Medium-Term Downside Risk: Insights from Textual Analysis of News Charles W. Calomiris and Harry Mamaysky Columbia Business School 1
Introduction • Automated processing of natural language is opening a previously unavailable window into market behavior • It may fundamentally transform finance practice • Prior work has been very short-term focused • But isn’t news (in aggregate) important for longer horizon outcomes? • We look at • Longer term country-level risk and return responses to news • How to measure news at the country level? 2
Our approach and a peak at findings… • We develop a theory-neutral approach to map country news into market outcomes, which measures word flow and examines connections of word flow to risk and return. • We apply this (for the first time, we think) outside the U.S., to 52 countries. • EMs vs. DMs treated separately, given differences in returns processes. Key Findings : 1. Many measures relevant (sentiment, frequency, entropy), EMs/DMs differ. 2. Topical context matters. 3. Results change over time importantly. 4. News generally has opposite implications for return and risk. 5. Drawdown is useful as a measure of risk, especially for EMs. 6. We capture more than a popular a priori measure, in and out of sample . 3
1. Theory-neutral vs. a priori word identifiers What word flow? • Theory-neutral vs. a priori approaches (Baker Bloom Davis 2016) • Theory-neutral does not require advance knowledge of what is important, and avoids data mining risks. • But is it possible to construct a comprehensive, parsimonious, and flexible theory-neutral model of word flow? 4
2. What aspects of news are important? • Sentiment • Frequency • Unusualness (entropy) • Topical context interacted with above • How are topics different from EM to DM? • How does effect of news, and interpretation of news, differ by topic? 5
3. Regime changes over time? • Principal components indicate shift point around Global Crisis • A priori shift point lines up with second principal component • Out of sample properties of forecasting in light of this change 6
4. How to identify topical context? • Identifying topic-relevant words and their characteristics • Louvain method 7
5. Is all news relevant for both returns and risk? • Will we find opposite signs when an effect is statistically significant for return , if it is also statistically significant for sigma or drawdown ? 8
6. How to measure risk? • Especially in EMs, returns are not normal and there is momentum in returns. • In addition to sigma , we use drawdown (which allows longer term effects from momentum, skew, and kurtosis to be expressed). 9
7. How to analyze countries, together or not? • We separate EMs and DMs and analyze each as a panel. 10
8. What news source? • Thomson-Reuters provides a common platform, English language, and large sample of relevant countries, for which there are other data on returns and on various relevant variables. 11
Text measures defined Data • Thomson-Reuters digital news archive from 1996 — 2015 • 5mm EM and 12mm DM articles • 52 countries (list next page) Text measures: • artcount – number of articles per country per month • entropy – “unusualness” of an article j (Glasserman and Mamaysky 2016) 𝐼 𝑘 = − 𝑞 𝑗 log 𝑛 𝑗 𝑗 ∈ {4−grams} • Effectively the average log probability of a word conditional on preceding words • sentiment – the difference of positive and negative words divided by total words in article j : 𝑘 = 𝑄𝑃𝑇 𝑘 − 𝑂𝐹𝐻 𝑘 𝑡 𝑏 𝑘 • Word sentiment comes from Loughran – McDonald dictionary 12
13
Topics Intuition : Find groups of words that co-occur together in articles Details : • 1240 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 World Bank • We have 2 document term matrixes: • 5mm x 1,240 for EM and 12mm x 1,240 for DM • Compute cosine similarity matrix ( 1,240 x 1,240 ) • Then do community detection (using Louvain method for modularity maximization) • Out topics are mutually exclusive (not necessary) 14
We find 5 topics for each group of countries • The Louvain algorithm returns ~40 word clusters with the following numbers of words • Place words from small clusters into big clusters 15
Topics for EMs 16
Topics for DMs 17
Topic similarity across EM and DM 18
Context specific sentiment • Let 𝑔 𝜐,𝑘 be the fraction of econ For a given country, we have 12 daily text measures: words in article j that are about topic τ • entropy • Topic sentiment is given by: • article count 𝑡 𝜐,𝑘 = 𝑔 𝜐,𝑘 × 𝑡 𝑘 • sMkt / fMkt • Aggregate the article level • sGovt / fGovt measures into daily measures (weighted by number of overall • sCorp / fCorp words in an article) • sComms / fComms • DM/EM specific: • sMacro / fMacro (EM) • sCredit / fCredit (DM) 19
Principal Components EM EM Sentiment • For 140 EM sentiment series (28 countries x 5) we look at first 2 principal components • PC2 – relative sentiment of Markets to Government • Some evidence of a regime shift in PC2 a little before the financial crisis 20
Principal components EM DM Sentiment • For 120 DM sentiment series (24 countries x 5) we look at first 2 principal components • PC2 – relative sentiment of Markets to Government (again!) • Some evidence of a regime shift in PC2 a little before the financial crisis 21
Event Studies • High-frequency top and bottom deciles of sentiment • Middle as placebo • Returns lead major sentiment indicators at high frequency • Some post-event drift for positive and negative events 22
Event studies – EM • Cumulative abnormal return around deciles of daily news events • Middle column is control for boring news • Some topics show post event drift: Mkt (both), Comms (negative) • This is very different from single name results, where there is little evidence of drift post negative news (only post positive)! 23
Event studies – DM • Some topics show post event drift: Mkt (negative, both?), Corp (positive), Credit (both) 24
Regression results • We run panel regression with dependent variables given by • return • return 12 • sigma • drawdown • We control for many variables that have been shown to have some forecasting power for future returns (next page) • The no-text measure regression is our Baseline model • All text measures (except entropy ) are normalized to unit variance • We run full sample, 1 st and 2 nd half of the sample 25
Control variables 26
Summary of regression results • News matters for EM and DM! • Results differ across EM and DM (e.g., artcount matters in EMs) • 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) almost always is consistent across return , sigma , and drawdown • Context matters: positive sentiment in Govt , Corp – bad news; positive sentiment in Mkt – good news • Incremental explanatory power largest for return 12 and drawdown ; explanatory power lower for return and sigma • Evidence of state dependence, especially for entropy • Goes from a “bad” pre-crisis to a “good” post-crisis 27
28
29
Out-of-sample testing • Do we have too many explanatory variables? • What about regime shifts? • Check out-of-sample forecasting performance • Run rolling 5-year regressions in t- 60,…,t -1 for forecasting month t outcomes Lasso (least absolute shrinkage and selection operator) 1 2 + 𝜇 𝛾 1 ′ min 2𝑂 𝑧 𝑗,𝑢 − 𝑦 𝑗,𝑢−1 𝛾 𝛾 1,𝑢 • Lasso does shrinkage and model selection • Amount of shrinkage given by 𝛾 1 / 𝛾 𝑃𝑀𝑇 1 30
Rolling lasso for DM drawdown 31
Rolling lasso for EM drawdown 32
Out-of-sample performance • Naïve model forecasts using country fixed effects • Base model includes only the non-text variables • CM includes all text measures • All models estimated using lasso 33
Comparison to Baker, Bloom and Davis • The two types of measures are correlated. • BBD has incremental value over Baseline for three market measures only for DMs. • In the in-sample panel regressions, our measures subsume BBD for explaining return , sigma and drawdown . 34
Out-of-sample comparisons to EPU • EPU counts articles from 10 major papers that contain triplets from uncertainty x economic x {policy terms} • For 5 EM and 11 DM countries where we have EPU data, compare out-of- sample performance of Base vs Base + alternative text measures 35
Conclusions • Useful information in text for medium-term country-level outcomes (returns and cumulative downside risk) • Different dimensions of text matter • In particular, context matters for sentiment • Effects differ across EM and DM, and over time • Evidence of out-of-sample forecasting ability • Next: • Currency effects • Connect to GDP nowcasting (Jungian subconscious?), Fed Beige Book 36
Recommend
More recommend
