How News and Its Context Drive Risk and Returns Around the World

Transcription

1 How News and Its Context Drive Risk and Returns Around the World Charles Calomiris and Harry Mamaysky Columbia Business School Q Group Spring 2018 Meeting

2 Outline of talk Introduction Data and text measures Observations Event studies Empirical results The role of entropy Out-of-sample testing Conclusion 5/8/2018 Mamaysky 2

3 Introduction Automated processing of natural language allows: Practitioners to monitor information in real time Researchers to systematically study how markets react to information Prior work studies individual US stock reaction to news News all articles about a stock on a given day Finding short-term underreaction to news Horizon stock reaction the next day Explanations bounded rationality or microstructure effects What about macro (country-level) responses to news? What types of news matter? Do news forecast future outcomes, and in what direction? Do price responses occur over the same horizon? Is it bounded rationality/microstructure, or narrative economics? 5/8/2018 Mamaysky 3

4 How to measure country level news? Our proposed news measures: Frequency Sentiment News topics Unusualness We are particularly interested in news topics: What are country level news topics? Does importance of news differ by topic? Do topics differ from EM to DM? How do topics evolve over time? In this paper, we scratch the surface 5/8/2018 Mamaysky 4

5 What are we trying to forecast? Four price-based measures at the country level return return 12 sigma drawdown next month s return next 12-month returns next month s realized volatility next 12-month maximum drawdown Look at countries stock market indexes e.g. DAX, SP500, Bovespa, Shanghai, etc. Returns measured in US$ terms from perspective of a USbased investor Analyze responses separately for emerging market (EM), and developed market (DM) countries 5/8/2018 Mamaysky 5

6 Some findings Over the short term, country stock markets drift into news with some evidence of continuation Causality appears to run in both directions (news returns, returns news) EM is different from DM news yield more incremental information for EM Topics matter effect of sentiment depends on topic Evidence of regime shifts in coefficients Evidence of out-of-sample predictability Our approach compares favorably to a priori approaches, like Baker, Bloom and Davis economic policy uncertainty 5/8/2018 Mamaysky 6

7 More questions Why is positive sentiment good in some topics, but bad in other topics? Are news about EM s different, or are EM s different, or both? How much of predictability is from currency effect? Is the effect underreaction? Is it causal? Now that people are studying this, will it change? 5/8/2018 Mamaysky 7

8 Outline of talk Introduction Data and text measures Observations Event studies Empirical results The role of entropy Out-of-sample testing Conclusion 5/8/2018 Mamaysky 8

9 Data and text measures Data: Thomson-Reuters digital news archive from mm EM and 12mm DM articles 52 countries, 28 DM and 24 EM (country list) Text measures: Article counts Unusualness (entropy) Sentiment Topics 5/8/2018 Mamaysky 9

10 Text measures 1. artcount number of articles per country per month 2. entropy unusualness of an article j (Glasserman and Mamaysky 2016) H j = i {4 grams} p i log m i Average per word log probability E.g. Order imbalance of 10,000 vs Order imbalance of 2,000,000 shares 3. sentiment the difference of positive and negative words divided by total words in article j: s j = POS j NEG j a j Word sentiment comes from Loughran McDonald dictionary (examples) 5/8/2018 Mamaysky 10

11 Text measures (cont d) 4. topics find groups of words that tend to cooccur together in articles Collect 1242 econ words Start w/ 237 words from index of Beim and Calomiris (2001) and find other words, bigrams and trigrams from EM corpus based on cosine similarity E.g. barriers, currency, parliament, macroeconomist, and soybean We calculate 2 document term matrixes: EM: 5mm articles x 1,240 words DM: 12mm articles x 1,242 words Compute cosine similarity matrixes (1,242 x 1,242) Each word/phrase represented by 5mm or 12mm length vector Then do community detection (via modularity maximization) Our topics are mutually exclusive (LDA gives similar answers) 5/8/2018 Mamaysky 11

12 5 Topics for EM Sample articles 5 topics Topic similarity 5/8/2018 Mamaysky 12

13 Sample EM articles 5 Topics for EM 5 Topics for DM 5/8/2018 Mamaysky 13

14 5 Topics for DM Sample articles 5 topics Topic similarity 5/8/2018 Mamaysky 14

15 Sample DM articles 5 Topics for EM 5 Topics for DM 5/8/2018 Mamaysky 15

16 Decomposing article sentiment Now that we have topics, how to classify an article s sentiment into topics? Article sentiment is s j Let f τ,j be the fraction of econ words in article j that are about topic τ Topic sentiment is defined as: s τ,j f τ,j s j Aggregate the article level measures s τ,j into daily measure s c,τ,t i.e. the sentiment in topic τ for country c on day t Weight each article by number of overall words 5/8/2018 Mamaysky 16

17 The country-level text measures For a given country, we have 12 daily text measures: entropy article count fmkt / smkt fgovt / sgovt fcorp / scorp fcomms / scomms fmacro / smacro (EM) fcredit / scredit (DM) EM or DM specific 5/8/2018 Mamaysky 17

18 Outline of talk Introduction Data and text measures Observations Event studies Empirical results The role of entropy Out-of-sample testing Conclusion 5/8/2018 Mamaysky 18

19 Observations EM Sentiment For 140 EM sentiment series (28 countries x 5 topics) we look at first 2 principal components PC2 relative sentiment of Markets to Government Some evidence of a regime shift in PC2 around the financial crisis 5/8/2018 Mamaysky 19

20 More principal components DM Sentiment For 120 DM sentiment series (24 countries x 5 topics) we look at first 2 PCs PC2 relative sentiment of Markets to Government (again!) Some evidence of a regime shift in PC2 a little before the financial crisis 5/8/2018 Mamaysky 20

21 Outline of talk Introduction Data and text measures Observations Event studies Empirical results The role of entropy Out-of-sample testing Conclusion 5/8/2018 Mamaysky 21

22 Event studies How do stock market indexes behave around news? A news event is defined as: Negative country-day sentiment in bottom 10% Positive country-day sentiment in top 10% Neutral sentiment between 45 th and 55 th percentile Look at abnormal performance of country s stock market in 10 days prior to the event, on the event day, and 10 days after the event Abnormal returns relative to either an EM or DM single factor model estimated over the entire sample Our main analysis is at a lower frequency than the event studies Lots of caveats 5/8/2018 Mamaysky 22

23 Event studies EM Observations Stock index prices drift into news (news not exogenous) No drift around neutral news Topics with post event drift: Mkt (both) Comms (negative) Different from single name results, where there is little evidence of drift post negative news (e.g. Tetlock et al., Henderschott et al.) Mechanisms must differ Underreaction to information Microstructure effects 5/8/2018 Mamaysky 23

24 Event studies DM Observations News is more of a surprise in DM s Bigger event-day jumps Some topics show post event drift: Mkt (negative, both?) Corp (positive) Credit (both) 5/8/2018 Mamaysky 24

25 Outline of talk Introduction Data and text measures Observations Event studies Empirical results The role of entropy Out-of-sample testing Conclusion 5/8/2018 Mamaysky 25

26 Empirical results For EM and DM samples, run panel regressions with country fixed effects to forecast t+1 observations of: return return 12 sigma Drawdown For regressions we aggregate data by month (t=month) Control for many variables that have been shown to have forecasting power for future returns The no-text measure regression is our Baseline model All text measures (except entropy) are normalized to unit variance Run regressions in the 1 st and 2 nd half of the sample, as well as for the overall sample EM drawdown residuals DM drawdown residuals 5/8/2018 Mamaysky 26

27 Summary of results News matters for EM and DM! But results differ across EM and DM Baseline R 2 lower for EM % increase in R 2 from text measures larger for EM Sign of effects (i.e. good news or bad) is consistent across return, return 12, sigma, and drawdown Context matters: positive sentiment in Govt, Corp positive sentiment in Mkt > bad news > good news Incremental explanatory power largest for return 12 and drawdown; and lower for return and sigma Evidence of state dependence For example entropy: bad pre-crisis to good post-crisis Results detail 5/8/2018 Mamaysky 27

28 Summary of panels Emerging markets Developed markets Ø = not in regression +/- = significant Std errors clustered by time for return and sigma By both for return 12 and drawdown Entropy coefficient switches signs from bad early to good later 5/8/2018 Mamaysky 28 Summary of results EM panel summary DM panel summary

29 Outline of talk Introduction Data and text measures Observations Event studies Empirical results The role of entropy Out-of-sample testing Conclusion 5/8/2018 Mamaysky 29

30 The role of entropy We find that entropy goes from a bad pre-crisis to a good post-crisis Look at country-days that are in the top 5% by entropy in the full sample, and in the two subsamples Measure the difference in sentiment and frequency by topic in high entropy days vs the full-sample average The numbers are normalized by the full-sample standard deviation 5/8/2018 Mamaysky 30

31 For EM Sentiment Frequency 5/8/2018 Mamaysky 31

32 For DM Sentiment Frequency 5/8/2018 Mamaysky 32

33 Outline of talk Introduction Data and text measures Observations Event studies Empirical results The role of entropy Out-of-sample testing Conclusion 5/8/2018 Mamaysky 33

34 Out-of-sample testing Do we have too many text-based explanatory variables? How to account for regime shifts? Check out-of-sample forecasting performance Run rolling 5-year elastic net regressions Use these to forecast returns and volatilities Compare three models Naïve rolling country fixed effects Baseline include returns, vol, value, credit-gdp, rate CM Baseline plus our text measures (except comms) 5/8/2018 Mamaysky 34

35 Out-of-sample portfolio formation Elastic net specification is: 1 min β 2N 1,t y i,t x i,t 1 β 2 + λ α β 1 + Τ 1 α β Combine lasso (least abs. shrinkage & selection operator) with ridge regression Smoothed series of coefficient estimates Amount of shrinkage given by β 1 / β OLS 1 α = 0.25 and λ is chosen to minimize cross-validation error Use elastic net to come up with time t return and variance forecast Portfolio weight for country i at time t is i w i E t r t+1 r f,t t = i γ var t (r t+1 r f,t ) With γ = 5 and weights are capped at ±1.5 Form equal weighted portfolio of all weights i at time t 5/8/2018 Mamaysky 35

36 Out-of-sample results We evaluate the performance of the three candidate models Risk benchmark is the Brusa, Ramadorai, and Verdelhan factor model There is evidence of statistical and economic outperformance The differences between CM and Base α s are statistically significant 5/8/2018 Mamaysky 36

37 Outline of talk Introduction Data and text measures Observations Event studies Empirical results The role of entropy Out-of-sample testing Conclusion 5/8/2018 Mamaysky 37

38 Conclusion Useful information in text for medium-term countrylevel outcomes! Different dimensions of text matter In particular, context (i.e. topics) Effects differ across EM and DM Effects differ over time Evidence of out-of-sample forecasting ability Next: Currency Other markets 5/8/2018 Mamaysky 38

39 Appendix 5/8/2018 Mamaysky 39

40 Country list Data and text measures 5/8/2018 Mamaysky 40

41 Entropy or unusualness The entropy of an article is H j = i 4 grams in j p i log m i p i is 4-gram i s fraction of all 4-grams in article j m i is the historical conditional probability of observing the 4 th word in i conditional on the first 3 words For month t, we estimate this over time window t- 27,,t-4 H j is a measure of how different the language constructs in article j are relative to a historical sample Text measures 5/8/2018 Mamaysky 41

42 Network modularity Network modularity measures the amount of connectivity in the network above what would be expected by chance Define a network by cosine similarity similarity of i and j D j D i where D i is the i th column of the document term matrix Consider a partition C of this network with k communities Let e be a k k symmetric matrix whose (i,j) th element is equal to the fraction of all edge weights in the network that connect members of communities i and j The modularity of the network is Q = Trace e e 2 2 = ดe ii ดa i i fraction of a i = fraction edges entirely ofall edges ingroup i that involve i where indicates the sum of matrix elements Community detection using LDA yields similar results, though is computationally more demanding D i D j Topics 5/8/2018 Mamaysky 42

43 We find 5 topics The Louvain algorithm returns ~40 word clusters with the following numbers of words Place words from small clusters into big clusters How do topics evolve over time? (topic stability) 5 Topics for EM 5 Topics for DM 5/8/2018 Mamaysky 43

44 Topic stability Given two partitions C and C over a set of N words, let X i,j be the number of common words between cluster i from C and cluster j from C Let m(i) be an injective mapping from C into C, then the overlap between the two clusterings is max m( ) 1 X i,m(i) N i This measures the fraction of the N words that are placed in the same paired categories in C and C If this measure is 100%, it means a perfect match between the two clusterings Time EM 72% 77% 74% 78% 67% DM 70% 79% 80% 84% 78% 5 topics 5/8/2018 Mamaysky 44

45 Topic similarity across EM and DM 5 Topics for EM 5 Topics for DM 5/8/2018 Mamaysky 45

46 Examples of sentiment words NEGATIVE: closed, fears, disrupting, losses, accused UNCERTAIN: almost, approximately, assumed, contingent, believed POSITIVE: leading, gained, strengthened, boosted, prosperity Text measures defined 5/8/2018 Mamaysky 46

47 Baker, Bloom and Davis measure To investigate the role of policy uncertainty, we first develop an index of economic policy uncertainty (EPU) for the United States and examine its evolution since Our index reflects the frequency of articles in 10 leading US newspapers that contain the following triple: economic or economy ; uncertain or uncertainty ; and one or more of congress, deficit, Federal Reserve, legislation, regulation or White House. -- Baker, Bloom and Davis (2015) Overview of results 5/8/2018 Mamaysky 47

48 Event study econometric issues Our return model is estimated over entire sample Evidence of alpha/news correlations in rolling windows News event is not independent of stock prices Event windows overlap Serial correlation within a country Cross-sectional correlation and asynchronicity Cross serial correlations Event studies 5/8/2018 Mamaysky 48

49 Control variables Empirical results 5/8/2018 Mamaysky 49

50 DM panel summary 5/8/2018 Summary of panels Mamaysky 50

51 EM panel summary 5/8/2018 Summary of panels Mamaysky 51

52 DM drawdown residuals Empirical results 5/8/2018 Mamaysky 52

53 EM drawdown residuals Empirical results 5/8/2018 Mamaysky 53

54 Elastic net estimates return 12 Out-of-sample analysis 5/8/2018 Mamaysky 54

55 Elastic net estimates drawdown Out-of-sample analysis 5/8/2018 Mamaysky 55