## 1. Sample of Crises Episode

## Source details

**Canonical URL:** [1. Sample of Crises Episode](https://www.imf.org/-/media/files/publications/wp/2019/wpiea2019273-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2019/wpiea2019273-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2019/wpiea2019273-print-pdf.pdf.json)

---

### Introduction and objective
- The paper assesses whether changes in sentiment precede banking crises as defined in Kaminsky and Reinhart (K&R, 1999).
- Main innovation: develop text-based early warning indicators (EWIs) that capture sentiment in financial news using computational linguistics and evaluate their performance as EWIs.
- Scope: construct sentiment indices using a large Financial Times (FT) corpus and assess them as EWIs for crisis episodes identified by K&R for 20 countries between 1970 and 1998.

### Data
- Corpus: over 3,114,080 news articles from the Financial Times (FT), daily, from 1980 to 2019.
- All articles in English covering business, finance, and economics.
- Average coverage:
  - about 6859 news articles per month (corpus-wide average).
  - about 48 articles per month for each of the 20 countries in the sample (country-month average).
- Temporal training selection for word vectors: FT text published between 1980 and 2018.
- Dictionary cutoff: words with a frequency of at least 50 in the FT corpus.

### Methodology: word vectors and semantic clustering
- Vector Space Semantics using the word2vec family.
- Implementation specifics:
  - window size used to compute term co-occurrence: 5
  - vector dimensionality: 200
  - context definition: 2 words preceding and 2 words following every single word in the corpus
  - package used: “genism” in platform Python
- Semantic similarity operationalization: cosine similarity between word-vectors.
- Semantic clusters: linguistically-determined groups of words associated with a specific sentiment or concept; when non-economic words appear, a “sentiment vector” is constructed using only a subset of words corresponding to appropriate (economic and financial) seed words.

### Sentiment indices: seeds, clusters, and aggregation
- Five abstract sentiment classes targeted: “fear”, “risk”, “hedging”, “opinion”, and “crisis”.
- Construction: triangulate the top “n” most similar terms to a set of seed terms for each semantic concept.
  - Fear seed terms: “fear”, “concern”, “afraid”, “anxious”.
  - Risk seed terms: “warn”, “risk”, “threat”, “financial&hazard”, “financial&contagious”, “impact”, “financial&infect”, “terror”, “danger”.
  - Hedging seed terms: “may”, “possibly”, “uncertain”, “maybe”, “can”, “perhaps”, “doubt”, “unsure”.
  - Opinion seed terms: “say”, “feel”, “predict”, “tell”, “believe”, “think”, “suggest”, “decide”, “propose”, “advise”, “hint”, “clue”, “speak”, “announce”.
  - Crisis seed terms: “financial crisis”, “depression”.
- Positive and negative sentiment clusters built using seed lists from Correa et. al. 2018 and FT-trained word vectors.
- Aggregated indices:
  - “All Sentiment (w/o pos. and neg.)”: combines “fear”, “risk”, “hedging”, “opinion”, and “crisis”.
  - “Negative Sentiment +”: adds the negative sentiment index to the five indices above.
  - “All Sentiment”: adds the positive sentiment index to “Negative Sentiment +”.

### Index construction and monthly measurement
- Country-specific article assignment: an article is about a country if the country name appears in the title, abstract, or first paragraph.
- Monthly frequency extraction: for each country and month, extract monthly frequency of each sentiment cluster and normalize by the total number of words present in all articles for that month and country.
- Index formula:
  - I_{...} = ( # [total number of words within a semantic cluster] / # [total number of words for a given country within all articles in a given month] ) * 1000
- EWS trigger rule: for every month compute a backward looking average of 24 months and assess whether the index is above 2 standard deviations from the average. A spike meeting this criterion triggers an EWS.

### Evaluation approach and metrics
- True positives: spikes that precede the start of a crisis (within a 24-months window).
- False positives: spikes not followed by a crisis start within a 24-months window.
- False negatives: crises not preceded by a spike within a 24-months window.
- Signals triggered between crisis start and crisis peak counted neither as true positives nor false positives.
- Prediction metrics:
  - Precision (P) = TP / (TP + FP)
  - Recall (R) = TP / (TP + FN)
  - F2 score used (weighs recall higher than precision per source wording).
- Benchmark: random benchmark average F2-score: 0.39 (constructed by randomly generating 20 seed words per country, repeating 100 times and averaging).

### Box 2 — Dataset and construction highlights
- Word vectors built on 3,114,080 FT articles between 1980 and 2019.
- Seed terms for the five main sentiment classes and predetermined negative/positive lists enable dataset-specific semantic clusters (top 15 semantically related terms retained per seed).
- Sentiment vector correction applied when clusters included non-economic words.
- Normalization and monthly country-level index construction follow the index formula above.

### Aggregate performance statistics (selected, preserve values exactly)
- Reported aggregate precision / recall / F2 (as presented):
  - Positive sentiment: Recall 0.92, Precision 0.26, F2 0.61
  - All sentiment: Recall 0.90, Precision 0.27, F2 0.61
  - Negative sentiment: Recall 0.92, Precision 0.24, F2 0.59
  - Opinion sentiment: Recall 0.92, Precision 0.24, F2 0.59
  - Crisis sentiment: Recall 1.00, Precision 0.21, F2 0.58
  - All sentiment (w/o pos & neg): Recall 0.87, Precision 0.24, F2 0.58
  - Negative sentiment +: Recall 0.86, Precision 0.24, F2 0.57
  - Risk sentiment: Recall 0.92, Precision 0.22, F2 0.56
  - Hedging sentiment: Recall 0.84, Precision 0.22, F2 0.54
  - Fear sentiment: Recall 0.86, Precision 0.19, F2 0.51
  - Average: Recall 0.90, Precision 0.23, F2 0.57
  - Min: Recall 0.84, Precision 0.19, F2 0.51
  - Max: Recall 1.00, Precision 0.27, F2 0.61

### Country-level aggregate F2 highlights (selected values)
- Argentina: Negative 0.52, Positive 0.66, All sentiment (w/o pos & neg) 0.61, Negative + 0.52, All 0.52, Average 0.56, Min 0.52, Max 0.66
- Brazil: Negative 0.69, Positive 0.91, All sentiment (w/o pos & neg) 0.80, Negative + 0.74, All 0.88, Average 0.80, Min 0.69, Max 0.91
- Mexico: Negative 0.83, Positive 0.67, All sentiment (w/o pos & neg) 0.75, Negative + 0.86, All 0.86, Average 0.79, Min 0.67, Max 0.86
- Turkey: Negative 0.86, Positive 0.91, All sentiment (w/o pos & neg) 0.87, Negative + 1.00, All 1.00, Average 0.93, Min 0.86, Max 1.00
- Aggregate (across countries): Negative 0.51, Positive 0.54, All sentiment (w/o pos & neg) 0.49, Negative + 0.50, All 0.54, Average 0.52, Min 0.49, Max 0.54
- Aggregate summary statistics: Average 0.53 (Negative), 0.53 (Positive), 0.47 (All w/o pos & neg), 0.49 (Negative +), 0.53 (All), 0.51 (Average), Min 0.00, Max 1.00, St. Dev. 0.24

### Specific-sentiment F2 highlights (selected values)
- Brazil: Fear 0.83, Risk 0.73, Hedging 0.83, Opinion 0.76, Crisis 0.73, Average 0.78, Min 0.73, Max 0.83
- Mexico: Fear 0.71, Risk 0.77, Hedging 0.75, Opinion 0.77, Crisis 0.88, Average 0.78, Min 0.71, Max 0.88
- Sweden: Fear 0.57, Risk 0.57, Hedging 0.68, Opinion 0.68, Crisis 0.83, Average 0.67, Min 0.57, Max 0.83
- Turkey: Fear 0.74, Risk 0.76, Hedging 0.93, Opinion 0.71, Crisis 0.37, Average 0.70, Min 0.37, Max 0.93
- Aggregate (across countries): Fear 0.43, Risk 0.48, Hedging 0.49, Opinion 0.51, Crisis 0.46, Average 0.47, Min 0.00, Max 0.83

### Main empirical findings (enumerated)
- 1) On average across countries, general sentiment indices (e.g., positive sentiment) perform better than specific sentiment indices (e.g., fear sentiment).
  - At aggregate level, “Positive sentiment” and “All sentiment” perform best in forecasting power (highest F2), followed by negative and opinion sentiments.
  - On average across all countries, EWS in all sentiments precede between 57 (F2 Score) and 90 percent (Recall) of identified crises, and on average 23 percent of the EWS identified are true positives.
  - Among the five specific indices, “crisis”, “opinion”, and “risk” display the highest recall, followed by “fear” and “hedging”.
  - Wider clusters improve the F2 score, though not necessarily the recall.
- 2) For a given country, some sentiment indices perform better than others (example: hedging sentiment for Turkey performs better than crisis sentiment for Turkey).
- 3) Some sentiment indices perform better for some countries than for others (examples and country-level best performers listed in source).
  - Brazil, Mexico, Sweden, and Turkey show higher average forecasting power among specific indicators.
- 4) At individual country level, general sentiment indices do not always outperform specific indices.
  - Five more general indicators show better forecasting power for Brazil, Mexico, and Turkey.
  - Mexico, Finland, and Denmark: “crisis” sentiment is best performer across indices for Mexico; Norway: “risk sentiment” best performer.
- 5) The crisis sentiment index, for most countries, triggered EWS at the right time and would have captured historical crises listed in the source (examples: Brazil 1995 and 1999; Denmark 1987-89; Finland 1991-92; Israel 1983-84; Malaysia 1985 and 1998-99; Mexico 1982-84; Norway 1988-91; Peru 1983; Sweden 1992; Thailand 1984; Turkey 1991).
- 6) Considering a battery of different sentiments increases chances of detecting crises; variation across indices and episodes means multiple indicators or aggregate indices increase detection probability.
- 7) Sentiment indices performed well in recent crisis (or near crisis) episodes.
  - Argentina: “crisis” and “negative” indices would have triggered EWS in an 18 months window prior to the 2018 currency crisis.
  - Brazil: “negative” and other indices triggered EWS ahead of Lava Jato (2014) and Rousseff impeachment (2016).
  - Turkey: “negative” sentiment triggered EWS ahead of the 2016 coup attempt and reached an all-time high in early 2018.
  - Spain: “crisis” index would have triggered six EWS ahead of the Global Financial Crisis (GFC).

### Illustrative case: Argentina — negative sentiment (selected timeline notes)
- After dropping temporarily in Q4 2018, the negative sentiment was 11 percent higher in April 2019 than during its peak before the program request (April 2018).
- The negative sentiment index increased sharply in April (ahead of the IMF Stand-By Arrangement request in May).
- The index increased by almost 30 percent between June and August, ahead of the request for an early release of funds within the IMF SBA and ahead of a 30 percent depreciation in September.
- The index fell at the end of 2018, then increased again between January and April (source text continues beyond provided excerpt).

### Robustness checks and alternative approaches (selected)
- Focus on word/term clusters rather than isolated economic words/terms, because isolated terms generally do not display spikes ahead of crises.
- Broader sentiment indices perform better than predetermined, isolated economic words/terms (reference to Tables 3 and 6).
- Results sensitive to different time windows after an EWS is triggered: longer windows increase probability window includes a crisis and improve the F2 score.
- Weighting cluster words by relative proximity to the seed does not materially change forecasting performance for selected seeds.
- Topic-modeling (machine-driven) selection of topics produces indices with slightly worse performance compared to the sentiment (semantic) indices.

### Selected quantitative robustness and additional results (preserve values)
- Economic term clusters (Table 6) — selected F2 results:
  - Corporate earnings — Recall 0.87 Precision 0.20 F2 Score 0.52
  - Banking sector losses — Recall 0.80 Precision 0.21 F2 Score 0.51
  - Asset impairments — Recall 0.81 Precision 0.20 F2 Score 0.50
  - Nonperforming assets/loans — Recall 0.81 Precision 0.19 F2 Score 0.50
  - Uncertainty — Recall 0.89 Precision 0.17 F2 Score 0.48
  - Balance sheets deterioration — Recall 0.86 Precision 0.16 F2 Score 0.46
- Different time windows — selected F2 scores by country (Table 7, columns 36 m 24 m 12 m 6 m 3 m):
  - Argentina — 0.78 0.64 0.34 0.18 0.09
  - Brazil — 0.84 0.74 0.45 0.19 0.09
  - Mexico — 0.57 0.47 0.28 0.17 0.14
  - Sweden — 0.48 0.43 0.25 0.18 0.09
  - Turkey — 0.41 0.39 0.29 0.18 0.08
  - Average — 0.36 0.30 0.18 0.09 0.05
- Weighting within clusters (Table 8) — average F2 Score (unweighted) by window:
  - 0.52 0.45 0.27 0.15 0.08
  - Average F2 Score (weighted) by window: 0.52 0.45 0.27 0.15 0.08
- Topic modeling (Table 9) — topic-level results (Recall Precision F2):
  - Regulatory — 0.84 0.21 0.53
  - Financial Analysis — 0.79 0.20 0.50
  - Investment — 0.78 0.19 0.49
  - RealEstate — 0.75 0.18 0.46
  - Average — 0.79 0.20 0.49

### Limitations
- Time-varying language: sentiment, word-topic associations, and word/topic frequencies may change over time; indices that rely on time-invariant word sets could miss relevant shifts.
- Country coverage: expanding country coverage may require rethinking relevant events for testing forecasting power, especially for advanced economies with infrequent systemic crises.
- Crisis dating and definition: using alternative sources for crisis dates could expand training samples and change identified word clusters relevant for prediction.
- Other event types: the approach could be applied to non-crisis but large events (e.g., large depreciations and recessions following terms-of-trade shocks).

### Extensions and potential uses
- Using country-specific news sources could improve accuracy and increase precision by capturing changes in cluster composition.
- The sentiment dataset can be used to:
  - potentially improve performance of traditional forecasting models given sentiment spikes ahead of crises or severe stress;
  - examine similarities/differences in cross-country sentiment and subsequent economic outcomes during common global shocks;
  - study the role of sentiment in affecting exchange rate markets, capital flows swings, and whether sentiment in financial news precedes movements in higher-frequency financial variables like sovereign spreads, CDS, or interest rates (example noted: Argentina in early 2018).

### Conclusion — practical takeaways
- Sentiment indices constructed from Financial Times news articles (January 1980 to May 2019) contain useful information and tend to spike ahead of financial crises or severe economic stress episodes.
- Predicting crises remains challenging because crises are rare and data coverage is often scarce, but the new news-based dataset can expand available early warning indicators and be useful for researchers and policy makers for multiple applications.

*Source: IMF Working Paper chapter "1. Sample of Crises Episode" (wpiea2019273-print-pdf).*

### 1. Sample of Crises Episode ............................................................................................

### 1. Sample of Crises Episode

### Introduction and objective
- The paper assesses whether changes in sentiment precede banking crises as defined in Kaminsky and Reinhart (K&R, 1999).
- Main innovation: develop text-based early warning indicators (EWIs) that capture sentiment in financial news using computational linguistics and evaluate their performance as EWIs.
- Paper scope: construct sentiment indices using a large Financial Times (FT) corpus and assess them as EWIs for crisis episodes identified by K&R for 20 countries between 1970 and 1998.

### Data
- Corpus: over 3 million news articles from the Financial Times (FT), daily, from 1980 to 2019.
- All articles in English covering business, finance, and economics.
- Average coverage:
  - about 6859 news articles per month (corpus-wide average).
  - about 48 articles per month for each of the 20 countries in the sample (country-month average).
- Temporal training selection for word vectors: FT text published between 1980 and 2018.
- Dictionary cutoff: words with a frequency of at least 50 in the FT corpus.

### Methodology: word vector representation and semantic clustering
- Use Vector Space Semantics and the word2vec family of vector space models to create word vector representations.
- Implementation specifics:
  - window size used to compute term co-occurrence: 5
  - vector dimensionality: 200
  - context definition: 2 words preceding and 2 words following every single word in the corpus
  - package used: “genism” in platform Python
- Semantic similarity operationalization: cosine similarity between word-vectors.
- Semantic clusters: linguistically-determined groups of words associated with a specific sentiment or concept (e.g., risk, fear); clusters reflect how concepts are used within the FT corpus.
- When non-economic words appear in a cluster, a “sentiment vector” is used: constructed using only a subset of words corresponding to appropriate (economic and financial) seed words.

### Sentiment indices: seed terms and semantic clusters
- Purpose: identify five abstract sentiment classes whose evolution may trigger EWS ahead of crises: “fear”, “risk”, “hedging”, “opinion”, and “crisis”.
- Construction approach: triangulate the top “n” most similar terms to a set of seed terms for each semantic concept of interest.

- ‘Fear’ sentiment
  - Seed terms: “fear”, “concern”, “afraid”, “anxious”.
  - Example top 15 semantically related terms shown for each seed (Table excerpt):
    - fear → worry,concern,anxiety,fret,suspicion,fearful,speculation,hope,doubt,jitter,warn,anger,con_cerns,concems,nervousness
    - concern → worry,con_cerns,concems,anxiety,fear,con_cern,unease,disquiet,nervousness,concem,oncern,scepticism,misgiving,suspicion,doubt
    - afraid → ashamed,fearful,glad,sure,think,aware,averse,pretend,mindful,oblivious,stupid,proud,suppose,hate,terrify
    - anxious → keen,eager,reluctant,try,want,loath,disincline,desperate,hesitant,acutely_sensitive,strive,able,unable,incline,urge
  - Note: some words appear as “partial” or “broken” due to scanned or imperfect digital copies (e.g., con_cerns, concems).

- ‘Risk’ sentiment
  - Seed terms: “warn”, “risk”, “threat”, “financial&hazard”, “financial&contagious”, “impact”, “financial&infect”, “terror”, “danger”.
  - Appendix tables (A1–A4) provide top 15 terms for “risk”, “hedging”, “opinion”, and “crisis” clusters.

- ‘Hedging’ sentiment
  - Seed terms: “may”, “possibly”, “uncertain”, “maybe”, “can”, “perhaps”, “doubt”, “unsure”.

- ‘Opinion’ sentiment
  - Seed terms: “say”, “feel”, “predict”, “tell”, “believe”, “think”, “suggest”, “decide”, “propose”, “advise”, “hint”, “clue”, “speak”, “announce”.

- ‘Crisis’ sentiment
  - Seed terms: “financial crisis”, “depression”.

- Positive and negative sentiment
  - Semantic clusters for positive and negative sentiment constructed using seed lists from Correa et. al. 2018 and the FT-trained word vectors (detailed tables A5 and A6 in Appendix 1).

### Aggregated sentiment indices
- Individual indices combined into broader aggregated indices to potentially increase robustness in detecting anomalies:
  - “All Sentiment (w/o pos. and neg.)”: combines “fear”, “risk”, “hedging”, “opinion”, and “crisis” semantic clusters.
  - “Negative Sentiment +”: adds the negative sentiment index to the five indices above (captures all sentiments considered except the positive sentiment).
  - “All Sentiment”: adds the positive sentiment index to “Negative Sentiment +” to obtain the largest set of words encompassing all individual sentiments.

### Evaluation approach
- Focus: evaluate whether sentiment indices spike or trend up ahead of banking crises (K&R crisis episodes).
- Performance assessment: sentiment indices evaluated as EWIs (innovation relative to prior studies that focus on contemporaneous correlations between sentiment and events).

*Source: IMF Working Paper chapter "1. Sample of Crises Episode" (excerpt).*

### Box 2. Constructing Sentiment Indices based on Semantic Clustering

### Box 2. Constructing Sentiment Indices based on Semantic Clustering

### Methodology: building semantic clusters and indices
- Corpus and embeddings:
  - Word vector representations built using a large corpus of news articles: 3,114,080 articles from the FT between 1980 and 2019.
  - The word vectors form a semantic map to quantify semantic similarity between any two words.
- Seed words and semantic concepts:
  - Seed words that characterize specific sentiment: “fear”, “risk”, “hedging”, “opinion”, “crisis”.
  - In addition, a predetermined set of negative and positive words (following Correa et. al. 2017) are used as seeds to capture negative and positive sentiment in the FT.
  - Seeds enable obtaining a larger set of relevant and semantically similar words specific to the dataset (Appendix 1).
- Country-specific article assignment:
  - An article is defined to be about a country if the name of that country appears in either the title, abstract, or first paragraph.
- Measuring monthly sentiment frequency and normalization:
  - For each country and month, extract the monthly frequency of each sentiment cluster and normalize by the total number of words present in all articles for that month and country.
  - The resulting relative frequency for each word cluster is treated as a proxy for the prevalence of the sentiment it represents over time.
- Index formula (as presented):
  - I_{...} = ( # [total number of words within a semantic cluster] / # [total number of words for a given country within all articles in a given month] ) * 1000
- Sentiment vector correction:
  - When the algorithm associated non-economic words to a seed, a “sentiment vector” was used, constructed by adding vectors corresponding to appropriate (economic and financial) seed words.

### Early warning signal (EWS) definition and evaluation approach
- EWS trigger rule:
  - For every month, compute a backward looking average of 24 months and assess whether the index is above 2 standard deviations from the average. A spike meeting this criterion triggers an EWS.
- Prediction metric definitions:
  - Precision (P) = TP / (TP + FP)
  - Recall (R) = TP / (TP + FN)
  - F2 score used (weighs recall higher than precision per source wording): F2 = (1 + β^2) * (P * R) / ((1 + β^2) * P + R) (source uses F2 where β^2 is set to reflect weighting).
  - True positives: spikes that precede the start of a crisis (within a 24-months window).
  - False positives: spikes not followed by a crisis start within a 24-months window.
  - False negatives: crises identified by K&R not preceded by a spike within a 24-months window.
  - Signals triggered between crisis start and crisis peak were counted neither as true positives nor false positives.
- Sample windowing for evaluation:
  - When computing metrics, the full sample 1980-2018 is not always considered because after 1999 there are no more crises in the K&R sample. Example noted: the 2001 Argentina crisis is not taken into account to compute evaluation metrics.
- Benchmark:
  - Random benchmark constructed by, for each month, randomly generating 20 seed words per country, building a time series 1980–2019, running the same evaluation, repeating 100 times and using the average F2 score.
  - Overall average F2-score for the random benchmark: 0.39.
  - All sentiment indices perform better than the random benchmark; broader sentiment clusters perform better relative to the benchmark.

### Key quantitative performance statistics (aggregate and index-level)
- Reported aggregate precision / recall / F2 (as presented in the source block):
  - Positive sentiment: Recall 0.92, Precision 0.26, F2 0.61
  - All sentiment: Recall 0.90, Precision 0.27, F2 0.61
  - Negative sentiment: Recall 0.92, Precision 0.24, F2 0.59
  - Opinion sentiment: Recall 0.92, Precision 0.24, F2 0.59
  - Crisis sentiment: Recall 1.00, Precision 0.21, F2 0.58
  - All sentiment (w/o pos & neg): Recall 0.87, Precision 0.24, F2 0.58
  - Negative sentiment +: Recall 0.86, Precision 0.24, F2 0.57
  - Risk sentiment: Recall 0.92, Precision 0.22, F2 0.56
  - Hedging sentiment: Recall 0.84, Precision 0.22, F2 0.54
  - Fear sentiment: Recall 0.86, Precision 0.19, F2 0.51
  - Average: Recall 0.90, Precision 0.23, F2 0.57
  - Min: Recall 0.84, Precision 0.19, F2 0.51
  - Max: Recall 1.00, Precision 0.27, F2 0.61
- Aggregate sentiment F2 scores by country (selected aggregate table values as presented):
  - Argentina: Negative 0.52, Positive 0.66, All sentiment (w/o pos & neg) 0.61, Negative + 0.52, All 0.52, Average 0.56, Min 0.52, Max 0.66
  - Brazil: Negative 0.69, Positive 0.91, All sentiment (w/o pos & neg) 0.80, Negative + 0.74, All 0.88, Average 0.80, Min 0.69, Max 0.91
  - Mexico: Negative 0.83, Positive 0.67, All sentiment (w/o pos & neg) 0.75, Negative + 0.86, All 0.86, Average 0.79, Min 0.67, Max 0.86
  - Turkey: Negative 0.86, Positive 0.91, All sentiment (w/o pos & neg) 0.87, Negative + 1.00, All 1.00, Average 0.93, Min 0.86, Max 1.00
  - Aggregate (across countries): Negative 0.51, Positive 0.54, All sentiment (w/o pos & neg) 0.49, Negative + 0.50, All 0.54, Average 0.52, Min 0.49, Max 0.54
  - Aggregate summary statistics: Average 0.53 (Negative), 0.53 (Positive), 0.47 (All w/o pos & neg), 0.49 (Negative +), 0.53 (All), 0.51 (Average), Min 0.00, Max 1.00, St. Dev. 0.24
- Specific-sentiment F2 table highlights (selected values as presented):
  - Brazil: Fear 0.83, Risk 0.73, Hedging 0.83, Opinion 0.76, Crisis 0.73, Average 0.78, Min 0.73, Max 0.83
  - Mexico: Fear 0.71, Risk 0.77, Hedging 0.75, Opinion 0.77, Crisis 0.88, Average 0.78, Min 0.71, Max 0.88
  - Sweden: Fear 0.57, Risk 0.57, Hedging 0.68, Opinion 0.68, Crisis 0.83, Average 0.67, Min 0.57, Max 0.83
  - Turkey: Fear 0.74, Risk 0.76, Hedging 0.93, Opinion 0.71, Crisis 0.37, Average 0.70, Min 0.37, Max 0.93
  - Aggregate (across countries): Fear 0.43, Risk 0.48, Hedging 0.49, Opinion 0.51, Crisis 0.46, Average 0.47, Min 0.00, Max 0.83, St. Dev. values provided per column in source.

### Main empirical findings (7 summarized points from the source)
1. On average across countries, general sentiment indices (e.g., positive sentiment) perform better than specific sentiment indices (e.g., fear sentiment).
   - At aggregate level, “Positive sentiment” and “All sentiment” perform best in forecasting power (highest F2), followed by negative and opinion sentiments.
   - On average across all countries, EWS in all sentiments precede between 57 (F2 Score) and 90 percent (Recall) of identified crises, and on average 23 percent of the EWS identified are true positives.
   - Among the five specific indices (“crisis”, “opinion”, “risk”, “fear”, “hedging”), “crisis”, “opinion”, and “risk” display the highest recall, followed by “fear” and “hedging”.
   - Wider clusters improve the F2 score, though not necessarily the recall; broader sets of words can improve accuracy (precision).
2. For a given country, some sentiment indices perform better than others (example: hedging sentiment for Turkey performs better than crisis sentiment for Turkey).
3. Some sentiment indices perform better for some countries than for others (example: crisis sentiment for Mexico compared to Malaysia).
   - Among specific indicators, Brazil, Mexico, Sweden, and Turkey show higher average forecasting power (high F2).
   - Brazil: “fear” and “hedging” best performers; Mexico and Sweden: “crisis” performs better; Turkey: “hedging” clearly best; Argentina and Bolivia: “opinion” performs well; Finland: “crisis” language performs well.
4. At individual country level, general sentiment indices do not always outperform specific indices.
   - Five more general indicators (positive/negative, All w/o pos & neg, Negative +, All) show better forecasting power for Brazil, Mexico, and Turkey.
   - Mexico, Finland, and Denmark: “crisis” sentiment is the best performer across the 10 indices for Mexico; Norway: “risk sentiment” best performer.
5. The crisis sentiment index, for most countries, triggered EWS at the right time.
   - “Crisis” sentiment would have captured specific historical crises listed in the source (e.g., Brazil 1995 and 1999; Denmark 1987-89; Finland 1991-92; Israel 1983-84; Malaysia 1985 and 1998-99; Mexico 1982-84; Norway 1988-91; Peru 1983; Sweden 1992; Thailand 1984; Turkey 1991).
   - EWS triggered after crisis starts but before crisis peaks were noted as still effective in flagging crises ahead of peaks.
6. Considering a battery of different sentiments increases chances of detecting crises.
   - Sentiment ahead of crises can vary across time, country, macro imbalances, and crisis types; using multiple sentiment indicators or aggregate indices increases detection probability.
   - Appendix 3 shows most indices trigger EWS ahead of severe financial crises (examples: Turkey 1991: hedging triggered 5 strong EWS ahead of crisis start; fear triggered 4; crisis triggered 2).
   - Variation across indices and episodes: Thailand indices triggered ahead of the 1990s peak; Brazil 1994-96: “risk” did not trigger while “crisis” did; Brazil 1999: both flagged several EWS.
7. Sentiment indices performed well in recent crisis (or near crisis) episodes.
   - Argentina: “crisis” and “negative” sentiment indices would have triggered EWS in an 18 months window prior to the 2018 currency crisis.
   - Brazil: “negative” and other sentiment indices triggered EWS ahead of Lava Jato (2014) and Rousseff impeachment (2016).
   - Turkey: “negative” sentiment triggered EWS ahead of the 2016 coup attempt and reached an all-time high in early 2018, ahead of later market pressures.
   - Spain: “crisis” index would have triggered six EWS ahead of the Global Financial Crisis (GFC).

### Illustrative case: Negative sentiment in Argentina (Box 3 highlights)
- The “negative sentiment” index was an effective EWS in Argentina across historical episodes.
- Sharp movements crossing 2 standard deviations above the previous 24 months average are highlighted as good EWS.
- For Argentina, the negative sentiment index would have successfully predicted:
  - 2 financial crises in the 1980s,
  - the severe recession in 1998,
  - the 2001 crisis,
  - the currency crisis in 2018.
- Specific 2018 timeline notes (as presented):
  - The index increased sharply in April, ahead of the IMF Stand-By Arrangement (SBA) request in May.
  - The index increased by almost 30 percent between June and August, ahead of the request for an early release of funds within the IMF SBA and ahead of a 30 percent depreciation in September.
  - After falling at the end of 2018, the index increased again between January and April (source text continues beyond provided excerpt).

*Source: Box 2. Constructing Sentiment Indices based on Semantic Clustering (wpiea2019273-print-pdf).*

### 2019.  After dropping temporarily in Q4 2018, the negative sentiment was 11 percent higher in

### wpiea2019273-print-pdf - 2019.  After dropping temporarily in Q4 2018, the negative sentiment was 11 percent higher in

### Main findings
- After dropping temporarily in Q4 2018, the negative sentiment was 11 percent higher in April 2019 than during its peak before the program request (April 2018).
- The jump in negative sentiment likely reflects an increase in inflation and inflation expectations, and a depreciation of the peso.
- Sentiment levels for most indices are higher after the GFC, potentially reflecting a permanent shift in Financial Times language, persistent concerns about crises and vulnerabilities after the GFC, and secular stagnation concerns after 2010.
- The authors construct 10 new sentiment indices for 20 countries, from January 1980 to May 2019, using Financial Times’ news articles, and find that indices tend to spike and/or trend up ahead of financial crises or severe economic stress episodes (e.g., Brazil 1999, Turkey 2018).

### Robustness checks — overview
- Focus on word/term clusters rather than isolated economic words/terms, because isolated terms generally do not display spikes ahead of crises.
- Broader sentiment indices perform better than predetermined, isolated economic words/terms (reference to Tables 3 and 6).
- Results are sensitive to different time windows after an EWS is triggered: longer windows increase the probability that the window includes a crisis and improve the F2 score.
- Weighting cluster words by relative proximity to the seed does not materially change forecasting performance for selected seeds.
- Topic-modeling (machine-driven) selection of topics produces indices with slightly worse performance compared to the sentiment (semantic) indices.

### Selected quantitative results — economic term clusters (Table 6)
- Corporate earnings — Recall 0.87 Precision 0.20 F2 Score 0.52
- Banking sector losses — Recall 0.80 Precision 0.21 F2 Score 0.51
- Asset impairments — Recall 0.81 Precision 0.20 F2 Score 0.50
- Nonperforming assets/loans — Recall 0.81 Precision 0.19 F2 Score 0.50
- Uncertainty — Recall 0.89 Precision 0.17 F2 Score 0.48
- Negative GDP  growth — Recall 0.85 Precision 0.17 F2 Score 0.48
- State intervention — Recall 0.84 Precision 0.17 F2 Score 0.48
- Exc e s s ive   vola tility — Recall 0.85 Precision 0.17 F2 Score 0.47
- Balance sheets deterioration — Recall 0.86 Precision 0.16 F2 Score 0.46

### Different time windows — selected F2 scores by country (Table 7)
- Columns correspond to windows: 36 m 24 m 12 m 6 m 3 m
- Argentina — 0.78 0.64 0.34 0.18 0.09
- Bolivia — 0.50 0.36 0.18 0.06 0.03
- Brazil — 0.84 0.74 0.45 0.19 0.09
- Chile — 0.00 0.00 0.00 0.00 0.00
- Colombia — 0.10 0.10 0.10 0.10 0.02
- Denmark — 0.49 0.36 0.27 0.15 0.09
- Finland — 0.44 0.40 0.28 0.14 0.04
- Indonesia — 0.44 0.30 0.14 0.06 0.06
- Israel — 0.28 0.28 0.17 0.07 0.05
- Malaysia — 0.43 0.32 0.17 0.08 0.04
- Mexico — 0.57 0.47 0.28 0.17 0.14
- Norway — 0.42 0.33 0.11 0.03 0.02
- Peru — 0.26 0.26 0.26 0.07 0.03
- Philippines — 0.00 0.00 0.00 0.00 0.00
- Spain — 0.00 0.00 0.00 0.00 0.00
- Sweden — 0.48 0.43 0.25 0.18 0.09
- Thailand — 0.24 0.24 0.14 0.06 0.04
- Turkey — 0.41 0.39 0.29 0.18 0.08
- Uruguay — 0.00 0.00 0.00 0.00 0.00
- Venezuela — 0.46 0.35 0.19 0.11 0.07
- Average — 0.36 0.30 0.18 0.09 0.05

### Different weights within clusters (Table 8) — summary
- For a selected set of seeds, using weighted clusters versus unweighted clusters does not materially change results.
- Average F2 Score (unweighted) across seeds by window: 0.52 0.45 0.27 0.15 0.08
- Average F2 Score (weighted) across seeds by window: 0.52 0.45 0.27 0.15 0.08
- Selected seed examples (unweighted F2 scores by window 36 m 24 m 12 m 6 m 3 m):
  - ['fear'] — 0.52 0.45 0.33 0.17 0.10
  - ['worry'] — 0.48 0.40 0.25 0.12 0.08
  - ['concern'] — 0.51 0.42 0.25 0.14 0.07
  - ['uncertain'] — 0.59 0.50 0.33 0.19 0.09
  - ['recession'] — 0.47 0.38 0.23 0.12 0.06
  - ['depression'] — 0.51 0.47 0.34 0.23 0.13
  - ['shock'] — 0.49 0.43 0.28 0.18 0.11

### Topic modeling results (Table 9)
- On average, selected topics’ frequency increases preceded between 49 percent (F2) and 79 percent (Recall) of crises, and about 20 percent of the signals identified were true positives (precision ~0.20).
- Topic-level results (Recall Precision F2):
  - Regulatory — 0.84 0.21 0.53
  - Financial Analysis — 0.79 0.20 0.50
  - Investment — 0.78 0.19 0.49
  - RealEstate — 0.75 0.18 0.46
  - Average — 0.79 0.20 0.49
  - Min — 0.75 0.18 0.46
  - Max — 0.84 0.21 0.53

### Limitations
- Time-varying language: sentiment, word-topic associations, and word/topic frequencies may change over time; indices that rely on time-invariant word sets could miss relevant shifts.
- Country coverage: expanding country coverage may require rethinking relevant events for testing forecasting power, especially for advanced economies with infrequent systemic crises.
- Crisis dating and definition: using alternative sources for crisis dates could expand training samples and change identified word clusters relevant for prediction.
- Other event types: the approach could be applied to non-crisis but large events (e.g., large depreciations and recessions following terms-of-trade shocks).

### Extensions and potential uses
- Using country-specific news sources could improve accuracy and increase precision by capturing changes in cluster composition.
- The sentiment dataset can be used to:
  - Potentially improve performance of traditional forecasting models given sentiment spikes ahead of crises or severe stress.
  - Examine similarities/differences in cross-country sentiment and subsequent economic outcomes during common global shocks.
  - Study the role of sentiment in affecting exchange rate markets, capital flows swings, and whether sentiment in financial news precedes movements in higher-frequency financial variables like sovereign spreads, CDS, or interest rates (e.g., Argentina in early 2018).

### Conclusion — practical takeaways
- Sentiment indices constructed from Financial Times news articles (January 1980 to May 2019) contain useful information and tend to spike ahead of financial crises or severe economic stress episodes.
- Predicting crises remains challenging because crises are rare and data coverage is often scarce, but the new news-based dataset can expand available early warning indicators and be useful for researchers and policy makers for multiple applications.

*Source: https://www.imf.org/-/media/files/publications/wp/2019/wpiea2019273-print-pdf.pdf*

### Appendix 1. Semantic Clusters

### Appendix 1. Semantic Clusters

### Construction method
- Select one or several “seed terms” for each concept. These seed terms are meant to be those most archetypically related to the semantic concepts of interest.
- Measure the (cosine) similarity between the word embeddings representing the seed terms and all other word embeddings trained on the corpus.
- Select the top 15-word embeddings with the highest (cosine) similarity to the seed terms (i.e. the 15 words closest to the seed terms on our “semantic map” of the corpus) and add those to the seed terms in order to form the specific set of words that function as proxy for the semantic concepts of interest.

### Risk language (Table A1)
- Seed Term: warn
  - Top 15 most closely semantically related terms: argue,concede,wamed,acknowledge,admit,say,suggest,predict,insist,complain,warning,conclude,stress,confirm,fear
- Seed Term: r is k
  - Top 15 most closely semantically related terms: reputational_risk,counterparty_risk,danger,systemic_risk,probability,downside_risk,nsks,likelihood,nsk,volatility,moral_hazard,risk_premia,illiquidity,leverage,vulnerability
- Seed Term: threat
  - Top 15 most closely semantically related terms: danger,existential_threat,threat_pose,onslaught,spectre,challenge,obstacle,pressure,vulnerability,provocation,threaten,consequence,hostility,problem,backlash
- Seed Term: financial&hazard
  - Top 15 most closely semantically related terms: flnancial,sinan_cial,finan_cial,financiai,sinancial,tinancial,financia,finacial,inancial,fiancial,s ina n_ c ia l,lina nc ia l,f ina nda l,f ina nia l,f ina noia l
- Seed Term: financial&contagious
  - Top 15 most closely semantically related terms: sinancial,flnancial,finan_cial,systemic,tinancial,sinan_cial,linancial,sinan_cial,inancial,financia,financiai,fnancial,anancial,nancial,self_fulfill
- Seed Term: impact
  - Top 15 most closely semantically related terms: effect,adverse_effect,adverse_impact,mpact,detrimental_impact,consequence,repercussion,spillover_effect,implication,detrimental_effect,ripple_effect,ramification,inflict,damage_inflict,dd
- Seed Term: financial&infect
  - Top 15 most closely semantically related terms: f lna nc ia l,f ina n_ c ia l,s ina nc ia l,s ina n_ c ia l,ina nc ia l,f ina nc ia i,f ina nc ia ,tina nc ia l,s ina n_ c ia l,lina nc ia l,nancial,fnancial,virus,avian_influenza,finandal
- Seed Term: terror
  - Top 15 most closely semantically related terms: te r r or is m,te r r or is t,is la mic _ e xtr e mis m,e xtr e mis m,ba r ba r ity,is la mis t_ te r r or is m,a tr oc ity,is la mis t_ te r r or is t,jiha dis m,viole nc e ,e xtr e mis t,is la mis t_ e xtr e mis m,ba nditr y,jiha d,br uta lity
- Seed Term: danger
  - Top 15 most closely semantically related terms: threat,possibility,likelihood,consequence,risk,spectre,problem,hazard,fear,peril,reality,difficulty,s tigma ,vulne r a bility,ins ta bility

### Hedging language (Table A2)
- Seed Term: may
  - Top 15 most closely semantically related terms: should,can,will,would,must,doe,merely,shall,mav,ought,yet,ould,likely,expedient,.may
- Seed Term: possibly
  - Top 15 most closely semantically related terms: conceivably,perhaps,potentially,easily,probably,even,maybe,almost_certainly,hopefully,dangerous,–,costly,plausibly,either,counterproductive
- Seed Term: uncertain
  - Top 15 most closely semantically related terms: bleak,optimistic,pessimistic,sanguine,precarious,benign,volatile,unsure,unclear,jittery,unpredictable,uncertainty,gloomy,fragile,relatively_benign
- Seed Term: maybe
  - Top 15 most closely semantically related terms: perhaps,anyway,really,probably,surely,think,remember,…,here,crazy,definitely,obviously,forget,lucky,wonder_whether
- Seed Term: can
  - Top 15 most closely semantically related terms: should,will,would,must,may,could,doe,ould,have,often,ought,couldn,want,need,usually
- Seed Term: perhaps
  - Top 15 most closely semantically related terms: maybe,indeed,surely,probably,certainly,even,obviously,arguably,possibly,self_indulgent,presumably,quite,hardly,sometimes,too
- Seed Term: doubt
  - Top 15 most closely semantically related terms: worry,concern,question,sceptical,scepticism,fear,misgiving,suspicion,fret,believe,anxiety,wonder,qualm,widespread_scepticism,think
- Seed Term: unsure
  - Top 15 most closely semantically related terms: unclear,sceptical,undecided,unaware,wary,apprehensive,ignorant,ponder,unconvinced,deeply_sceptical,skeptical,ambivalent,mindful,know,unconcerned

### Opinion language (Table A3)
- (Table A3 present in source but detailed rows not provided in supplied content.)

### Crisis language (Table A4)
- Seed Term: say
  - Top 15 most closely semantically related terms: concede,argue,insist,admit,tell,acknowledge,believe,reckon,warn,conclude,suggest,add,stress,tell_reporter,explain
- Seed Term: feel
  - Top 15 most closely semantically related terms: think,remember,imagine,believe,seem,wonder_whether,forget,realise,know,wonder,sense,grumble,behave,resentful,disorientate
- Seed Term: predict
  - Top 15 most closely semantically related terms: suggest,foresee,reckon,anticipate,believe,argue,warn,confidently_predict,concede,acknowledge,say,speculate,indicate,conclude,forecast
- Seed Term: tell
  - Top 15 most closely semantically related terms: remind,say,explain,ask,complain,admit,recall,convince,assure,acknowledge,observe,believe,inform,reassure,argue
- Seed Term: believe
  - Top 15 most closely semantically related terms: argue,reckon,acknowledge,think,suggest,convince,concede,predict,insist,say,recognise,admit,wonder_whether,conclude,assume
- Seed Term: think
  - Top 15 most closely semantically related terms: believe,feel,wonder_whether,imagine,know,remember,guess,realise,forget,wonder,convince,really,understand,afraid,worry
- Seed Term: suggest
  - Top 15 most closely semantically related terms: argue,indicate,concede,believe,acknowledge,predict,conclude,confirm,imply,reckon,emphasise,say,admit,insist,warn
- Seed Term: decide
  - Top 15 most closely semantically related terms: agree,choose,ask,determine,consider,reconsider,instruct,ponder,persuade,reexamine,refuse,con_sider,discuss,advise,clarify
- Seed Term: propose
  - Top 15 most closely semantically related terms: approve,pro-_pose,proposal,agree,announce,recommend,plan,envisage,outline,oppose,unanimously_approve,draft,consider,abolish,reject
- Seed Term: advise
  - Top 15 most closely semantically related terms: instruct,consult,behalf,recommend,inform,decide,congratulate,rely,insist,ask,tell,oversee,appoint,foist,urge
- Seed Term: hint
  - Top 15 most closely semantically related terms: suggestion,dismiss_suggestion,reject_suggestion,indicate,concede,signal,indication,suggest,acknowledge,warn,express_disappointment,argue,acknowledgment,whiff,acknowledgement
- Seed Term: clue
  - Top 15 most closely semantically related terms: pointer,explanation,reassurance,indication,salutary_lesson,answer,puzzle,inkling,succour,object_lesson,glimpse,cautionary_tale,timely_reminder,pleasant_surprise,consolation
- Seed Term: speak
  - Top 15 most closely semantically related terms: hear,listen,tell,summon,admonish,harangue,upbraid,berate,observe,practise,lecture,shout,pontificate,declaim,recite
- Seed Term: announce
  - Top 15 most closely semantically related terms: unveil,confirm,reveal,finalise,launch,approve,propose,outline,nnounced,warn,conclude,anounced,agree,declare,reaffirm
- Seed Term: f ina nc ia l_ c r is is
  - Top 15 most closely semantically related terms: credit_crunch,turmoil,flnancial,sinancial,finan_cial,meltdown,liquidity_crunch,sinan_cial,upheaval,systemic,turbulence,cri_sis,linancial,recession,cnsis
- Seed Term: depression
  - Top 15 most closely semantically related terms: recession,stagflation,trauma,stagnation,great_depression,malaise,epidemic,reces_sion,illness,psychosis,affliction,privation,deflation,disorder,famine

### Positive sentiment (Table A5)
- Seed Term: a ble
  - Top 15 most closely semantically related terms: unable,oblige,try,need,eager,allow,keen,enable,ability,ready,enough,want,wish,anxious,incline
- Seed Term: enable
  - Top 15 most closely semantically related terms: allow,help,facilitate,require,ensure,able,encourage,assist,complement,empower,need,incentivise,seek,want,use
- Seed Term: adequately
  - Top 15 most closely semantically related terms: properly,inadequately,adequate,sufficiently_robust,appropriately,poorly,fully,sufficiently,prudently,sufficient,rigorously,accurately,correctly,adequacy,susficient
- Seed Term: benign
  - Top 15 most closely semantically related terms: relatively_benign,favourable,favorable,robust,positive,subdue,unfavourable,accommodative,uncertain,unstable,bleak,predictable,perverse,mute,deflationary
- Seed Term: buoyant
  - Top 15 most closely semantically related terms: subdue,sluggish,depress,lacklustre,resilient,weak,robust,strong,stagnant,severely_depress,buoyancy,volatile,upbeat,doldrums,favourable
- Seed Term: buoyancy
  - Top 15 most closely semantically related terms: softness,sluggishness,weakness,resilience,steadiness,firmness,strength,tightness,upswing,upturn,fragility,buoyant,upward_trend,bullish_sentiment,bullish_mood
- Seed Term: comfortable
  - Top 15 most closely semantically related terms: comsortable,happy,uncomfortable,luxurious,uncomsortable,pleasant,agreeable,risk_averse,relax,unsatisfied,comfy,closely_align,attractive,safe,hospitable
- Seed Term: confident
  - Top 15 most closely semantically related terms: consident,hopeful,optimistic,sceptical,cautiously_optimistic,convince,reasonably_confident,believe,adamant,assure,hope,supportive,con-_cerned,unconvinced,doubt
- Seed Term: enhance
  - Top 15 most closely semantically related terms: improve,strengthen,maximise,greatly_enhance,bolster,optimise,maximize,undermine,boost,complement,underpin,reinforce,impair,maintain,facilitate
- Seed Term: favorable
  - Top 15 most closely semantically related terms: unfavorable,favourable,unfavourable,relatively_benign,benign,positive,worrisome,disadvantageous,accommodative,remunerative,reliable_indicator,propitious,adverse,important_determinant,stable
- Seed Term: favourably
  - Top 15 most closely semantically related terms: unfavourably,sceptically,favourable,warily,positively,sympathetically,favorable,cautiously,optimistic,harshly,sanguine,kindly,keenly,toppy,positive
- Seed Term: healthy
  - Top 15 most closely semantically related terms: robust,decent,strong,solid,respectable,stable,resilient,buoyant,unhealthy,sustainable,satisfactory,good,unexciting,positive,steady
- Seed Term: improve
  - Top 15 most closely semantically related terms: enhance,strengthen,deteriorate,boost,mprove,reduce,bolster,increase,mproved,improvement,maintain,optimise,restore,accelerate,maximise
- Seed Term: improvement
  - Top 15 most closely semantically related terms: deterioration,upturn,reduction,acceleration,recovery,mprovement,increase,uplift,improve,rebind,turnaround,decline,uptick,slippage,slight_improvement
- Seed Term: mitiga te
  - Top 15 most closely semantically related terms: minimise,counteract,ameliorate,alleviate,offset,lessen,negate,minimize,neutralise,outweigh,quantify,magnify,avoid,reduce,eliminate
- Seed Term: positive
  - Top 15 most closely semantically related terms: negative,bullish,upbeat,favourable,bearish,downbeat,benign,strong,disappoint,cautious,robust,favorable,gloomy,optimistic,unfavourable
- Seed Term: pos itive ly
  - Top 15 most closely semantically related terms: cautiously,favourably,warmly,sympathetically,profoundly,naturally,thoroughly,somewhat,c le a r ly,pos itive ,mildly,dis tinc tly,s c e ptic a lly,inw a r dly,dif f e r e ntly
- Seed Term: profits
  - Top 15 most closely semantically related terms: proflts,prots,prolts,profis,proits,profi,prolits,prolt,proflt,prfits,profl,profts,prost,proit,prols
- Seed Term: rebound
  - Top 15 most closely semantically related terms: bounce_back,rebound_strongly,recover,tumble,fall,slump,dip,rebind,slip,plummet,retreat,slide,plunge,climb,bounce
- Seed Term: recover
  - Top 15 most closely semantically related terms: bounce_back,rebound,recoup,claw_back,regain,stabilise,suffer,erase,subside,rebound_strongly,recede,fall,slip,evaporate,plummet
- Seed Term: recovery
  - Top 15 most closely semantically related terms: rebind,upturn,upswing,growth,slowdown,downturn,revival,nascent_recovery,turnround,turnaround,improvement,bounceback,cyclical_upswing,recession,ecovery
- Seed Term: r e s ilie nc e
  - Top 15 most closely semantically related terms: strength,robustness,fragility,buoyancy,weakness,vulnerability,dynamism,solidity,steadiness,durability,vitality,softness,toughness,sensitivity,success
- Seed Term: r e s ilie nt
  - Top 15 most closely semantically related terms: robust,buoyant,surprisingly_resilient,remarkably_resilient,strong,stable,sluggish,subdue,healthy,volatile,weak,profitable,benign,vulnerable,solid
- Seed Term: s olid
  - Top 15 most closely semantically related terms: strong,robust,impressive,healthy,stable,solidly,stellar,resilient,respectable,uninspiring,sturdy,unexciting,decent,enviable,buoyant
- Seed Term: sound
  - Top 15 most closely semantically related terms: mawkish,dissonant,tinny,tuneful,plangent,synthesiser,appealingly,bouncy,whimsy,unearthly,astringent,soporific,sublimely,portentous,quirky
- Seed Term: s ta bilis e
  - Top 15 most closely semantically related terms: stabilize,weaken,normalise,restore_confidence,falter,recover,strengthen,regain_momentum,reflate,restore,deteriorate,rebalance,revive,stabilisation,soften
- Seed Term: s ta bilize
  - Top 15 most closely semantically related terms: stabilise,normalise,reflate,stabilization,restore_confidence,regain_momentum,choke_off,readjust,deleverage,rebalance,recalibrate,deflate,weaken,stabilisation,kickstart
- Seed Term: success
  - Top 15 most closely semantically related terms: suc_cess,achievement,uccess,sucess,accomplishment,succes,breakthrough,resilience,prowess,ambition,dynamism,outperformance,progress,survival,strength
- Seed Term: successful
  - Top 15 most closely semantically related terms: succesful,profitable,sucessful,successfu,ambitious,suc_cessful,innovative,uccessful,hugely_successful,excite,prestigious,sophisticate,cessful,successul,impressive
- Seed Term: successfully
  - Top 15 most closely semantically related terms: effectively,aggressively,vigorously,successful,skilfully,satisfactorily,energetically,intensively,robustly,swiftly,efficiently,rapidly,dynamically,profitably,consistently

### Negative sentiment (Table A6)
- Seed Term: abrupt
  - Top 15 most closely semantically related terms: sudden,dramatic,ignominious,untimely,unexpected,abruptly,precipitous,sharp,drastic,catmic,disorderly,hasty,gradual,suddenness,painful
- Seed Term: adverse
  - Top 15 most closely semantically related terms: unfavourable,unfavorable,favourable,negative,currency_fluctuation,unforeseen,adverse worsen,favorable,severe,unforeseeable,deterioration,recessionary,positive,unanticipate
- Seed Term: adversely
  - Top 15 most closely semantically related terms: materially_affect,adversely_affect,adverse,severely_affect,magnify,impair,unfavourab negative,severely_dent,affect,detrimental,severely_curtail,ameliorate,currency_fluctua
- Seed Term: depress
  - Top 15 most closely semantically related terms: aggravate
- Seed Term: aggravate
  - Top 15 most closely semantically related terms: exacerbate,beset,hamper,affect,accentuate,bedevil,adversely_affect,magnify,worsen,c plague,hinder,wrack,ameliorate,severely_affect
- Seed Term: bad
  - Top 15 most closely semantically related terms: dreadful,lousy,terrible,awful,horrible,nasty,wrong,atrocious,rotten,unpleasant,miserable,unfortunate,scary,ugly,stupid
- Seed Term: burden
  - Top 15 most closely semantically related terms: administrative_burden,strain,constraint,pressure,dependency,stigma,dependence,red_ta risk,cost,pain,onus,reliance,downward_pressure,hardship
- Seed Term: challenge
  - Top 15 most closely semantically related terms: dilemma,chal_lenge,obstacle,problem,threat,conundrum,difficulty,daunt_task,ethical_di ,task,headwind,hurdle,uphill_battle,headache,uphill_struggle
- Seed Term: closure
  - Top 15 most closely semantically related terms: mothballing,redundancy,shutdown,lie_offs,rationalisation,layoff,closing,stoppage,compulsory_redundancy,pit_closure,shut_down,demolition,disposal,gartcosh,disruption
- Seed Term: contraction
  - Top 15 most closely semantically related terms: slowdown,deceleration,acceleration,deterioration,stagnation,decline,shrinkage,downturn,recession,upturn,upswing,downward_revision,rebind,retrenchment,upward_revision
- Seed Term: c os tly
  - Top 15 most closely semantically related terms: expensive,cumbersome,painful,problematic,timeconsuming,complicate,disruptive,inefficient,burdensome,dangerous,unnecessary,troublesome,wasteful,onerous,complex
- Seed Term: damage
  - Top 15 most closely semantically related terms: harm,reputational_damage,severely_damage,irreparable_damage,collateral_damage,cripple,tarnish,destroy,devastate,endanger,catastrophic,irreparable,undermine,hurt,corrode
- Seed Term: danger
  - Top 15 most closely semantically related terms: threat,possibility,likelihood,consequence,risk,spectre,problem,hazard,fear,peril,reality,difficulty s tigma ,vulne r a bility,ins ta bility
- Seed Term: de f ic it
  - Top 15 most closely semantically related terms: budget_deficit,surplus,defcit,borrow_requirement,imbalance,shortfall,budget,psbr,external_imbalance,gdp,fiscal,expenditure,unemployment,inflation,outlay
- Seed Term: dent
  - Top 15 most closely semantically related terms: severely_dent,hurt,tarnish,undermine,bolster,boost,erode,weaken,dampen,adversely_affect,severely_damage,crimp,depress,affect,batter
- Seed Term: de s ta bilis e
  - Top 15 most closely semantically related terms: undermine,destabilisation,disrupt,inflame,imperil,neutralise,instability,aggravate,paralyse,cripple,counterproductive,choke_off,unstable,hinder,jeopardise
- Seed Term: deteriorate
  - Top 15 most closely semantically related terms: worsen,deterioration,weaken,improve,sharp_deterioration,rapid_deterioration,depress,shrink,stabilise,stagnate,decline,normalise,weak,dire,diverge
- Seed Term: deterioration
  - Top 15 most closely semantically related terms: sharp_deterioration,rapid_deterioration,improvement,deteriorate,slowdown,weakness,decline shrinkage,erosion,contraction,deceleration,slippage,worsen,upturn,downturn
- Seed Term: dif f ic ult
  - Top 15 most closely semantically related terms: e xtr e me ly_ dif f ic ult,dif f c ult,dif ic ult,dif f i_ c ult,ha r d,tr ic ky,dimc ult,dif c ult,impos s ible ,dif f lc ult,e a s y,virtually_impossible,tough,problematic,notoriously_difficult
- Seed Term: discourage
  - Top 15 most closely semantically related terms: deter,dissuade,encourage,inhibit,prevent,preclude,hinder,stifle,prohibit,restrain,impede,penalise,shy_away,induce,protect
- Seed Term: downgrade
  - Top 15 most closely semantically related terms: rating_downgrade,downgradings,upgrade,lower,rating_agency,junk_status,fitch,default,fitch_rating,moody,underweight,downward_revision,tumble,writedowns,fall
- Seed Term: drag
  - Top 15 most closely semantically related terms: knock,stumble,weigh,push,turn,hobble,limp,pull,weigh_heavily,propel,lurch,rumble,slip,drive,suck
- Seed Term: erode
  - Top 15 most closely semantically related terms: undermine,diminish,weaken,steadily_erode,hurt,whittle_away,severely_dent,exacerbate,curtail,crimp,decimate,reduce,constrain,outweigh,erosion
- Seed Term: erosion
  - Top 15 most closely semantically related terms: deterioration,steady_erosion,diminution,shrinkage,erode,appreciation,degradation,weakness,decline,destruction,weaken,precipitous_drop,intensification,downward_spiral,reassessment
- Seed Term: exacerbate
  - Top 15 most closely semantically related terms: aggravate,cause,accentuate,hamper,magnify,beset,affect,heighten,plague,worsen,overshadow,trigger,undermine,bedevil,hurt
- Seed Term: expose
  - Top 15 most closely semantically related terms: vulnerable,susceptible,prone,lay_bare,confine,exploit,hurt,perceive,ignore,affect,afflict,suffer,confront,cause,induce
- Seed Term: fear
  - Top 15 most closely semantically related terms: worry,concern,anxiety,fret,suspicion,fearful,speculation,hope,doubt,jitter,warn,anger,con_cerns,concems,nervousness
- Seed Term: force
  - Top 15 most closely semantically related terms: threaten,orce,troop,orces,persuade,army,oblige,urge,civilian,reluctant,threat,rearm,unilaterally,yugoslav_army,military
- Seed Term: f r a gility
  - Top 15 most closely semantically related terms: vulnerability,precariousness,resilience,robustness,frailty,sluggishness,strength,solidity,parlous_state,weakness,inadequacy,buoyancy,centrality,cyclicality,interconnectedness
- Seed Term: gloomy
  - Top 15 most closely semantically related terms: downbeat,bleak,pessimistic,upbeat,grim,bullish,rosy,optimistic,bearish,glum,sanguine,sombre,dismal,gloomy_outlook,mistic
- Seed Term: hurt
  - Top 15 most closely semantically related terms: affect,adversely_affect,hamper,exacerbate,undermine,hobble,crimp,erode,hit,aggravate,hinder,harm,cripple,dent,decimate
- Seed Term: illiquid
  - Top 15 most closely semantically related terms: risky,mispriced,highly_leverage,liquid,cdos,illiquidity,opaque,securitised,cdo,synthetic_cdo,volatile,sivs,synthetic_cdos,overpriced,abcp
- Seed Term: impa ir me nt
  - Top 15 most closely semantically related terms: impairment_charge,writedown,write_offs,writedowns,goodwill_impairment,writeoff,writeoffs,unrealise_loss,npls,delinquency,loss,contingent_liability,unrealise_gain,intangible_asset,exceptional_item
- Seed Term: ina bility
  - Top 15 most closely semantically related terms: unwillingness,failure,ability,reluctance,willingness,apparent_inability,determination,desire,refusal,eagerness,attempt,unable,effort,readiness,uncanny_ability
- Seed Term: jeopardise
  - Top 15 most closely semantically related terms: endanger,imperil,undermine,hinder,impede,derail,hamper,fatally_undermine,stymie,severely_hamper,adversely_affect,impair,disrupt,stifle,scupper
- Seed Term: los e
  - Top 15 most closely semantically related terms: regain,shed,slip,win,drop,disappear,earn,climb,gain,recover,fall,miss,claw_back,cede,destroy
- Seed Term: negative
  - Top 15 most closely semantically related terms: positive,bearish,adverse,unfavourable,benign,bullish,weak,favourable,unfavorable,contraction ary,downbeat,volatile,adversely,short_term,gloomy
- Seed Term: pose
  - Top 15 most closely semantically related terms: face,confront,beset,create,exacerbate,solve,aggravate,tackle,sidestep,sacing,overcome,cause,provoke,expose,magnify
- Seed Term: question
  - Top 15 most closely semantically related terms: ques_tions,ques_tion,doubt,thorny_question,dilemma,unanswered_question,query,debate,conundrum,concern,thorny_issue,vex_question,vex_question,inquire
- Seed Term: repercussion
  - Top 15 most closely semantically related terms: ramification,implication,consequence,ripple_effect,impact,reverberation,spillover_effect,adverse_consequence,fallout,adverse_impact,adverse_effect,dire_consequence,collateral_damage,aftershocks,effect
- Seed Term: r is ky
  - Top 15 most closely semantically related terms: illiquid,risk_averse,unattractive,highly_leverage,problematic,distress,safe,opportunistic,highrisk,expensive,volatile,plain_vanilla,toxic,costly,attractive
- Seed Term: severely
  - Top 15 most closely semantically related terms: severe,greatly,badly,irreparably,grievously,hamstring,cripple,severely_affect,drastically,gravely,severely_damage,exacerbate,irretrievably,aggravate,massively
- Seed Term: shortfall
  - Top 15 most closely semantically related terms: surplus,deficit,loss,gap,deficiency,discrepancy,shortage,outflow,deterioration,outlay,cutback,black_hole,imbalance,overspend,downward_revision
- Seed Term: s pir a l
  - Top 15 most closely semantically related terms: downward_spiral,escalate,vicious_cycle,uncontrollable,balloon,deflationary_spiral,creep,uncontrolled,skyrocket,runaway,vicious_spiral,excessive,deflation,hyper_inflation,inflationary_spiral
- Seed Term: squeeze
  - Top 15 most closely semantically related terms: constraint,downward_pressure,downturn,recession,crunch,strain,suck,downward_spiral,drain,upward_pressure,spiral,slump,drag,flatten,intense_competition
- Seed Term: stagnate
  - Top 15 most closely semantically related terms: stagnant,sluggish,decelerate,falter,stagnation,shrink,anaemic,slow,slacken,flatlining,depress,slump,deteriorate,lag_behind,stabilise
- Seed Term: s tr a in
  - Top 15 most closely semantically related terms: severe_strain,pressure,upward_pressure,constraint,downward_pressure,pres_sure,burden,ressure,tension,pres-_sure,squeeze,difficulty,fragility,instability,presure
- Seed Term: stress
  - Top 15 most closely semantically related terms: emphasise,acknowledge,insist,concede,argue,admit,say,suggest,believe,warn,indicate,conclude,reject_suggestion,recognise,acutely_aware
- Seed Term: struggle
  - Top 15 most closely semantically related terms: scramble,try,unable,battle,strive,uphill_struggle,uphill_battle,attempt,eager,fail,fight,difficult,grapple,anxious,flounder
- Seed Term: suffer
  - Top 15 most closely semantically related terms: susfered,susfering,susser,afflict,hurt,hit,recover,benesited,susfers,reel,plague,bounce_back,face,inflict,endure
- Seed Term: threaten
  - Top 15 most closely semantically related terms: imperil,cripple,endanger,disrupt,undermine,severely_damage,jeopardise,paralyse,decimate,destabilise,avert,force,threat,vow,fear
- Seed Term: turbulent
  - Top 15 most closely semantically related terms: tumultuous,torrid,stormy,traumatic,eventful,tempestuous,rocky,miserable,choppy,relative_calm,volatile,febrile,perilous,roller_coaster,tumult
- Seed Term: unable
  - Top 15 most closely semantically related terms: able,unwilling,reluctant,fail,oblige,loath,try,unlikely,refuse,inability,struggle,eager,impossible,anxious,virtually_impossible
- Seed Term: undermine
  - Top 15 most closely semantically related terms: jeopardise,erode,hinder,reinforce,bolster,hamper,imperil,stifle,impede,fatally_undermine,endanger,stymie,destabilise,underpin,exacerbate
- Seed Term: unease
  - Top 15 most closely semantically related terms: disquiet,anxiety,nervousness,unhappiness,uneasiness,wariness,apprehension,dissatisfaction,scepticism,frustration,misgiving,discontent,irritation,anger,concern
- Seed Term: unexpectedly
  - Top 15 most closely semantically related terms: thanexpected,unusually,surprisingly,unexpected,noticeably,exceptionally,after,unprecedentedly,alarmingly,persistently,sharply,firstquarter,month_onmonth,unspectacular,quarterly
- Seed Term: vulnerable
  - Top 15 most closely semantically related terms: susceptible,expose,prone,resistant,heavily_expose,likely,resilient,problematic,sensitive,fragile,immune,unstable,unattractive,unlikely,impervious
- Seed Term: weakness
  - Top 15 most closely semantically related terms: strength,sluggishness,deterioration,softness,slowdown,weaken,buoyancy,malaise,resilience,overvaluation,fragility,underperformance,firmness,weak,turbulence
- Seed Term: worsen
  - Top 15 most closely semantically related terms: deteriorate,aggravate,exacerbate,deterioration,weaken,deepen,dire,depress,rapid_deterioration,sharp_deterioration,severe,heighten,intensify,stabilise,improve
- Seed Term: writedowns
  - Top 15 most closely semantically related terms: writedown,write_offs,impairment_charge,impairment,writeoffs,unrealise_loss,loss,writeoff,markdowns,subprime,redemptions,capital_raisings,rating_downgrade,divestments,goodwill_impairment

*Source: Appendix 1. Semantic Clusters (wpiea2019273-print-pdf)*

### Appendix 2. EWS

### Appendix 2. EWS

### Specific Sentiment — Table A1: aggregate and sample statistics
- Aggregate row (All metrics, Specific Sentiment):
  - Aggregate recall: 0.79
  - Aggregate precision: 0.15
  - Aggregate F2 score: 0.43
  - Aggregate (Risk) recall: 0.83
  - Aggregate (Risk) precision: 0.18
  - Aggregate (Risk) F2 score: 0.48
  - Aggregate (Hedging) recall: 0.78
  - Aggregate (Hedging) precision: 0.19
  - Aggregate (Hedging) F2 score: 0.49
  - Aggregate (Opinion) recall: 0.86
  - Aggregate (Opinion) precision: 0.19
  - Aggregate (Opinion) F2 score: 0.51
  - Aggregate (Crisis) recall: 0.85
  - Aggregate (Crisis) precision: 0.16
  - Aggregate (Crisis) F2 score: 0.46
  - Aggregate (Average) recall: 0.82
  - Aggregate (Average) precision: 0.18
  - Aggregate (Average) F2 score: 0.47
  - Aggregate (min): 0.78
  - Aggregate (max): 0.86
  - Aggregate (overall) recall: 0.15
  - Aggregate (overall) precision: 0.19
  - Aggregate (overall) F2 score: 0.43
  - Aggregate (final metric): 0.51
- Average row (across countries, Specific Sentiment):
  - Average recall: 0.61
  - Average precision: 0.14
  - Average F2 score: 0.34
  - Average (Risk) recall: 0.70
  - Average (Risk) precision: 0.17
  - Average (Risk) F2 score: 0.40
  - Average (Hedging) recall: 0.61
  - Average (Hedging) precision: 0.19
  - Average (Hedging) F2 score: 0.37
  - Average (Opinion) recall: 0.68
  - Average (Opinion) precision: 0.18
  - Average (Opinion) F2 score: 0.41
  - Average (Crisis) recall: 0.71
  - Average (Crisis) precision: 0.16
  - Average (Crisis) F2 score: 0.36
  - Average (Average) recall: 0.66
  - Average (Average) precision: 0.17
  - Average (Average) F2 score: 0.38
  - Average (min): 0.50
  - Average (max): 0.78
  - Average (overall) recall: 0.09
  - Average (overall) precision: 0.27
  - Average (overall) F2 score: 0.23
  - Average (final metric): 0.52
- Min and Max rows (Specific Sentiment):
  - Min: 0.00 (recall), 0.00 (precision), 0.00 (F2 score) for multiple subindices.
  - Max: 1.00 (recall) for multiple subindices; other max values include 0.50, 0.83, 0.71, 0.93.
  - St. Dev. row: 0.48, 0.15, 0.29, 0.44, 0.16, 0.28, 0.47, 0.22, 0.32, 0.46, 0.16, 0.31, 0.43, 0.18, 0.29, 0.39, 0.16, 0.27, 0.48, 0.40, 0.12, 0.22, 0.26, 0.31

### General Sentiment — Table A2: aggregate and sample statistics
- Aggregate row (All metrics, General Sentiment):
  - Aggregate (Negative sentiment) recall: 0.83
  - Aggregate (Negative sentiment) precision: 0.20
  - Aggregate (Negative sentiment) F2 score: 0.51
  - Aggregate (Positive sentiment) recall: 0.83
  - Aggregate (Positive sentiment) precision: 0.22
  - Aggregate (Positive sentiment) F2 score: 0.54
  - Aggregate (Negative sentiment +) recall: 0.78
  - Aggregate (Negative sentiment +) precision: 0.20
  - Aggregate (Negative sentiment +) F2 score: 0.49
  - Aggregate (All sentiment) recall: 0.79
  - Aggregate (All sentiment) precision: 0.21
  - Aggregate (All sentiment) F2 score: 0.50
  - Aggregate (Average) recall: 0.81
  - Aggregate (Average) precision: 0.23
  - Aggregate (Average) F2 score: 0.54
  - Aggregate (min): 0.78
  - Aggregate (max): 0.83
  - Aggregate (overall) recall: 0.20
  - Aggregate (overall) precision: 0.23
  - Aggregate (overall) F2 score: 0.49
  - Aggregate (final metric): 0.54
- Average row (across countries, General Sentiment):
  - Average (Negative) recall: 0.71
  - Average (Negative) precision: 0.19
  - Average (Negative) F2 score: 0.42
  - Average (Positive) recall: 0.67
  - Average (Positive) precision: 0.21
  - Average (Positive) F2 score: 0.42
  - Average (Negative +) recall: 0.56
  - Average (Negative +) precision: 0.19
  - Average (Negative +) F2 score: 0.38
  - Average (All sentiment) recall: 0.61
  - Average (All sentiment) precision: 0.22
  - Average (All sentiment) F2 score: 0.39
  - Average (Average) recall: 0.62
  - Average (Average) precision: 0.24
  - Average (Average) F2 score: 0.42
  - Average (min): 0.63
  - Average (max): 0.73
  - Average (overall) recall: 0.15
  - Average (overall) precision: 0.26
  - Average (overall) F2 score: 0.33
  - Average (final metric): 0.49
- Min and Max rows (General Sentiment):
  - Min: 0.00 across multiple metrics and countries.
  - Max: 1.00 for several recall metrics; other maxima include 0.56 and 0.86.
  - St. Dev. row: 0.43, 0.17, 0.29, 0.46, 0.21, 0.32, 0.48, 0.20, 0.33, 0.47, 0.25, 0.33, 0.48, 0.26, 0.35, 0.43, 0.21, 0.31, 0.48, 0.44, 0.16, 0.26, 0.30, 0.33

### Topic Indices — Table A3: Investment, Regulatory, Financial Analysis, Real Estate
- Aggregate row (All metrics, Topic Indices):
  - Investment recall: 0.78
  - Investment precision: 0.19
  - Investment F2 score: 0.49
  - Regulatory recall: 0.84
  - Regulatory precision: 0.21
  - Regulatory F2 score: 0.53
  - Financial Analysis recall: 0.79
  - Financial Analysis precision: 0.20
  - Financial Analysis F2 score: 0.50
  - Real Estate recall: 0.75
  - Real Estate precision: 0.18
  - Real Estate F2 score: 0.46
  - Average recall: 0.80
  - Average precision: 0.19
  - Average F2 score: 0.48
  - Aggregate (min): 0.75
  - Aggregate (max): 0.84
  - Aggregate (overall) recall: 0.16
  - Aggregate (overall) precision: 0.21
  - Aggregate (overall) F2 score: 0.45
  - Aggregate (final metric): 0.53
- Average row (across countries, Topic Indices):
  - Average recall: 0.60
  - Average precision: 0.19
  - Average F2 score: 0.37
  - Regulatory average recall: 0.66
  - Regulatory average precision: 0.19
  - Regulatory average F2 score: 0.41
  - Financial Analysis average recall: 0.58
  - Financial Analysis average precision: 0.19
  - Financial Analysis average F2 score: 0.36
  - Real Estate average recall: 0.63
  - Real Estate average precision: 0.19
  - Real Estate average F2 score: 0.38
  - Topic Average recall: 0.64
  - Topic Average precision: 0.18
  - Topic Average F2 score: 0.38
  - Average (min): 0.40
  - Average (max): 0.77
  - Average (overall) recall: 0.10
  - Average (overall) precision: 0.29
  - Average (overall) F2 score: 0.23
  - Average (final metric): 0.51
- Min and Max rows (Topic Indices):
  - Min: 0.00 for multiple topic metrics.
  - Max: 1.00 for many recall metrics; other maxima include 0.71, 0.93, 0.87.
  - St. Dev. row: 0.47, 0.22, 0.32, 0.46, 0.18, 0.31, 0.49, 0.24, 0.34, 0.45, 0.24, 0.29, 0.38, 0.18, 0.27, 0.48, 0.40, 0.13, 0.28, 0.28, 0.31

### Country presence and visualization notes
- Country lists repeated in figures and index visualizations include: Argentina, Bolivia, Brazil, Chile, Colombia, Denmark, Finland, Indonesia, Israel, Malaysia, Mexico, Norway, Peru, Philippines, Spain, Sweden, Thailand, Turkey, Uruguay, Venezuela.
- Figures referenced include sentiment time-series and topic-specific signals (for example, "Figure A1. Crises and Signals from the Regulation Topic Index"), spanning years shown in axis labels from 1980 through 1999.

*Source: Appendix 2. EWS (from supplied content).*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2019/wpiea2019273-print-pdf.pdf_