Survey
* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project
Mapping the State of Financial Stability 14th Annual DNB Research Conference 3rd November 2011 DNB, Amsterdam Peter Sarlin (Åbo Akademi / TUCS) & Tuomas Peltonen (European Central Bank) 1 An example of a SOM output at certain time Mexico ArgentinaTaiwan Malaysia Singapore Turkey Brazil Philippines South Africa India Tranquil Pre crisis Post crisis Crisis UK Indonesia Thailand Japan Canada Czech Republic China Poland Hong Kong Switzerland Norway Denmark Euro area US Korea Sweden Russia Australia New Zealand Hungary 2 1. Introduction - What do we do in the paper? • Create the Self-Organizing Financial Stability Map (SOFSM) • A model that can visualize multidimensional macro-financial vulnerabilities and the state of financial stability across countries and over time • A model that has good out-of-sample predictive capabilities of future systemic events / financial crises 3 2. Self-Organizing Financial Stability Map (SOFSM) • • Building blocks for creating the SOFSM: • Self-Organizing Maps • Identifying systemic events • Vulnerability indicators • Model training • Model evaluation Mapping the State of Financial Stability 4 2.1 Self-Organizing Maps (SOMs) – what are they? • • SOM is an Exploratory Data Analysis (EDA) technique by Kohonen (1981) ➨ Viscovery SOMine It is a clustering and projection technique: – – – • • • Spatially constrained form of k-means clustering Preserves the neighbourhood relations of the data (instead of trying to preserve the distances between data) Projects data onto a grid of nodes (rather than projecting data into a continuous space) Enables visualization of high-D data ➨ 2D grid of nodes without losing the topological relationships of data and sight of individual indicators. Enables a flexible distribution and interactions. Kohonen’s group has continuously reviewed the SOM literature • • The SOM has been used in approx. 10 000 works Applied to currency and debt crises: Arciniegas and Arciniegas Rueda (2009), Resta (2009), Sarlin (2011) and Sarlin and Marghescu (2011) 5 2.1 Self-Organizing Maps (SOMs) – training algorithm xj mb Radius of the neighborhood σ 1. 2. 3. • Compare all data points xj with all nodes mi to find for each data point the nearest node mb (i.e., best-matching unit, BMU) Update each node mi to averages of the attracted data, including data located in a specified neighbourhood σ Repeat steps 1 and 2 a specified number of times. The SOM parameters are radius of the neighbourhood σ, number of nodes M, map format (ratio of X and Y dimensions), and number of training iterations t. 6 2.1 Self-Organizing Maps (SOMs) – interpreting the output Indicator 1 Indicator 2 • This is a 2D map that represents multi-D data with a 2-level clustering • For each indicator, we create a „feature plane“ where the color coding represents the distribution of its values on the 2D map. Indicator 3 Indicator 4 7 2.2 Identifying systemic events and creating financial stability cycle • Use the data set from Lo Duca and Peltonen (2011): 28 countries (18 EMEs & 10 AEs), Quarterly data 1990Q1-2010Q3 • Identification of systemic events: • • The Financial Stress Index (FSI) includes 5 components for each country, measuring volatilities and sharp declines in key market segments (stock, foreign exchange and money markets) A systemic event occurs when the FSI is above the 90th percentile of the country-specific distribution (on average, negative real consequences) Using the FSI, we identify four classes to describe the financial stability cycle: • • • • • Pre-crisis periods (18 months before the systemic event) Crisis periods (systemic events defined by a financial stress index) Post-crisis periods (18 months after the systemic event) Tranquil periods (all other periods) 8 2.3 Vulnerability indicators • 14 indicators of country-level macro-financial vulnerabilities: • • • Domestic = inflation, GDP growth, CA deficit, budget balance, credit growth, leverage, equity price growth, equity valuation Global = inflation, GDP growth, credit growth, leverage, equity price growth, equity valuation Test several transformations of the indicators (over 200 transformations of the indicators tested). • Select best-performing (as a leading indicator) transformations of the variables 9 2.4 Model training Tranquil • Pre crisis Euro area „Static“ model, i.e. model is not re-estimated recursively: • • • Post crisis Crisis • C18 • Pre crisis Tranquil Euro area Post crisis Data as an input to the SOFSM • • • Class variables + vulnerabilities for training Only vulnerabilities for mapping and evaluating Crisis probabilities as an output of the SOFSM • Crisis Training set (estimation sample): 1990Q4 - 2005Q1 Test set (out-of-sample): 2005Q2 - 2009Q2 In the benchmark, we use 18 months as a forecast horizon Account for policymakers’ preferences when evaluating the performance as in Alessi and Detken (2011) (benchmark μ=0.5) Map data onto SOFSM and retrieve a crisis probability 0.01 0.10 0.19 0.28 0.38 0.47 0.56 0.65 10 2.5 Model evaluation • Defining early warning nodes • When calibrating the policymakers’ preferences, we vary the thresholds.This changes the number of “early warning nodes”. µ =0.4 µ =0.5 µ =0.6 11 2.5 Model evaluation • Training the SOM: • While a higher number of nodes M improves in-sample performance, it decreases generalization, i.e. out-of-sample performance. • We increase M and find the first model with Usefulness ≥ 0.25 (logit model). • In terms of “Usefulness”, when µ=0.5, the models are by definition very similar on in-sample data, while the SOM performs better on out-of-sample data • Robustness is tested with respect to three aspects • SOM parameters: radius of neighborhood and number of nodes • Policymakers’ preferences • Forecast horizon Crash periods Precision Recall Tranquil periods Precision Recall Model Data set μ Logit T rain 0.5 SOM T rain 0.5 TP FP TN FN 162 190 190 314 830 706 73 45 0.46 0.38 0.69 0.81 0.92 0.94 Logit SOM 77 112 57 89 249 217 93 58 0.57 0.56 0.45 0.66 0.73 0.79 T est T est 0.5 0.5 Accuracy U AUC 0.81 0.69 0.79 0.71 0.25 0.25 0.81 0.83 0.81 0.71 0.68 0.69 0.13 0.18 0.72 0.75 Reminder: Recall positives = TP/(TP+FN), Recall negatives = TN/(TN+FP), Precision positives = TP/(TP+FP), Precision negatives = TN/(TN+FN), FP rate = FP/(FP+TN), TP rate = TP/(FN+ TP), Accuracy=(TP+TN)/(FN+FP+TN+ TP). 12 3. Mapping the State of Financial Stability – The two dimensional SOFSM • This is the 2D SOFSM that represents multi-D data. • The stages of the financial stability cycle are derived using only the class variables (pre-crisis, crisis, post-crisis and tranquil periods) Tranquil Pre crisis Post crisis Crisis 13 3. Mapping the State of Financial Stability – Constructing the four clusters according to the financial stability cycle C24 C18 Tranquil Pre crisis Tranquil Pre crisis Post crisis 0.50 0.75 0.01 0.22 0.44 Tranquil 0.47 0.70 0.00 • 0.33 Tranquil 0.01 0.02 0.29 Tranquil Post crisis Crisis 0.18 0.35 0.52 0.06 0.25 0.44 0.63 Tranquil Pre crisis Post crisis 0.68 0.32 Pre crisis Tranquil Post crisis Crisis 0.50 0.21 PPC0 Pre crisis Crisis 0.11 Post crisis 0.22 Tranquil 0.00 P18 Crisis 0.11 Post crisis 0.53 Pre crisis T0 Pre crisis 0.35 Tranquil P24 0.32 0.18 P12 Crisis 0.23 0.00 Post crisis Crisis 0.13 0.65 Pre crisis Post crisis Post crisis Crisis P6 Pre crisis Tranquil Pre crisis Crisis C0 0.00 Tranquil Post crisis Crisis 0.26 C6 Pre crisis Post crisis Crisis 0.01 C12 Crisis 0.55 0.82 0.00 0.09 0.18 0.27 Clustering is performed using hierarchical clustering based on class variables. The map is partitioned into four clusters according to the financial stability cycle: a pre-crisis, crisis, post-crisis and tranquil cluster. 14 3. Mapping the State of Financial Stability – The distribution of 14 indicators across the 4 clusters 0.17 0.41 0.29 0.52 0.64 0.76 0.14 0.27 0.41 0.55 0.68 0.80 0.19 0.32 0.46 0.59 0.86 0.08 0.61 0.76 0.25 0.14 0.30 0.60 0.75 0.57 0.73 0.90 0.13 0.28 0.14 0.85 0.18 0.52 0.14 0.83 0.17 0.57 0.71 0.86 0.15 0.31 0.56 0.46 0.61 0.76 0.92 0.11 0.25 0.39 0.52 0.80 Tranquil Pre crisis Post crisis Post crisis 0.18 0.66 Tranquil periods Tranquil 0.06 0.81 Crisis Post-crisis periods 0.70 0.68 Tranquil Crisis Crisis 0.42 0.55 Post crisis Pre crisis 0.28 0.43 Global real equity growth Crisis 0.42 0.30 Pre crisis Tranquil 0.00 0.70 Tranquil Post crisis 0.65 0.57 Global real credit growth Crisis 0.39 0.44 Post crisis Pre crisis 0.27 0.31 Pre crisis Tranquil 0.01 0.71 Crisis periods Post crisis 0.91 0.57 Crisis 0.41 Crisis Tranquil Crisis 0.45 0.43 Post crisis Pre crisis Post crisis 0.91 0.30 Global real GDP growth Pre-crisis periods Crisis 0.46 0.16 Pre crisis Tranquil Crisis 0.31 0.72 Pre crisis Post crisis 0.85 Crisis Tranquil Pre crisis 0.71 Post crisis Global equity valuation Global leverage 0.58 Tranquil Crisis 0.55 0.45 Pre crisis Post crisis Crisis 0.43 0.32 Global inflation Pre crisis Post crisis 0.16 0.18 Tranquil Tranquil Pre crisis 0.31 0.83 Government deficit CA deficit 0.19 0.69 Post crisis Crisis Crisis Crisis Crisis Crisis Pre crisis Post crisis Post crisis Post crisis Post crisis Post crisis Tranquil Tranquil Pre crisis Pre crisis Pre crisis Pre crisis Tranquil Tranquil Tranquil Tranquil Pre crisis Equity valuation Leverage Real equity growth Real credit growth Real GDP growth Inflation 0.29 0.40 0.52 0.63 0.02 0.18 0.34 0.50 0.66 0.82 • Domestic: early signs of crisis - equity growth and valuation, budget deficit, followed by real GDP and credit growth, leverage, budget surplus, and CA deficit. •Global: early signs of crisis - equity growth and level, followed by real GDP growth, while global credit growth and leverage are more concurrent with crises. 15 3. Mapping the State of Financial Stability – Temporal dimension Evolution of macro-financial conditions (all 14 indicators) for the United States and the Euro area (2002-10, first quarter) US 2006 Pre crisis US 2002 2010 US 2007 Euro 2006 Euro 2007 Tranquil US 2008–09 Euro 2008 Crisis Euro 2002 US 2004–05 Euro area aggregate, did not Euro reflect the 2004–05 crisis in GR, IE, PT. Euro 2010 Post crisis US 2003 Euro 2003 Financial Stress Index also decreased for the euro area aggregate Euro 2009 16 3. Mapping the State of Financial Stability – Cross section Visualizing current macro-financial vulnerabilities in key advanced and emerging economies (2010Q3) Mexico ArgentinaTaiwan Malaysia Singapore Turkey Brazil Philippines South Africa India Contagion through similarities in macro-financial vulnerabilities SOFSM enables identifying events surpassing historical experience Tranquil Pre crisis Post crisis Crisis UK Indonesia Thailand Japan Canada Czech Republic China Poland Hong Kong Switzerland Norway Denmark Euro area US Korea Sweden Russia Australia New Zealand Hungary 17 3. Mapping the State of Financial Stability – Regional evolution Evolution of the macro-financial conditions in Emerging Market Economies and Advanced Economies (2002-10 , first quarter) EME 2005 Tranquil Pre-crisis Tranquil AE 2004 Pre crisis EME 2004 AE 2005 Post-crisis EME AE 2006 2006 2007 AE 2007 AE 2008 EME Crisis 2008 Post crisis EME 2010 AE 2010 AE 2002 EME 2002 AE 2009 Crisis EME 2009 AE 2003 EME 2003 18 4. Conclusions • Self-Organizing Financial Stability Map is a useful model for financial stability surveillance: • mapping the state of financial stability and visualizing multidimensional macro-financial vulnerabilities • has good out-of-sample predictive capabilities of future systemic events / financial crises (EWS) • the SOFSM is flexible with respect to, e.g., events of interest, vulnerability indicators, forecast horizons, policymaker‘s preferences 19 Thank you for your attention! 20