix News analysis - forecasting professional forecasters I forecast the median of the surveys conducted by Bloomberg at the moment of the release, using the model prediction updated up to the previous macroeconomic release. RMSFE relative to random walk Capacity Utilization.81 Housing Starts.67 Building Permits.72 Trade Balance 1.21 Change in Nonfarm Payrolls.87 U. of Mich. Sentiment.75 Unemployment Rate.73 CPI.63 PPI 1.9 Consumer Confidence Index.62 ISM Manufacturing.44 GDP 1.27 real time out of sample, Go back 7 / 19

71 References Appendix Different kind of indexes and correlations As market weights I take the coefficients of the regression of news on 1y bonds (as in Altavilla et al., 217). Regression of News on daily difference of 1y bonds Model-based AGM (217) Industrial Production Capacity Utilization: Industry ** Housing Starts New Private Housing Units Authorized.952 *.65 Trade Balance.897 * 1.3 ** Total Nonfarm Employment ** University of Michigan: Consumer Sentiment ** Unemployment Rate Consumer Price Index Producer Price Index *** -.13 Cons Conf ** ISM Mfg: PMI *** 2.62 ** GDP ** NB Sample size: Model: / AGM: AGM do a regression with 41 variables. Go back 8 / 19

72 References Appendix Nowcasting Surprise Index, S&P 5 and Citi Index 9 / 19

73 References Appendix Correlations - different frequencies Scotti Index Correlations 1-month 2-months Quarterly Change in 1y yields S&P 5 excess returns Citi Index Correlations 1-month 2-months Quarterly Change in 1y yields S&P 5 excess returns Go back 1 / 19

74 References Appendix Regression - different frequencies Scotti Index OLS - R 2 1-month 2-months Quarterly Change in 1y yields S&P 5 excess return.2.4 Citi Index OLS - R 2 1-month 2-months Quarterly Change in 1y yields S&P 5 excess return Go back 11 / 19

75 References Appendix News Analysis: Now-Casting news Data description: Histograms I start from Now-Casting news 12 / 19

76 4 IP 4 Cap Ut 4 Invent to sales ratio New Ord 4 Hous St 3 New Private H RDPI RPCE N Ret Sal

77 Trade Bal Nonfarm Empl Empl Tot Priv Ind UoM 3 Bus Inv 3 Unempl Mfg New Ord Hours PCE Prices

78 5 CPI 8 ShortTermRate 4 LongTermRate PPI 4 PPI Fin Goods 5 Money Loans 4 Cons Conf 4 PMI Mfg

79 15 GDP

80 References Appendix News Analysis: Now-Casting news Data description: Histograms Bloomberg news 17 / 19

81 Industrial Production MoM Capacity Utilization Housing Starts Building Permits Trade Balance Change in Nonfarm Payrolls U. of Mich. Sentiment Unemployment Rate CPI MoM

82 3 PPI MoM 25 Consumer Confidence Index 4 ISM Manufacturing GDP Annualized QoQ A