Stock market firm-level information and real economic activity

Transcription

1 Stock market firm-level information and real economic activity F. di Mauro, F. Fornari, D. Mannucci Presentation at the EFIGE Associate Partner Meeting Milano, 31 March 2011 March 29, 2011

2 The Great Recession key facts Overall large failure in anticipating negative GDP growth in 2008Q4 and 2009Q1 almost equally split between central banks, academics and market participants. especially serious considering the size and global nature of the decline in activity Going forward Need to re-consider our approach to forecasting?

3 Approaches to predicting economic fluctuations Mainly through aggregate information: macro variables: labor market, money, credit, lagged growth, cpi, confidence indicators...; financial indicators: aggregate stock market returns and variances, slope of the yield curve, credit spreads. Debate on the importance of financial variables in leading to improvements Overall helpful, but best combination of financial indicators (slope, short rate, stock market volatility) not always the best globally Roots of recessions change from episode to episode F. Fornari and A. Mele. Financial volatility and economic activity. Working Paper, LSE FMG. J. Stock and M. W. Watson. Forecasting output and inflation: the role of asset prices. Journal of Economic Literature

4 Can we take a different perspective? How to improve our ability to track business cycles? If recessions are different shouldn t we look at a larger set of predictors or change more often the way we combine predictors? possibly a real time performance criterion can be used to select regressors and weights this is the way forecasting has evolved in meteorology to handle so-called uncertain instabilities

5 Granular Explanation of Aggregate Fluctuations X. Gabaix. The granular origins of aggregate fluctuations Aggregated shock to rate of growth of sales of the 100 largest US firms anticipates the GDP growth in the US 1-quarter ahead. Possible? Sensible? Idiosyncratic shocks to firms should wash out in the aggregate. But this results does not consider the fact that actual firm size distribution has fat tails. Think of the increasing number of multinationals or industrial conglomerates. F. Fornari and A. Mele. Some firms are better than others. firm-level based predictions of aggregate fluctuations. Conditioning upon the equity prices of a set of US firms improves the prediction of the US IP growth

6 Tentative framework for idiosyncratic shocks Importance of idiosyncratic shocks to firms also compatible with Barsky and Sims interpretation of innovation to consumer confidence (CC) as news shocks Shocks to CC in the US found to be orthogonal to current productivity and output but associated with persistent rises in these variables over time Why should shocks to firms equities be/look like news shocks?

7 Firms as filtering devices Consider a firm monitored by some analysts which announces negative developments, can be delay or postponement in the launch of new products we may not even know the reason, maybe lower demand, maybe credit availability gets tighter Analysts revise the price target down, the current equity price drops This price shock is irrelevant in the aggregate but it may be capturing the first signs of a coming macroeconomic or financial shock which eventually will spread through the economy In practice, the firm price maps a set of unobservable shocks A theoretical model to conceptualize this still needed (but see some references in the literature list)

8 Firm-level information Hypothesis we want to test Can real economic activity be better anticipated when one looks at firm level information in addition to aggregate information? Firm-level variables we use Returns of equity prices Variances of equity prices

9 Methodology The main equation we use to forecast changes in IP Δ h ln(ip) t = α + m β 1,j Δ h ln(ip) t f (h) j=1 m m + β 2,j Term t f (h) + β 3,j MktVar t f (h) + j=1 m j=1 γ (i) 1,j Ret(i) j=1 t f (h) + m j=1 γ (i) (i) 2,j Var t f (h) + ε t, h, the forecast horizon, is equal to 6, 12, 18 or 24 months; f (h) = {h + 6, h + 12, h + 18} is the lag structure; ip t is the Industrial Production index; Term t and MktVar t are the term spread and variance of the overall stock market index; Ret (i) t and Var (i) t are the stock market return and variance of the selected companies.

10 Key problem with this regression overlapping data especially significant influence at longer horizons significance controlled with Newey-West, but maybe more care needed

11 Data Selected Economies United States, United Kingdom, Japan, euro area. Aggregate Variables Lagged Industrial Production, Term Spread, Aggregate Variance. Firm level Variables Equity returns and variances of all the companies that have been continuously listed in the respective stock exchanges since Sample Period January 1973 through December Monthly frequency.

12 Sectoral structure of the firms in the sample, by country Sector US UK JP EA Tot 1) Oil & Gas ) Basic Materials ) Industrials ) Consumer Goods ) Health Care ) Consumer Services ) Telecommunications ) Utilities ) Financials ) Technology Total

13 In-sample all firms have something to say of course in sample can be a biased standpoint increases in R-squared relative to aggregate information is for some firms exceptional

14 United States Aggregate Variables Aggregate Variables and Individual Returns Aggregate Variables and Individual Variances Aggregate Variables, Individual Returns and Variances United Kingdom Regression R 2 Regression R Regression R Euro Area 0.5 Regression R Japan

15 Out-of-sample Analysis We estimate model (1) in the paper over ten-year windows; and make prediction of IP growth rate over the subsequent 6, 12, 18, 24 months; Ranking Standing in month t h we rank the firms according to their forecasting performance at the given horizon h exhibited over the previous 6 months, as measured by the RMSE. Ranking is of course irrelevant to produce the IP forecasts It merely represents a criterion to choose one forecast or a subset of forecasts out of a large number of forecasts. G. Amisano and J. Geweke. Optimal prediction pools. ECB Working Paper, No. 1017, 2009.

16 Domestic Results key message: a few firms are much better than random selection, a few firms are much worse 80 United States 80 United Kingdom # of times in first decile # of times in first decile Domestic firms 80 Euro Area Domestic firms 80 Japan # of times in first decile Domestic Firms # of times in first decile Domestic firms

17 Against randomness i) US firms ranking from 1st to 10th monthly data from September 2009 to August 2010 noticeable short term persistence, similar pattern for other countries (see Table 2 in the paper) 1 st 2 nd 3 rd 4 th 5 th 6 th 7 th 8 th 9 th 10 th SUNOCO DOMTAR PREC.CAST APPLIED MATS. NATIONAL ARCHER INTEL ASHLAND RPM INTL. MDU SUNOCO DOMTAR ARCHER PREC.CAST APPLIED MATS. MDU NATIONAL SCHLUMBERGER SPX RPM INTL. SUNOCO DOMTAR MDU ARCHER SCHLUMBERGER APPLIED MATS. NATIONAL JOHNSON PREC.CAST AQUA MDU SUNOCO ARCHER DOMTAR SCHLUMBERGER EXXON EATON AQUA ALCOA NATIONAL MDU SUNOCO ARCHER DOMTAR SCHLUMBERGER EXXON AQUA PERKINELMER ALCOA EATON MDU PERKINELMER SUNOCO HESS AQUA AMERICA JACOBS SPX SCHLUMBERGER DOMTAR ALCOA MDU PERKINELMER HESS MOTOROLA SUNOCO AQUA VALMONT SPX SCHLUMBERGER VULCAN MDU PERKINELMER VALMONT MOTOROLA HESS VULCAN AQUA SUNOCO SPX SCHLUMBERGER MDU VALMONT PERKINELMER VULCAN MOTOROLA RADIOSHACK ASHLAND AQUA HESS HELMERICH VALMONT VULCAN MDU PERKINELMER RADIOSHACK MOTOROLA HELMERICH ASHLAND AQUA HESS VALMONT PERKINELMER VULCAN MDU RADIOSHACK ASHLAND MOTOROLA HELMERICH HESS SKYWORKS MDU VALMONT PERKINELMER ASHLAND HELMERICH RADIOSHACK SKYWORKS SCHLUMBERGER AQUA SPX

18 Against randomness ii) Compute the forecast range in the top and bottom quintiles consider 10 percent and 90 percent values the difference in fit across the two ranges is evident Randomness would make predictions more similar

19 IP forecasts by quintile, United States

20 IP forecasts by quintile, euro area

21 Spillover Do foreign firms matter for domestic real developments? Quick answer: Yes, relative to aggregate information No, relative to domestic firms, nearly as much information content as those When looking at RMSE or Diebold Mariano test note the change in wording: models, not firms A model is a collection of firms ranking at a given place through time

22 Root Mean Squared Error

23 Diebold Mariano Diebold Mariano tests United States euro area domestic firms significance threshold domestic firms significance threshold foreign firms foreign firms firms firms United Kingdom Japan domestic firms significance threshold domestic firms significance threshold foreign firms foreign firms firms firms

24 Country patterns the relative weights of the four countries have recorded large swings over time for a given domestic countries, changes in weights have been at times common across foreign countries but at times some foreign country has tended to gain importance for example Japanese firms were key to anticipate real developments in Japan around end 90s......but irrelevant for recent developments Also for US, the 2001 recession could have been anticipated equally likely by US, UK or euro area firms but standing in Dec 2006 more UK firms than US firms were helpful to anticipate developments in Dec 2007 (see chart)

25 Country patterns

26 Sectoral patterns A small number of firms, domestic and foreign, helps a lot to forecast real developments What makes them special? Do sectors matter? On average, i.e. taking unconditional standpoint, no sector matters more than its weight in the sample looking at each month in turn, situation changes noticeably Some sectors are better to capture recessions rather than expansions it was the case of financial firms during last recession but in the 2001 US recession, mostly Consumer Goods and Consumer Services mattered (see chart)

27 Sectoral patterns

28 Balance sheet items Beyond sectors, do other characteristics of the firm matter in determining their predictive power? For the United States only, data being richer, we look at a number of balance sheet items Then we build a balance-sheet gap variable in the following way: mean of balance sheet item No. i in the top N firms (N=10) MINUS mean of balance sheet item No. i in bottom N firms This gap variables is used in the following regression:

29 Balance sheet items 2 Δ h ln(ip) t = α + + m β 1,j Δ h ln(ip) t f (h) + j=1 m j=1 m β 2,j Term t f (h) j=1 β 3,j MktVar t f (h) + γ i BS (i) t + ε t, Basically this regression allows us to verify whether there are gaps in balance sheet items across firms that can account for the higher predictive power that some firms have (1)

30 Balance sheet items 3 Note that: although balance sheet items are available only yearly, the ranking of the firms according to their predictive power changes potentially in each month and therefore records a noticeable monthly variation. The cross sectional differences in some of the selected items are indeed connected to real developments. For instance, the Inventory to Total Assets ratio, the Foreign Assets to Total Assets ratio and the Assets Turnover are highly significant for the four forecasting horizons considered. As their coefficients are positive, positive gaps in inventories or high assets turnover are associated to future expansions. On the other hand, positive gaps across the two types of firms for balance sheet items as the Quick Ratio (a measure of short term liquidity), Operative Profit Margin and Sales tend to predict declining real growth. BS (i) t

31 Balance Sheet items 1 Size Assets Liabilities Cash Fundamentals 6-month 12-month 18-month 24-month Capitalization Employees Ebitda External Finance Other Investments Assets Asset per Employee Asset Turnover Book Value per Share Liabilities Debt/Capital Interest Expenses/Debt Capital Expenditure/Total Assets Cash/Total Investments Cash per Share Funds from Operations

32 Table: Estimated coefficients from the regression of the rate of growth of US Industrial Production, over h months, on gaps in balance sheet fundamentals across firms. Data are from January 1985 to December denotes significant at 1% level, at 5%, at 10%. Balance sheet items continued Liquidity Global Inventory Sales Fundamentals 6-month 12-month 18-month 24-month Current Ratio Operative Profit Margin Quick Ratio Foreign Assets/Total Assets Foreign Income/ Total Income Foreign Sales/ Total Sales Inventory Inventory Turnover Inventory/Total Assets Net Sales Sales per Share Cost of Goods/Sales

33 Macro - Financial stability In each month, 970 predictions available for IP growth (per country and h) Compute their gaussian kernel density function and take 1 percent and 5 percent values as economic VaR Also scenario analysis can be designed through ranges for predictive quintiles Possibly use RMSE to attach probabilities to ranges Co-VaR easily computable by integrating out firms - empirically - from quintiles (helpful to know the prob of finding firm j in top decile together with firm i, j)

34 5 percent VaR for the four areas 0.1 US UK euro area Japan

35 1 and 5 percent VaR for the United States IP growth rate 0.10 IP VaR 5% VaR 1%

36 Key facts looking at individual firms seems to increase our possibility to anticipate business cycle phases evidence rather homogeneous across four economic areas Foreign firms matter as much as domestic firms and much more than aggregate information Conditionality matters a lot: country and sectoral weights change in relation to the business cycle position Possible use within macro-financial stability analyses worth exploring key balance sheet items of the different firms

37 What next? more testing on the significance of the actual gain in predictive power co-var use covariance information, although tons of series to be analysed

38 THE END - Thanks a lot for your attention