## wpiea2023203-print-pdf

## Source details

**Canonical URL:** [wpiea2023203-print-pdf](https://www.imf.org/-/media/files/publications/wp/2023/english/wpiea2023203-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2023/english/wpiea2023203-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2023/english/wpiea2023203-print-pdf.pdf.json)

---

### 1.1 Literature — methodological foundations and contributions
- Two broad approaches to decomposing demand and supply: Dynamic Stochastic General Equilibrium (DSGE) and SVAR models.
- SVAR / sign-restriction tradition:
  - Supply shocks identified as those with a permanent effect on output (Blanchard and Quah (1989)).
  - Sign restrictions identify supply shocks as those moving prices and quantities with opposite signs, demand shocks with the same sign (Uhlig (2005)).
  - Shapiro (2022) classifies quarterly movements using the sign of forecast errors in single-equation models; price and quantity errors same sign → demand driven; opposite sign → supply driven.
- Advantages of a bottom-up sectoral approach:
  - Prevents conflation of aggregate demand shocks with large sectoral supply shocks (Keynesian Supply shocks; Guerrieri and others (2022)).
  - Granular data point to specific sectors explaining aggregate demand/supply mix.
  - Distinguishes idiosyncratic industry movements from majority-sector movements, important for COVID-19 interpretation.
  - Extends Shapiro (2022) by using a wide range of machine learning models and selecting the optimal model per country to incorporate large information sets, account for non-linearities, and forecast across 218 sectors in 14 countries/regions without identical predictor sets.
- Machine learning in macro forecasting and structural analysis:
  - Literature indicates potential improvements (Coulombe and others (2019)); examples include forecasting inflation, unemployment, and combined unemployment/inflation studies.
  - Growing interest in using high-performance forecasting for structural analysis (Athey and Imbens (2015); Joseph (2019)).
- Keynesian supply shocks and sectoral interactions:
  - Theory (Guerrieri and others (2022)): large supply shocks in few sectors can induce demand-like aggregate responses via complementarities and income effects.
  - Empirical evidence: Cesa-Bianchi and Ferrero (2021); paper finds Keynesian supply shock evidence in a minority of Asia sample countries and for Europe and the USA, but results are fully explained by the COVID-19 event.
  - Policy implication: asymmetry across sectors implies a central bank raising rates to reduce inflation faces a larger sacrifice ratio than when inflation is broad-based demand driven (Guerrieri, Lorenzoni and Straub (2021)).
- Constructing aggregate shocks from industry-level shocks:
  - Industry shocks may correlate with common shocks; solution: granular instrumental variables following Gabaix and Koijen (2021) to purge common shocks and isolate idiosyncratic large-sector movements.
- Evidence on growth spillovers:
  - Literature finds spillovers from a 1 percentage point shock to China: 0.2 pp (Maliszewski and others (2016)); 0.4 pp after 3 years (Furceri and others (2017)); 0.3 pp from a 1 percent of GDP expansion of credit (Barcelona and others (2022)).
  - Paper finds medium-term spillovers of approximately 0.3 pp.
  - One standard deviation for the granular shock: 94 basis points of growth in China and 60 basis points in Europe and the USA.
  - Thus a 1 pp shock in the USA or Europe would imply spillovers close to double those from a 1pp shock in China.
  - Paper finds spillovers peaking at 0.4 pp after 6 quarters for the US and Europe; short-term (within 1 year) spillovers from China limited except to Asia.
- How this paper extends prior work:
  - Uses industry-level price and quantity indices to classify industries and aggregates to country-level decompositions as weighted averages.
  - Distinguishes idiosyncratic vs. majority-representative movements; employs machine learning model selection per country.
  - Applies decompositions to COVID-19 lockdowns and recovery, the 2021–22 inflation surge, Keynesian supply shock incidence, and growth spillovers using granular IVs.

### 2.1 Empirical Models — forecasting framework and classification rule
- Forecast objective and model:
  - ŷ_{i,t+h}_c = f(θ, X_{i,t}_c, X_{i,t−1}_c, ..., X_{i,t−4}_c), where ŷ_{i,t+h}_c forecasts year-on-year change in price or quantity index for industry i in country c.
  - Predictor matrix X_{i,t}_c includes lags of target y_{i,t}_c, values for target variables in industries other than i (x_{-i,t}_c), and global control variables common across countries/industries.
- Aggregate decomposition:
  - Aggregate quantity: y_t = ∑_{i=1}^N w^y_{t,i} y_{i,t}.
  - Aggregate deflator: π_t = ∑_{i=1}^N w^π_{t,i} π_{i,t}.
  - Forecast errors: ε^Q_{i,t} = ŷ_{i,t} − y_{i,t}; ε^P_{i,t} = π̂_{i,t} − π_{i,t}.
  - Classification rule: industry movement is demand if ε^Q_{i,t} ε^P_{i,t} > 0 and supply if ε^Q_{i,t} ε^P_{i,t} < 0.
  - Shares of demand and supply in aggregate GDP or deflator determined using industry weights.
- Models considered (eight total; baseline OLS):
  - Baseline OLS uses only lags of the target variable: X_{i,t}_c = {y_{i,t}_c}.
  - Linear models (3): Lasso (L1), Ridge (L2), Elastic-Net (L1+L2).
  - Non-linear models (4): Support Vector Regression (SVR), Random Forest, Ada-boosted trees, K-nearest-neighbor (KNN).
- Forecasting exercise setup:
  - Training period: 4 years (16 quarters) per country/industry/measure; out-of-sample forecasting begins next quarter.
  - Model selection metric: average root mean squared error (RMSE) RRRMSE_{c,m} computed over industries and price/quantity measures.
  - Forecast horizon for lead/lag event study: h = −3, ..., 3.

### 2.2 Data — coverage and predictors
- Coverage:
  - Real and nominal GDP value-added data by industry for 12 Asian countries, covering 74 percent of GDP in the Asia-Pacific Region.
  - Data source: Haver.
  - Quantity growth: year-on-year percentage change in real GDP value-added for each industry.
  - Deflator: growth rate in nominal GDP less growth rate in corresponding real data.
  - Median number of industries: 16.
  - China required imputations because one industry series not published since Q4 2017 (imputation details in annex I).
- Global predictors:
  - Aggregate measures of GDP growth, investment, unemployment and policy rates in advanced economies;
  - Global share price returns (Dow Jones and MSCI); global industrial production growth;
  - Export and import growth aggregates for the USA, Euro Area and advanced Asia;
  - Global policy uncertainty (GDP weighted measures from Baker, Bloom and Davis (2016));
  - Consumer confidence measures for the US and the EU (Conference Board and European Commission);
  - Global PMI composite, manufacturing and services from JP Morgan;
  - Global liquidity measures covering local and cross-border claims on the nonfinancial sector (Bank for International Settlements).
- Country coverage table (Country — Start Date — End Date — Number of Industries):
  - AUS 12/31/2008 12/31/2022 20
  - BRN 6/30/2016 12/31/2022 24
  - CHN 6/30/1999 12/31/2022 9
  - IDN 6/30/2014 3/31/2023 17
  - IND 9/30/2002 3/31/2023 8
  - KOR 6/30/2006 12/31/2022 15
  - MNG 6/30/2016 12/31/2022 9
  - MYS 6/30/1997 12/31/2022 6
  - PHL 6/30/2006 12/31/2022 16
  - SGP 6/30/1981 12/31/2022 11
  - THA 6/30/2006 12/31/2022 17
  - VNM 6/30/2016 3/31/2023 20

### 2.2.1 GDP Deflator — validation and leading indicator properties
- Aggregation consistency test error metric: ε_t = π_t^{NGDP} − ∑_i w_{i,t}^N π_{i,t}^{VAD}.
- Deflator aggregation errors (Table 2: Country — RMSE — MAE — Correlation*):
  - AUS 0.88 0.63 0.98
  - BRN 2.24 1.48 0.99
  - CHN 0.76 0.40 0.97
  - IDN 0.69 0.37 0.98
  - IND 0.55 0.31 0.98
  - KOR 0.26 0.17 0.99
  - MNG 3.50 2.34 0.95
  - MYS 0.81 0.54 0.99
  - PHL 0.20 0.15 1.00
  - SGP 0.65 0.45 0.99
  - THA 0.18 0.13 1.00
  - VNM 0.66 0.45 1.00
  - *Correlation between π_t^{NGDP} and ∑_i w_{i,t}^N π_{i,t}^{VAD}.
- Interpretation:
  - Errors moderate for most countries except BRN and MNG, where deflator-based inflation is on average 2 and 3 times higher, respectively, than other countries in the sample.
  - Correlations are very high; direction of price changes reliable for decomposition purposes.
- GDP deflator as a leading indicator for CPI:
  - Event study regression: Δy_{i,t+h} = α_i + γ_t + β_h Δπ_{i,t} + ε_{t+h}, h = −3, ..., 3.
  - Findings:
    - GDP-based inflation is a significant leading indicator for rises in CPI inflation in Asia.
    - The deflator increases by around 77 bps in the lead-up to a typical 100 bps increase in CPI inflation.
    - Signal positive and statistically significant in the 3 quarters preceding the CPI increase.
    - Adding controls reduces the GDP deflator increase in the quarters leading up to a CPI rise by around half on average, but effect remains significant and positive for all cases except when adding lags of wage inflation (wage data sparse and roughly halve the country sample).

### 2.3 Optimal Forecasting Models — performance and country optima
- Optimal model selection: m* = argmin_m RRRMSE_{c,m}.
- Empirical summary:
  - Tree-based models (Random Forest and Ada-boosted Trees) perform best on average.
  - Baseline OLS competitive and optimal in several cases: China, India, Malaysia, Singapore.
  - Random Forest and Ada-boost optimal for 4 countries (same count as OLS optimal cases) but with lower average RMSE where optimal.
  - Conclusion: OLS not optimal in majority, but using OLS unlikely to lead to very large differences in results.
- Table 3: Optimal Model by Country:
  - AUS Ada-boosted Trees
  - BRN Ada-boosted Trees
  - CHN OLS
  - IDN Random Forest
  - IND OLS
  - KOR Ada-boosted Trees
  - MNG KNN Regression
  - MYS OLS
  - PHL Random Forest
  - SGP OLS
  - THA Random Forest
  - VNM Ada-boosted Trees

### 3. Results: GDP and Inflation Decompositions — main findings
- Classification uncertainty:
  - Forecast errors near zero treated as unknown using parameter 훼훼.
  - Baseline 훼훼 = 0.1 implies 20 percent of forecast errors near zero are unclassified.
  - Alternative 훼훼 in annex (robustness) uses 훼훼 = 0.2.
- Aggregate GDP decomposition (periods: pre-COVID four years, 2020 lockdowns, post-COVID 2021–22):
  - Supply accounted for a significant share of growth prior to COVID-19; pronounced in Indonesia, Korea and the Philippines.
  - Major GDP drop during lockdowns 2020Q2: majority of decline attributed to demand in several countries.
    - Demand-driven drops during lockdowns observed in: Indonesia, India, Malaysia, the Philippines, Singapore, and Thailand.
  - Recovery late 2020 / early 2021 predominantly demand-driven; supply important in Australia, Malaysia and especially Korea (primarily manufacturing-driven).
- Inflation decomposition:
  - Mirrors GDP decomposition: supply more prominent in 2021 and demand picking up in 2022.
  - Supply-chain disruptions following lockdowns contributed substantially to price rises in 2021 (supply-driven inflation).
  - Demand components remain significant, indicating demand-supply mismatches and demand-pull elements in 2021–22.
  - Country-specific patterns:
    - Indonesia and Malaysia: much of demand-pull inflation driven by primary commodity price movements (internationally determined).
    - Philippines and Vietnam: broad-based price increases across sectors, more likely domestic demand-pull.
- Measurement and treatment of idiosyncratic industry shocks:
  - Idiosyncratic component measured via cross-sectional outliers relative to quarter median.
  - Industry-level components normalized to mean 0 and standard deviation 1 using pre-COVID sample values.
  - Outliers: values 4 standard deviations above cross-sectional median are reset to pre-COVID mean for that industry.
  - Result: much commodity-price-driven inflation in Indonesia during 2021/22 classified as idiosyncratic; Philippines and Vietnam show less idiosyncratic influence.
  - Idiosyncratic inflation less likely to signal overheating for central banks unless strong downstream transmission occurs.
- Selected country patterns:
  - Korea: GDP growth heavily supply-driven (manufacturing); recovery driven by manufacturing; semi-conductor price declines in 2022 impacted growth.
  - Brunei: idiosyncratic shocks from mining and LNG; supply movements primary for growth; large idiosyncratic demand shocks in 2020 from LNG manufacture.
  - Indonesia: apart from 2020Q2 lockdown, demand-driven GDP components span Manufacturing; Wholesale and Retail trade; Transportation and Warehousing.
  - India: pandemic slump and recovery primarily demand-driven (manufacturing, transport/hospitality); recent outturns show more supply components excluding 2022Q2.
- China timing and drivers:
  - Chinese GDP fell sharply in 2020Q1 due to Hubei lockdowns (January–April 2020) causing severe supply shocks.
  - Recovery: 2020Q2 rebound in supply across most sectors except Hotel and Catering; 2020Q3–2021Q1 demand-driven recovery (manufacturing); 2022Q2 onward growth primarily supply-driven in manufacturing.
  - Chinese CPI: mild deflation late 2020 / early 2021; around 2 percent in the last decade; very recent CPI data show CPI inflation touching zero in 2023Q2 (note in source).
  - Chinese GDP deflator: rose sharply in early 2021 based on supply shocks; decelerated from a peak of over 6 percent to zero in 2022Q4 despite stable GDP growth; movements driven by manufacturing/industry.
- USA and Europe comparison:
  - USA: significant role for supply shocks in 2021/22 inflation increase; half of initial inflation increase in 2021Q2 explained by idiosyncratic shocks; supply shocks across sectors accounted for 40 percent of inflation in 2022; idiosyncratic shocks significant in 2020Q2 lockdowns and recovery, fading in late 2022.
  - Europe: almost half of run-up in inflation explained by idiosyncratic demand shocks concentrated in manufacturing and Trade/Travel/Accommodation/Food from 2022Q1; supply role grows from 2022Q3 onward dominated by Industry excluding construction (includes electricity and gas production), likely linked to European gas price rise in 2022.
  - Policy implication: supply or idiosyncratic demand shock dominance makes monetary tightening costlier in output terms than when demand shocks dominate.

### Application I — Keynesian Supply Shocks: identification and country evidence
- Theoretical setup: Guerrieri and others (2022) — in a two-sector model with nominal wage rigidity, large negative supply shocks in one sector can reduce income and aggregate demand enough that aggregate demand declines more than aggregate supply depending on substitutability and MPC.
- Empirical approach:
  - Use industry-level price and quantity shocks classified as demand or supply.
  - Local projections estimated country-by-country with specification preserved from source; β_h measures cumulative change in demand shocks following a supply disturbance.
  - Negative β_h → Keynesian supply shocks (negative supply → negative aggregate demand); positive β_h → sectoral substitutability (demand reallocation).
- Findings:
  - Evidence of Keynesian supply mechanism (negative subsequent aggregate demand) found for Australia, Indonesia, India, and to a lesser degree, Singapore and Vietnam.
  - Remaining 7 Asian countries show limited evidence of declines in aggregate demand following sectoral supply shocks.
  - COVID-19 period important: pre-2020 average responses show limited Keynesian supply effects, indicating that very asymmetric and large supply shocks (e.g., COVID-19 lockdowns) likely needed to generate Keynesian aggregate demand effects.
  - USA and Europe: evidence of Keynesian supply mechanism present and stronger in the USA; dependent on inclusion of COVID-19 period — excluding COVID-19, negative correlation not evident.

### Application II — Growth Spillovers: granular IV methodology and spillover estimates
- Construction of granular country shocks:
  - Industry-level quantity forecast error decomposed: ε_{i,t}^Q = θ_i μ_t + z_{i,t}.
  - Aggregates: size-weighted ε_t^w = ∑_{i=1}^N w_{i,t} ε_{i,t}^Q; equally-weighted ε_t^E = (1/N) ∑_{i=1}^N ε_{i,t}^Q.
  - Granular residual: z_t = ε_t^w − ε_t^E.
  - Use z_t as instrument for aggregate quantity shocks ε_t.
  - First-stage: post-lasso (Chernozhukov and others (2015)); predicted ẑ_t used in IV local projections.
  - Joint F statistics (post-lasso OLS): 9.5 (USA), 14 (EUR), 11.8 (China).
- IV local projections panel specification (horizons h = 0...6):
  - y_{i,t+h} − y_{i,t−1} = α_i + ρ t + β_h ẑ_t + γ z_{t−1} + κ κ y_{i,t−1} + Γ' X_{i,t} + v_{i,t+h}.
  - Controls X_{i,t}: financial conditions indices, Chicago Board Options Exchange Volatility Index, two lags of world export-weighted GDP growth.
- Samples and timing:
  - China and Europe sample start: 2006Q2.
  - USA sample start: 2011Q2.
  - Sample end: 2019Q4 (to avoid COVID-19 lockdown shocks).
  - Impulse-response horizon: h = 0...6 quarters.
  - One standard deviation of z_t: 94 bps (China); 60 bps (EUR); 60 bps (USA).
- Empirical spillover findings:
  - China:
    - Limited short-term global spillovers from China idiosyncratic shocks.
    - Medium-run spillovers approximately 0.3 percent of GDP.
    - Asia-only: short-term spillover approximately 0.4 percent (short lived and fully offset by end of first year); medium-term spillovers approximately 0.3 percent.
  - Europe:
    - Global growth spillovers slightly larger than China's peak at 0.4 percent of GDP.
    - Spillovers more persistent and larger in the short term.
    - Spillovers to Asia broadly similar magnitude to rest of world but more short-term; approximately zero of 1.5 years .
  - United States:
    - IV estimates indicate a significant drop in global (ex-USA) growth in the short term, peaking around 0.4 percent after 6 quarters.
    - Impact on Asia more moderate at around 0.2 percent.
    - Caveats: USA z_t sample begins in 2011, covering a comparatively stable period; shocks exclude important financial channels (e.g., US monetary policy), likely moderating estimated effects on Asia.
- Robustness checks:
  - Two alternative IV local projection specifications:
    - (i) additional lags of the target variable.
    - (ii) lags of the shock variables and trend growth in the country.
  - Conclusions broadly unchanged; specification (ii) provides stronger evidence for near-term spillovers from China and the USA and similar medium-term conclusions; for Europe, (ii) suggests spillovers may build more slowly, peak slightly larger, and potentially be weaker in the medium term.
- Key statistics and sample details:
  - Joint F statistics (post-lasso OLS): 9.5 (USA); 14 (EUR); 11.8 (China).
  - One standard deviation of z_t: 94 bps (China); 60 bps (EUR); 60 bps (USA).
  - Sample start dates: China and Europe 2006Q2; USA 2011Q2.
  - Sample end date: 2019Q4.
  - Impulse-response horizon: h = 0...6 quarters.

### Annexes and robustness
- Annex IV: GDP and Inflation Decompositions by country (figures referenced: Figure A2, Figure A3; images not transcribed).
- Asia aggregate decompositions (Figure A4) notes:
  - "Deocmpositions for Asia based on GDP weighted average of individual country results."
  - "Uncertinaty parameter 훼훼=0.1 so that 20 percent of observations are unclassified."
  - "Asia sample exlcudes BRN and VNM to allow for longer sample."
- Sensitivity to higher model uncertainty — Annex V uses 훼훼 = 0.2 with figures:
  - "Figure A5: GDP Decompositions under 휶휶=ퟎퟎ.ퟐퟐ".
  - "Figure A6: Inflation Decompositions under 휶휶=ퟎퟎ.ퟐퟐ".
- Figures and panels span IMF WORKING PAPERS Monitoring Demand and Supply in Asia; page markers 37 through 45 for Annex IV and V figures.

*Source: wpiea2023203-print-pdf — "1.1 Literature", "2.1 Empirical Models", "2.2 Data", "2.2.1 GDP Deflator", "2.3 Optimal Forecasting Models", "3. Results: GDP and Inflation Decompositions", "5. Application II: Growth Spillovers", Annex IV and Annex V — IMF Working Paper.*

### 1.1 Literature .........................................................................................................

### 1.1 Literature

### Approaches to decomposing demand and supply
- Two broad approaches: Dynamic Stochastic General Equilibrium (DSGE) and SVAR models.
- DSGE:
  - Represent internally consistent representations of economic theory for policy experiments.
  - Imposes a large set of modelling assumptions that may not be well suited to novel shocks (e.g. COVID-19).
- SVAR and sign-restriction tradition:
  - Blanchard and Quah (1989): identify supply shocks as those with a permanent effect on output; temporary disturbances attributed to demand.
  - Sign restrictions (popularized by Uhlig (2005)): identify supply shocks as those that cause movements in prices and quantities with opposite signs, and demand shocks as those with the same sign.
  - Shapiro (2022) methodology: uses the sign of forecast errors in single-equation models to classify quarterly movements in output or prices as supply driven (price and quantity errors move in the opposite direction) or demand driven (errors move in the same direction). This approach is amenable to large sets of equations without cross-equation identifying restrictions.

### Advantages of a bottom-up sectoral approach (as pursued in the paper)
- Identification at the industry level prevents conflation of aggregate demand shocks with large sectoral supply shocks that induce demand-like responses in aggregate prices and quantities (Keynesian Supply shocks; Guerrieri and others (2022)).
- Granular data provide a narrative for the aggregate decomposition by pointing to specific sectors that explain the mix of demand or supply.
- Distinguishing idiosyncratic movements in demand or supply from movements representative of a majority of sectors is important for interpreting shocks around COVID-19.
- Extending Shapiro (2022) by considering a wide range of machine learning models and selecting the optimal model for each country allows:
  - Incorporation of a large information set when informative priors are difficult.
  - Accounting for non-linearities.
  - Efficient forecasting across 218 sectors in 14 countries/regions without applying identical predictor sets to all countries.

### Machine learning in macro forecasting and structural analysis
- A nascent literature shows potential improvements in macro forecasting using machine learning (Coulombe and others (2019) review).
- Examples of forecasting applications:
  - Inflation: Joseph and others (2021).
  - Unemployment: Smalter and Cook (2017).
  - Unemployment and inflation: Sermpinis et al. (2014).
- Comparative analyses of techniques: Chen (2019), Gianonne and others (2021), Kalamara and others (2022).
- Growing interest in leveraging high-performance forecasting to inform structural analysis: Athey and Imbens (2015), Joseph (2019).

### Keynesian supply shocks and sectoral interactions
- Guerrieri and others (2022): theory of Keynesian supply shocks—large supply shocks in a few or a single sector can induce demand-like aggregate responses via complementarities and income effects.
- Empirical evidence:
  - Cesa-Bianchi and Ferrero (2021) find the channel active in the US even outside COVID-19.
  - The paper finds evidence for Keynesian supply shocks in a minority of countries in the Asia sample, and for Europe and the USA, but these results are fully explained by the COVID-19 event.
- Policy implication noted:
  - If shocks are asymmetric across sectors, a central bank raising interest rates to reduce inflation will face a larger sacrifice ratio than when inflation is driven by broad-based demand (Guerrieri, Lorenzoni and Straub (2021)).

### Constructing aggregate shocks from industry-level shocks for causal analysis
- Industry-level shocks may be correlated with common shocks (global or country-wide), reducing suitability for causal analysis.
- Solution adopted: construct granular instrumental variables following Gabaix and Koijen (2021) to produce country-level shocks purged of common shocks and driven by idiosyncratic movements in large sectors.

### Evidence on growth spillovers in the literature and paper’s positioning
- Growth spillovers from a 1 percentage point shock to growth in China:
  - Literature: Maliszewski and others (2016) find 0.2 pp; Furceri and others (2017) find 0.4 pp after 3 years; Barcelona and others (2022) find 0.3 pp from a 1 percent of GDP expansion of credit in China.
  - Paper: finds medium-term spillovers of approximately 0.3 pp.
- One standard deviation for the granular shock: 94 basis points of growth in China and 60 basis points in Europe and the USA. Thus a 1 pp shock in the USA or Europe would imply spillovers close to double those from a 1pp shock in China.
- Spillovers from the US and Europe:
  - Canova (2005): 0.3-0.5 percent spillovers to Latin America for US demand and supply shocks.
  - Feldkircher and Huber (2016): demand and supply shocks driving US GDP have small real spillovers in other countries of approximately 0.1 percentage points, whereas US monetary shocks have effects more than twice that size.
  - Paper: finds spillovers peaking at 0.4 pp after 6 quarters for the US and Europe; short-term (within 1 year) spillovers from China limited except to Asia.
- Caveats:
  - Spillover estimates for the US are constrained by a short sample period (2011-2019) that covers a relatively placid business cycle period.
  - Europe and China samples start in 2006 and include the Global Financial Crisis (2007-8) and the European debt crisis (2009-10).

### How this paper extends prior work
- Uses industry-level price and quantity indices to classify industries as supply or demand driven and aggregates to country-level decompositions as weighted averages of industry growth rates.
- Extends Shapiro (2022) by:
  - Distinguishing idiosyncratic vs. majority-representative movements across sectors.
  - Employing a wide range of machine learning models and selecting the optimal model for each country.
- Applies decompositions to:
  - Diagnose the mix of demand and supply drivers of lockdowns in 2020 and the post-COVID recovery (supply factors important in 2021; demand factors more important in 2022).
  - Explain the sharp increase in inflation in 2021-22: large and advanced economies generally experienced more supply shocks (China, Australia, Korea), while emerging markets saw significant demand pressures (Indonesia, Malaysia, Philippines, Vietnam, Thailand). Much of Indonesia’s increase is idiosyncratic and driven by a single commodity-related sector.
  - Study Keynesian supply shock incidence and growth spillovers using granular instrumental variables.

*Source: wpiea2023203-print-pdf - 1.1 Literature*

### 2.1 Empirical Models

### 2.1 Empirical Models

### Model framework and forecasting objective
- Forecast for each country (c) and industry (i): 
  - ŷi,t+h_c = f(θ, X_i,t_c, X_i,t−1_c, ..., X_i,t−4_c) (equation (1)).
  - ŷi,t+h_c is the forecast of the year-on-year change in the price or quantity index for industry i in country c, h periods ahead.
  - f(.) is the forecasting function (model-specific); θ are hyperparameters chosen based on out-of-sample forecasting performance.
- Predictor matrix X_i,t_c includes:
  - lags of the target variable y_i,t_c;
  - values for target variables in industries other than i (denoted x_{-i,t}_c);
  - a set of global control variables that are the same for each country and industry.
- Decomposition of aggregate GDP and GDP deflator:
  - Aggregate quantity: y_t = ∑_{i=1}^N w^y_t_i y_{i,t} where w^y_t_i is industry weight in real GDP.
  - Aggregate deflator: π_t = ∑_{i=1}^N w^π_t_i π_{i,t} where w^π_t_i is industry weight in nominal GDP.
  - Forecast errors: ε^Q_{i,t} = ŷ_{i,t} − y_{i,t} (GDP forecast errors) and ε^P_{i,t} = π̂_{i,t} − π_{i,t} (GDP deflator forecast errors).
  - Classification rule by industry i: movements are classified as demand if ε^Q_{i,t} ε^P_{i,t} > 0 and supply if ε^Q_{i,t} ε^P_{i,t} < 0.
  - Shares of demand and supply in aggregate GDP or its deflator are determined using industry weights.

### Models considered
- Eight forecasting models in total; baseline is OLS (same as Shapiro (2022)):
  - Baseline OLS uses only lags of the target variable: X_i,t_c = {y_{i,t}_c}.
  - Remaining seven models use the full predictor matrices X_i,t_c.
- Linear models (3):
  - Lasso (L1 penalty; Tibshirani (1996)): performs predictor selection via an L1 norm penalty; hyperparameter is weight on L1 penalty.
  - Ridge (L2 penalty; Hoerl and Kennard (2000)): shrinks coefficients (L2 norm), places positive weight on all predictors; similar predictions to principal component regression.
  - Elastic-Net (Zou and Hastie (2005)): convex combination of L1 and L2 penalties; hyperparameter selects relative weight between L1 and L2.
- Non-linear models (4):
  - Support Vector Regression (SVR): linear SVR minimizes absolute distance of predictions subject to slope coefficients remaining small (L2 norm); non-linearity introduced via basis function transformations of predictors.
  - Random Forest regressions (Breiman (2001)): aggregates many simple bootstrap-fitted decision trees; prediction is average across trees.
  - Ada-boosted trees (Freund and Schapire (1997)): sequentially reweights observations to focus on large errors; hyperparameters include number of models and learning rate.
  - K-nearest-neighbor (KNN) regression: clusters data into K clusters (centers reduce within-cluster variance); cluster means used for prediction.

### Rationale and model behavior
- OLS baseline is competitive because including additional predictors can create overfitting; OLS performance is significantly higher if additional predictors are excluded.
- Ridge behaves similarly to factor models by shrinking coefficients of correlated predictors; elastic-net interpolates between Lasso and Ridge depending on hyperparameter.
- Random Forest and Ada-boost address decision tree overfitting by aggregation (bootstrapping or sequential reweighting).
- KNN is analogous to a factor model via dimensionality reduction by clustering.

### Forecasting exercise setup
- Training and evaluation:
  - For each country, industry, and price/quantity measure, models are trained on a period of 4 years (16 quarters).
  - Out-of-sample forecasting exercise begins on the following quarter’s prediction.
- Model selection criterion:
  - Average root mean squared error (RMSE) over all industries, prices and quantities is used to compute RMSE metric for each country (c) and model (m):
    - RRRMSE_{c,m} = ∑_{i=1}^N ∑_{j ∈ {P,Q}} (ŷ_{i,j,t+h}_c − y_{i,j,t+h}_c)^2 / N  (equation (2) described in text).
  - j ∈ {P (deflator), Q (real value-added growth)}.
- Forecast horizon for lead/lag event study: h = −3, ..., 3 (equation (3) used below).

---

### 2.2 Data

- Coverage:
  - Real and nominal GDP value-added data by industry for 12 Asian countries, covering 74 percent of GDP in the Asia-Pacific Region.
  - Data source: Haver.
  - Quantity growth: year-on-year percentage change in real GDP value-added for each industry.
  - Deflator: growth rate in nominal GDP less growth rate in corresponding real data.
  - Industries selected based on availability; median number of industries is 16.
  - China required imputations for industry-level decomposition because one industry series not published since Q4 2017 (imputation details in annex I).
- Global predictors included:
  - Aggregate measures of GDP growth, investment, unemployment and policy rates in advanced economies;
  - Global share price returns (Dow Jones and MSCI);
  - Global industrial production growth;
  - Export and import growth aggregates for the USA, Euro Area and advanced Asia;
  - Global policy uncertainty (GDP weighted measures from Baker, Bloom and Davis (2016));
  - Consumer confidence measures for the US and the EU (Conference Board and European Commission);
  - Global PMI composite, manufacturing and services from JP Morgan;
  - Global liquidity measures covering local and cross-border claims on the nonfinancial sector (Bank for International Settlements).
  - Data sources and transformations detailed in annex II.

- Table 1: Data Coverage (Country — Start Date — End Date — Number of Industries)
  - AUS 12/31/2008 12/31/2022 20
  - BRN 6/30/2016 12/31/2022 24
  - CHN 6/30/1999 12/31/2022 9
  - IDN 6/30/2014 3/31/2023 17
  - IND 9/30/2002 3/31/2023 8
  - KOR 6/30/2006 12/31/2022 15
  - MNG 6/30/2016 12/31/2022 9
  - MYS 6/30/1997 12/31/2022 6
  - PHL 6/30/2006 12/31/2022 16
  - SGP 6/30/1981 12/31/2022 11
  - THA 6/30/2006 12/31/2022 17
  - VNM 6/30/2016 3/31/2023 20

---

### 2.2.1 GDP Deflator: validation and leading indicator properties

- Aggregation consistency test:
  - Error term: ε_t = π_t^{NGDP} − ∑_i w_{i,t}^N π_{i,t}^{VAD}, where π_t^{NGDP} is aggregate nominal GDP deflator, π_{i,t}^{VAD} is industry-level deflator, and w_{i,t}^N are nominal GDP weights.
  - Measured RMSE, MAE, and correlation between π_t^{NGDP} and weighted average industry deflator.
- Table 2: Deflator Aggregation Errors (Country — RMSE — MAE — Correlation*)
  - AUS 0.88 0.63 0.98
  - BRN 2.24 1.48 0.99
  - CHN 0.76 0.40 0.97
  - IDN 0.69 0.37 0.98
  - IND 0.55 0.31 0.98
  - KOR 0.26 0.17 0.99
  - MNG 3.50 2.34 0.95
  - MYS 0.81 0.54 0.99
  - PHL 0.20 0.15 1.00
  - SGP 0.65 0.45 0.99
  - THA 0.18 0.13 1.00
  - VNM 0.66 0.45 1.00
  - *Correlation between π_t^{NGDP} and ∑_i w_{i,t}^N π_{i,t}^{VAD}.
- Interpretation:
  - Errors are moderate for most countries except BRN and MNG, where deflator-based inflation is on average 2 and 3 times higher, respectively, than other countries in the sample.
  - Correlations are very high across countries; direction of price changes is reliable for decomposition purposes.

- GDP deflator as a leading indicator for CPI inflation:
  - Event study regression: Δy_{i,t+h} = α_i + γ_t + β_h Δπ_{i,t} + ε_{t+h} (equation (3)), where y is GDP deflator-based inflation and π is CPI-based inflation, with h = −3, ..., 3.
  - Findings:
    - GDP-based inflation is a significant leading indicator for rises in CPI inflation in Asia.
    - The deflator increases by around 77 bps in the lead-up to a typical 100 bps increase in CPI inflation.
    - The signal is positive and statistically significant in the 3 quarters preceding the CPI increase.
    - Adding additional controls (Consensus Economics current year CPI expectations, PPI inflation rates, wage inflation, import prices) reduces the GDP deflator increase in the quarters leading up to a CPI rise by around half on average but the effect remains significant and positive for all cases except when adding lags of wage inflation (wage inflation data are sparse and roughly halve the sample of countries).

---

### 2.3 Optimal Forecasting Models

- Optimal model selection: for each country c, select model m* with lowest RMSE: m* = argmin_m RRRMSE_{c,m}.
- Empirical results (summary):
  - Tree-based model averaging (Random Forest and Ada-boosted Trees) perform best in terms of average RMSE overall.
  - Baseline OLS is competitive and optimal in several cases: China, India, Malaysia, Singapore.
  - Random Forest and Ada-boost are optimal for 4 countries (same count as OLS optimal cases) but with lower average RMSE where they are optimal.
  - Conclusion: while OLS is not optimal in the majority of cases, using OLS is unlikely to lead to very large differences in results (robustness section referenced).

- Table 3: Optimal Model by Country
  - AUS Ada-boosted Trees
  - BRN Ada-boosted Trees
  - CHN OLS
  - IDN Random Forest
  - IND OLS
  - KOR Ada-boosted Trees
  - MNG KNN Regression
  - MYS OLS
  - PHL Random Forest
  - SGP OLS
  - THA Random Forest
  - VNM Ada-boosted Trees

_Italic: Source — wpiea2023203-print-pdf, "2.1 Empirical Models", IMF Working Papers._

### 3. Results: GDP and Inflation Decompositions

### 3. Results: GDP and Inflation Decompositions

### Methodological note on classification uncertainty
- The decomposition treats forecast errors near zero as unknown using a parameter 훼훼.
- Baseline uses 훼훼 = 0.1, which implies that 20 percent of forecast errors where the errors are close to zero result in an unknown classification.
- Alternative results with higher 훼훼 are presented in the annex as robustness checks.
- Data for VNM and BRN only start in 2016, hence the full Asia aggregate results are limited to starting in 2016.

### Aggregate GDP decomposition findings (Asia and Asia excluding China)
- Results cover three periods: the four years prior to the COVID-19 shock in 2020, the lockdowns during the COVID-19 shock in 2020, and the post-COVID recovery in 2021-22.
- Supply accounted for a significant share of growth prior to the COVID-19 shock; this is pronounced in Indonesia, Korea and the Philippines.
- The major GDP drop during lockdowns occurred in 2020Q2 with a majority of the decline in GDP attributed to demand in several countries.
  - Demand-driven drops during lockdowns observed in: Indonesia, India, Malaysia, the Philippines, Singapore, and Thailand.
- Recovery in late 2020 and early 2021 was predominantly demand-driven, with a significant role for supply growth in Australia, Malaysia and especially Korea (the latter primarily driven by supply).

### Inflation decomposition findings
- Inflation broadly mirrors GDP decomposition:
  - Supply factors are more prominent in 2021 and demand picking up in 2022.
- Supply chain disruptions following lockdowns contributed substantially to price rises in 2021 (supply-driven inflation).
- Demand components remain significant, suggesting demand-supply mismatches and demand-pull elements in 2021-22.
- Country-specific patterns for demand-pull inflation:
  - Indonesia and Malaysia: much of demand-pull inflation driven by primary commodity price movements, which are typically determined internationally.
  - Philippines and Vietnam: evidence of broad-based price increases across sectors, more likely mapping into domestic conditions.

### Measurement and treatment of idiosyncratic industry shocks
- Idiosyncratic component of demand and supply measured via cross-sectional outliers relative to the median over industries in each quarter.
- Industry-level components normalized to mean 0 and standard deviation 1 using pre-COVID sample values.
- Outliers defined where the value is 4 standard deviations above the cross-sectional median.
  - For flagged outliers, the industry-quarter value is assigned equal to the pre-COVID mean for that industry.
- Rationale: median or expenditure-weighted median alternatives would not preserve the expenditure-weighted decomposition of the headline GDP deflator and produced excessive volatility in some applications.
- Consequence: Much of the commodity-price-driven inflation in Indonesia during 2021/22 is classified as idiosyncratic; Malaysia receives a different classification (possibly due to small sector count). Philippines and Vietnam show less idiosyncratic influence—interpreted as genuine demand-pull inflation.
- Idiosyncratic inflation is less likely to signal an overheating economy for central banks unless strong downstream transmission occurs.

### Selected country-level GDP and inflation patterns
- Korea:
  - GDP growth heavily driven by supply movements, primarily manufacturing.
  - Recovery post-pandemic driven by manufacturing, later impacted by sharp declines in semi-conductor prices in 2022.
- Brunei:
  - Idiosyncratic shocks driven by mining and manufacturing of Liquefied Natural Gas.
  - Supply movements primary for growth outturns; large idiosyncratic demand shocks (manufacture of LNG) in 2020.
- Indonesia:
  - Apart from the 2020Q2 lockdown drop, demand-driven GDP components arise from a wider group of sectors: Manufacturing; Wholesale and Retail trade; Transportation and Warehousing.
- India:
  - Slump and recovery around the pandemic primarily demand-driven with significant roles for manufacturing and transport/hospitality sectors.
  - Recent growth outturns (excluding 2022Q2) show more sizable supply components with trade, transport, communication and broadcasting contributing.

### China: timing and drivers
- Chinese GDP fell sharply in 2020Q1, one period prior to most other Asian nations, due to lockdowns in Hubei province in January–April 2020.
- Those lockdowns resulted in a severe supply shock across most sectors.
- Recovery pattern:
  - 2020Q2: rebound in supply in all sectors except Hotel and Catering.
  - 2020Q3–2021Q1: demand-driven recovery, especially in manufacturing.
  - 2022Q2 onward: growth primarily driven by supply in manufacturing.
- Chinese CPI:
  - Showed mild deflation in late 2020 and early 2021, then remained relatively close to its historic average rate in the last decade of 2 percent despite global price rises in 2021-2022.
  - Note: very recent CPI data for 2023 showed CPI inflation touching zero in 2023Q2.
- Chinese GDP deflator:
  - Rose sharply in early 2021 based on supply shocks.
  - With the exception of an aberrant 2021Q3, GDP deflator–based inflation has been supply driven and decelerated sharply in 2022 from a peak of over 6 percent to zero in 2022Q4 despite relatively stable GDP growth over this time.
  - Much of these movements driven by the manufacturing/industry sector.

### Comparison with USA and Europe (EU27)
- High inflation in the USA and Europe motivated policy rate increases by the Federal Reserve and the ECB.
- USA:
  - Significant role for supply shocks in driving the increase in inflation in 2021/22.
  - Half of the initial increase in inflation rates in 2021Q2 is explained by idiosyncratic shocks.
  - Supply shocks broadened across sectors (Manufacturing; Wholesale & Retail Trade; Transportation & Warehousing) accounting for 40 percent of inflation in 2022.
  - Idiosyncratic shocks played a significant role in output during lockdowns in 2020Q2 and during recovery, fading in late 2022.
- Europe:
  - Almost half of the run-up in inflation explained by idiosyncratic demand shocks concentrated in two sectors: manufacturing and Trade, Travel, Accommodation and Food from 2022Q1 onward.
  - Interpreted as related to unwinding of the COVID-19 shock and shifts in demand toward goods and hospitality rebounds.
  - Supply role grows from 2022Q3 onward, dominated by Industry excluding construction (includes electricity and gas production) — likely linked to the sharp rise in European gas prices in 2022 following the war in Ukraine.
- Policy implication noted: If supply shocks or idiosyncratic demand shocks predominate, monetary policy tightening may carry higher output costs for reducing inflation compared with cases where demand shocks dominate.

### Application I — Keynesian Supply Shocks: identification and results
- Theoretical background:
  - Guerrieri and others (2022) show in a two-sector model with nominal wage rigidity that large negative supply shocks in one sector can reduce income and aggregate demand sufficiently that aggregate demand declines more than aggregate supply, depending on sectoral substitutability and income effects (MPC).
- Empirical approach:
  - Industry-level price and quantity shocks (classified as demand or supply) are used to analyze correlations between industry-level supply shocks and subsequent demand responses.
  - Local projections estimated country-by-country:
    - Regression (notation preserved from source): 휀푖,푡+ℎ NDDmGDD − 휀푖,푡−1 NDDmGDD = 훼푖 + 훾푡 + 훽ℎ 휀푖,푡 SSSS Gy + ∆휀푖,푡−1 NDDmGDD + vt+ℎ
    - 훽ℎ measures cumulative change in demand shocks following a supply disturbance at the sectoral level.
    - Negative 훽ℎ consistent with Keynesian supply shocks (negative supply → negative aggregate demand); positive 훽ℎ consistent with sectoral substitutability (demand reallocation).
- Empirical findings:
  - Evidence of Keynesian supply mechanism (negative aggregate demand movements following asymmetric supply shocks) found for: Australia, Indonesia, India, and to a lesser degree, Singapore and Vietnam.
  - In the remaining 7 Asian countries in the sample, limited evidence of declines in aggregate demand following supply shocks at the industry level.
  - The COVID-19 period plays an important role: the dashed-line pre-2020 average responses generally show limited Keynesian supply effects — indicating very asymmetric and large supply shocks (such as COVID-19 lockdowns) are likely needed for supply shocks to generate Keynesian aggregate demand effects.
- USA and Europe comparison:
  - Evidence of Keynesian supply shock mechanism present in both USA and Europe, stronger in the USA (negative supply shocks followed by negative aggregate demand shocks).
  - As with Asian samples, this result is dependent on inclusion of the COVID-19 period; excluding COVID-19, the negative correlation is no longer evident.

*Source: wpiea2023203-print-pdf - 3. Results: GDP and Inflation Decompositions*

### 5. Application II: Growth Spillovers

### 5. Application II: Growth Spillovers

### Methodology: construction of granular country shocks
- Industry-level shocks: ε_{i,t}^Q decomposed as ε_{i,t}^Q = θ_i μ_t + z_{i,t} (aggregate component μ_t and idiosyncratic component z_{i,t}).
- Aggregate country-level shocks:
  - Size-weighted aggregate: ε_t^w = ∑_{i=1}^N w_{i,t} ε_{i,t}^Q
  - Equally-weighted aggregate: ε_t^E = (1/N) ∑_{i=1}^N ε_{i,t}^Q
- Granular residual (idiosyncratic country shock) defined as:
  - z_t = ε_t^w − ε_t^E = (Q_t^w − Q_t^E)  (equation (4))
- Use z_t as an instrumental variable for aggregate quantity shocks ε_t.
- First-stage estimation: post-lasso procedure of Chernozhukov and others (2015); predicted value denoted ẑ_t estimated by OLS using predictors selected by lasso.
  - Joint F statistics on the granular instruments in the post-lasso OLS estimation are 9.5 (USA), 14 (EUR), 11.8 (China).

### IV local projections panel specification
- Outcome and specification (for horizons h = 0...6):
  - y_{i,t+h} − y_{i,t−1} = α_i + ρ t + β_h ẑ_t + γ z_{t−1} + κ κ y_{i,t−1} + Γ' X_{i,t} + v_{i,t+h}  (equation (5))
- Variables and controls:
  - y_{i,t+h} is log GDP for country i.
  - α_i are country fixed effects; t captures a linear time trend.
  - X_{i,t} includes control variables: financial conditions indices for advanced and emerging economies, Chicago Board Options Exchange Volatility Index, two lags of world export-weighted GDP growth for the country.
  - GDP and financial conditions data sourced from IMF (2023).
- Samples:
  - China and Europe samples start in 2006Q2.
  - USA sample starts in 2011Q2.
  - Samples end in 2019Q4 to avoid COVID-19 lockdown shocks.
- External dataset: quarterly dataset from Copestake and others (Forthcoming) covering 51 economies.
  - Asia-only sample includes 12 economies: Japan, Australia, New Zealand, Taiwan Province of China, Hong Kong SAR, India, Indonesia, Korea, Malaysia, Philippines, Singapore, and Thailand.
- One standard deviation of z_t: 94 bps in China and 60 bps in EUR and USA.

### Empirical findings: growth spillovers by source economy
- China:
  - Limited evidence of short-term global growth spillovers from China idiosyncratic shocks.
  - Medium-run evidence of spillovers of approximately 0.3 percent of GDP.
  - Asia-only results: short-term spillover of approximately 0.4 percent (short lived and fully offset by the end of the first year); medium-term spillovers remain approximately 0.3 percent.
- Europe:
  - Global growth spillovers slightly larger than China's peak at 0.4 percent of GDP.
  - Spillovers from Europe are notably more persistent and larger in the short term.
  - Spillovers to Asia are broadly similar in magnitude to rest of world but more short-term; approximately zero of 1.5 years (text reports "approximately zero of 1.5 years .").
- United States:
  - IV estimates indicate a significant drop in global (ex-USA) growth in the short term, peaking around 0.4 percent after 6 quarters.
  - Impact on Asia is more moderate at around 0.2 percent.
  - Caveats:
    - USA z_t sample begins in 2011, covering the comparatively stable period after the Global Financial Crisis and before the COVID-19 shock; this likely contributes to smaller estimated global spillovers.
    - The shocks considered are idiosyncratic real-economy shocks and exclude important financial channels (e.g., US monetary policy, financial conditions) that have known global spillovers; excluding these channels may moderate estimated effects on Asia.

### Robustness checks for growth spillovers
- Two alternative IV local projection specifications considered:
  - (i) additional lags of the target variable (GDP growth for non-source countries).
  - (ii) lags of the shock variables and trend growth in the country.
- Results:
  - Conclusions broadly unchanged across specifications.
  - Specification (ii) provides stronger evidence for near-term spillovers from China and the USA while yielding similar medium-term conclusions.
  - For Europe, specification (ii) suggests spillovers may build more slowly, peak slightly larger, and potentially be weaker in the medium term.

### Key statistics and sample details (exact values preserved)
- Joint F statistics (post-lasso OLS): 9.5 (USA), 14 (EUR), 11.8 (China).
- One standard deviation of z_t: 94 bps (China); 60 bps (EUR); 60 bps (USA).
- Sample start dates: China and Europe: 2006Q2; USA: 2011Q2.
- Sample end date: 2019Q4.
- Impulse-response horizon: h = 0...6 quarters.

*Source: IMF Working Paper — Section 5, "Application II: Growth Spillovers".*

### Annex IV. GDP and Inflation Decompositions by

### Annex IV. GDP and Inflation Decompositions by Country

### GDP decompositions (Figure A2)
- Figure referenced: "Figure A2: GDP Decompositions".
- Presented as country-level decompositions of GDP (figure images not transcribed in source text).
- Panel location references: IMF WORKING PAPERS Monitoring Demand and Supply in Asia; INTERNATIONAL MONETARY FUND page markers 37–38.

### Inflation decompositions (Figure A3)
- Figure referenced: "Figure A3: Inflation Decompositions".
- Presented as country-level decompositions of inflation (figure images not transcribed in source text).
- Panel location references: IMF WORKING PAPERS Monitoring Demand and Supply in Asia; INTERNATIONAL MONETARY FUND page marker 39.

### GDP and inflation decompositions for Asia including China (Figure A4)
- Figure referenced: "Figure A4: GDP and Inflation Decompositions for Asia including China".
- Notes accompanying figure:
  - "Deocmpositions for Asia based on GDP weighted average of individual country results."
  - "Uncertinaty parameter 훼훼=0.1 so that 20 percent of observations are unclassified."
  - "Asia sample exlcudes BRN and VNM to allow for longer sample."
- Panel location references: IMF WORKING PAPERS Monitoring Demand and Supply in Asia; INTERNATIONAL MONETARY FUND page marker 40.

### Sensitivity to higher model uncertainty — Annex V (alpha = 0.2)
- Context: "Annex V. GDP and Inflation Decompositions Assuming Higher Model Uncertainty".
- Method variation: Results from main narrative repeated with alpha = 0.2.
- Figures:
  - "Figure A5: GDP Decompositions under 휶휶=ퟎퟎ.ퟐퟐ".
  - "Figure A6: Inflation Decompositions under 휶휶=ퟎퟎ.ퟐퟐ".
- Panel location references: IMF WORKING PAPERS Monitoring Demand and Supply in Asia; INTERNATIONAL MONETARY FUND page markers 41–44.

### Document navigation and location markers (figures and pages)
- Figures and related panels span IMF WORKING PAPERS Monitoring Demand and Supply in Asia; INTERNATIONAL MONETARY FUND page markers 37 through 45 for Annex IV and V figures.
- Final working paper markers continue into references section on pages 46–48 and the working paper identification line on page 48–49.

*Source: Annex IV and Annex V figures and accompanying notes from the provided IMF working paper content.*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2023/english/wpiea2023203-print-pdf.pdf_