## 1.     Augmented Error-Correction Mechanism

## Source details

**Canonical URL:** [1.     Augmented Error-Correction Mechanism](https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp06175.pdf)

## Other formats

- [Markdown version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp06175.pdf.md)
- [Structured JSON version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp06175.pdf.json)

---

### Introduction
- Problem statement:
  - Following adjustment programs in transition countries, inflation declined rapidly from hyperinflation to single-digit levels in the most successful ones, including the Kyrgyz Republic.
  - Monetary policy remains set on an adhoc basis due to instability of money demand and inadequate representation of the seasonal structure of the economy in inflation forecasts; unanticipated external shocks (for example, world oil price increases) also contributed.
  - Initial misspecification of seasonality in consumer prices and economic activity often yields monetary outcomes that amplify variations in prices and economic activity over the seasonal cycle.
- Objective and approach:
  - Two approaches to integrating the seasonal structure of consumer prices into inflation forecasting and simulations to search for an optimal monetary policy stance where seasonality is pronounced:
    - Approach 1: Model seasonality in consumer prices using general ARIMA models with stochastic seasonality — tracks seasonal pattern well and yields reliable forecasts but neglects the role of monetary policy in the inflation process.
    - Approach 2: Introduce the role of monetary policy in modeling and forecasting inflation; carry out tests for integration of consumer prices at all seasonal frequencies to establish constancy, or otherwise, of the seasonal structure over time.
- Model specification and forecasting framework:
  - Evidence suggests seasonal integration of consumer prices in the Kyrgyz Republic, indicating that seasonality could be modeled deterministically.
  - Specify an error-correction model of inflation where:
    - The error-correcting equation is represented by a simple long-run inflation equation.
    - The model is augmented with AR(1) data generation processes for the growth of money and exchange rate movements.
  - Develop a two-stage forecasting framework for two augmented ECMs — including and excluding seasonal dummies — and use response-surface-type simulation experiments to yield quarterly and annual inflation forecasts and corresponding optimal monetary policy stances.

### Inflation and Monetary Policy in the Kyrgyz Republic (stylized facts, 1995–2004)
- Quarterly seasonal pattern:
  - Third-quarter inflation rates have generally been negative except before 1996 and after the Russian financial crisis of 1998.
  - Highest inflation in second and fourth quarters; relatively modest in first quarter and price reductions in third quarter.
- Inflation trends:
  - Twelve-month consumer price inflation declined to 35 percent in 1996.
  - Inflation declined to 14 percent in 1997.
  - Inflation reached a little below 10 percent by mid-1998 (just before the Russian financial crisis).
- Money and exchange rate dynamics:
  - Prior to the Russian financial crisis, money growth was on a steady downward path and the rate of depreciation of the som had declined from earlier high levels.
  - Following the Russian financial crisis of late 1998:
    - Inflation picked up to around 40 percent between June and December 1999.
    - In the first half of 1999, the som depreciated on average by close to 50 percent compared to the same period in the previous year.
    - Deposit dollarization reached 63 percent in June 1999.
    - Money growth increased to 34 percent by end-1999, from a low of 17 percent at end 1998.
  - Subsequent developments:
    - Inflation declined quickly after late-1999 as monetary policy was tightened, allowing a gradual decline in nominal depreciation of the som.
    - Since 2001, the som tended to appreciate amid increasing foreign inflows, contributing to further decline in headline inflation despite high money growth.
    - The increase in money supply was demand-determined and had little inflationary impact; reserve money increased with net foreign assets of the central bank in a pseudo currency board manner.
    - The pass-through of exchange rate appreciations to inflation overshadowed monetary expansion effects.

### Seasonal characteristics of consumer prices and stochastic seasonality tests
- ARIMA/SARMA estimation (Box-Jenkins approach) — selected specifications:
  - Monthly CPI: ARIMA(2,0,1) with a two-period lag seasonal autoregressive process and a one-month lag moving average process.
  - Quarterly CPI: ARIMA(1,0,1) with a seasonal AR(1) process.
  - "Almost all the autoregressive seasonal coefficients for the four consumer price series considered are significant at the 1 percent significance level."
- Canova-Hansen (CH) seasonal integration tests using ln(CPI) and ∆ln(CPI):
  - Absence of seasonal unit roots at the biannual and annual frequencies—nonsignificance of the l2 and lπ statistics relative to the critical value of 1.01 (generalized Von Mises distribution with 3 degrees of freedom) at the 5 percent level.
  - CH/joint tests indicate absence of seasonal instability.
  - Individual dummy stability tests:
    - No seasonal inversion occurred.
    - Seasonal components remain significant in the inflation process.
    - Seasonal patterns have generally remained flat over the sample period (indicative of seasonal integration).
    - The third-quarter seasonal dummy remained negative for most of the sample, with its significance becoming stronger in later parts of the sample period.
    - The first and second quarters display the highest inflation rates; fourth quarters display relatively moderate rates.
- Selected numerical CH/Table 2 results (1995:01–2004:04):
  - Casanova-Hansen statistics: 0.4800, 0.1300, 0.1300, 0.4200, 0.6600, 0.6500.
  - R squared: 0.9924 and 0.4887.
  - F-ratio: 1103.08* and 8.1249*.
  - Durbin's h-statistics: 0.4849 and 0.4982.
  - Note: An asterisk (*) indicates statistical significance at the 1 percent level.

### Monetary model of inflation and long-run equation
- Price aggregation and model structure:
  - p_t = θ p^T_t + (1−θ) p^NT_t, 0<θ<1.
  - p^T_t = q_t s_t (import price in foreign currency times nominal exchange rate).
  - p^NT_t depends on excess money (s_t m_t − d_t m_t) and money demand (function of output and prices).
- Log-linear long-run specification (including gold price P_gold and trade-weighted foreign prices P_for):
  - ln p_t = constant + β1 ln P_gold_t + β2 ln P_for_t + β3 ln s_t + β4 ln m_t + μ_t.
- Long-run OLS estimates (1995–2004) — selected numbers:
  - Coefficient estimates: 1.321, 0.402, 0.062, 0.141, -0.043, 0.058, 0.062.
  - Corresponding p-values (parentheses): (0.022), (0.000), (0.335), (0.038), (0.001), (0.014), (0.000).
  - R-squared: 0.995.
  - Regression F (6,31): 1,529.717; Significance of F: 0.000.
  - Durbin's h-statistics: -0.206.
- Interpretation:
  - Estimates are consistent with expectations—foreign trade partner price increases, monetary expansion, and exchange rate depreciation increase inflation.

### VAR identification of shocks and relative contributions
- Six-variable VAR (two lags) with Choleski decomposition: x' = [ln Pgold, ln Pfor, ln y, ln m, ln s, ln cpi] plus seasonal dummies.
- Impulse responses to temporary 1 percent shocks:
  - Pgold: insignificant reductions in consumer prices throughout forecast horizon.
  - Pfor: significant increases in consumer prices in the first five quarters following the shock.
  - Output (ln y): temporary 1 percent positive output shock reduces consumer prices significantly in the first two quarters; effect tends to zero thereafter.
  - Money (ln m): temporary 1 percent increase in money growth induces an increase in consumer prices two quarters after the shock.
  - Exchange rate (ln s): depreciation passes through quickly to prices; effect peaks about four quarters after the shock.
- Contribution to forecast variation in consumer prices:
  - Monetary factors (changes in money and exchange rates) contribute two-thirds of the forecast variation in consumer prices over the forecast horizon after a two-period lag.
  - Monetary factors explain 36 percent of the variation in the immediate horizon; real factors explain 14 percent.
  - Effect of real factors increases to 43 percent of the second-quarter forecast error in consumer prices; monetary factors also increase slightly.
  - In the long run, monetary factors account for 77 percent of the 12-quarter forecast mean square error of consumer prices.

### Augmented error-correction models (ECMs) and short-run dynamics
- Preferred augmented ECM includes seasonal dummies, lagged inflation, money growth, exchange rate changes, and lagged error-correction term EC_{t-1} derived from the long-run equation.
- General short-run ECM form includes seasonal dummies dum_j, lagged ∆ln cpi terms, ∆ln m (money), ∆ln s (exchange rate), and lagged EC term.
- Estimation details and findings:
  - Two broad specifications estimated: with and without seasonal dummies; both assume AR(1) data generation for money growth and exchange rate changes in baseline estimations (alternative random walk specifications also considered in forecasting).
  - Real output was statistically insignificant in the short-run dynamics when seasonal dummies are included and was excluded from short-run regressors.
  - Joint nonlinear systems estimation indicates:
    - It takes between five and six months for deviations of inflation from the long-run path to be corrected under both model variants.
    - Symmetry in the adjustment mechanism (speed of adjustment upward similar to downward across model variants).
    - Instantaneous pass-through: exchange rate depreciation impact on inflation is faster and larger than instantaneous monetary expansion.
      - A 1 percent money growth increase raises inflation by 0.1 percent under both models (insignificant in model excluding seasonal components).
      - Instantaneous exchange rate depreciation increases inflation by about 0.3 percent under both models.
- Hypothesis tests on data-generation processes for money and exchange rate (5 percent significance level):
  - Null (i): change in money follows a random walk with no drift (γ10 = 0, γ11 = 1) — rejected.
  - Null (ii): change in exchange rate follows a random walk with no drift (δ10 = 0, δ11 = 1) — rejected.
  - Null (iii): levels of money and exchange rate series follow AR(1) processes with no constants (γ10 = δ10 = 0) — not rejected.
  - Forecasting exercise adopts an agnostic approach: considers AR(1) and random walk specifications for money and exchange rate changes, yielding four forecast models.

### Selected augmented ECM numerical results (selected Table 4 highlights)
- CPI equation and money/exchange rate equation parameter estimates under two specifications (Without Seasonal Dummies / With Seasonal Dummies):
  - Log Likelihood: 82.862 and 92.883.
  - Durbin's h-statistics: 0.7043* and 0.3186* (asterisk denotes statistical significance at the 5 percent level in the table note).
  - Hypothesis test chi-square statistics: 39.4729*, 39.4729*; 18.1046*, 18.1046*; and 4.7006, 4.7006 (degrees of freedom 2, 2, and 1 respectively); asterisk (*) indicates 5 percent significance for the first two.
- Interpretation summary from ECM estimates:
  - Error-correction speed implies 5–6 months to correct deviations.
  - Instantaneous exchange rate pass-through ≈ 0.3 percent inflation per 1 percent depreciation.
  - Instantaneous money-growth pass-through ≈ 0.1 percent inflation per 1 percent money growth.

### Forecast performance (out-of-sample) and comparative metrics
- Measures used: mean absolute forecast error, root mean square forecast error (RMSE), and Theil’s U inequality coefficient.
- Table 5 — Out-of-Sample Inflation Forecasts (selected numbers as presented):
  - Quarterly inflation (benchmarks and forecasts shown in table):  
    13.04 1.20 3.63 3.16 1.68 3.57 2.61 2.50  
    20.87 0.69 3.08 4.96 3.21 2.63 1.40 1.80  
    3 -0.90 -0.91 2.84 -0.01 -2.33 3.27 1.50 -0.90  
    41.44 1.60 4.14 4.56 2.15 3.96 2.21 1.70
  - RMSE0.270.431.222.271.132.511.27
  - Mean Absolute Error1.532.518.787.774.318.533.62
  - Theil's U-statistics0.150.240.691.280.641.420.72
  - Source line in table: "Source: Author's estimates."
- Principal comparative findings:
  - Models that include seasonal components outperformed those without seasonal components using all three measures.
  - A random walk process for exchange rates outperforms an AR(1) process in forecasting inflation.
  - Models with seasonal dummies track seasonality and duplicate turning points better than those without seasonal components.
  - Example: models with seasonal dummies project negative inflation in Q3 (consistent with historical data); models without seasonal dummies predict positive inflation between 1.5 percent and 3.3 percent in Q3.

### Simulation design, first-stage results and policy-stance implications
- Two-stage simulation design:
  - Stage 1: simulate the estimated augmented ECM under varying assumptions on the autoregressive coefficient for money growth while keeping the exchange rate data generation process fixed (either AR(1) or random walk).
  - Stage 2: simulate the system using the optimal money-growth AR(1) coefficients to determine the exchange rate data generation process consistent with the predetermined inflation forecast.
- First-stage optimal money-growth AR(1) coefficients to attain an ex ante annual inflation forecast of 4 percent:
  - Model with seasonal dummies: optimal AR(1) coefficient for money growth process is 0.7 for both AR(1) and random walk specifications for exchange rate movements.
  - Model excluding seasonal dummies: corresponding AR(1) coefficient for money growth is about 0.3 for both exchange rate data generation processes.
- Interpretation:
  - Inflation-expanding effects in Q2 and Q4 are higher under the model without seasonal components than with seasonal components.
  - Inflation-dampening effects in Q1 and Q3 are far higher under the model with seasonal components.
  - Model without seasonal components implies a higher annual inflation and the need for a relatively tighter monetary stance (consistent with a lower AR(1) coefficient for money growth).
- Second-stage simulations — exchange rate path, monetary policy stance, and seasonality:
  - Augmented EC model without seasonal components suggests bunching of positive quarterly inflation forecasts that do not disappear at any level of the AR(1) coefficient for nominal exchange rate changes—these are inconsistent with the historical seasonal pattern.
  - When seasonal components are included, simulation results are consistent with the historical pattern of inflation and its relationship with money and the exchange rate.
  - Conditional on an optimal money supply process with an AR(1) coefficient of 0.7, an AR(1) coefficient of below 0.4 for nominal exchange rate depreciation would be consistent with the observed inflation dynamics in the Kyrgyz Republic.
- Policy-stance examples for a 3 percent annual inflation forecast:
  - Seasonal-components model (line AA): required exchange rate change path corresponds to an AR(1) coefficient at 0.3.
    - Quarterly inflation along line AA: Q1 = 3.1 percent, Q2 = 4.9 percent, Q3 = -0.2 percent, Q4 = 4.5 percent.
  - No-seasonal-components model (line BB): achieving the same forecast requires a tighter monetary policy stance with a lower exchange rate depreciation (a lower AR(1) coefficient).
    - Quarterly inflation along line BB: Q1 = 3.4 percent, Q2 = 2.4 percent, Q3 = 2.9 percent, Q4 = 3.6 percent.
- Additional note:
  - Random walk exchange rate specification in the second-stage simulations yields lower quarterly and annual inflation forecasts under both model variants.

### Policy recommendations and concluding remarks
- Recommended framework:
  - Adopt an approach that accounts for seasonality, the monetary policy stance, and implied exchange rate movements in an encompassing setup to inform monetary policy consistent with short-term inflation forecasts or targets.
  - Use the proposed econometric "toolkit" for inflation forecasting that yields annual and quarterly inflation forecasts, corresponding monetary policy stance, and exchange rate movements consistent with that stance.
  - The framework is generally applicable to countries with pronounced seasonal fluctuations in economic activity and prices where authorities wish to determine monetary policy stance for short-term inflation forecasts or targets.
- Caveats and implementation considerations:
  - Efficacy of the approach depends critically on stability of the seasonal structure of consumer prices.
  - The approach may be subject to the "velocity instability problem" (the relationship between inflation and money may change over time), which mainly affects long-term forecasts more than short-term exercises.
  - Continual empirical updates and tests for stability of the money–inflation relationship are necessary before and during implementation.
- Operational note:
  - The two-stage procedure could be implemented simultaneously by monetary authorities determining foreign exchange market intervention and/or domestic open market operations (sterilized or nonsterilized) to yield the required monetary policy stance and corresponding exchange rate change consistent with the ex ante inflation forecast or target.

*Source: _wp06175 - 1.     Augmented Error-Correction Mechanism (PDF chapter/section).*

### 1.     Augmented Error-Correction Mechanism ................................................................... 13

### 1.     Augmented Error-Correction Mechanism ................................................................... 13

### Introduction
- Problem statement:
  - Following adjustment programs in transition countries, inflation declined rapidly from hyperinflation to single-digit levels in the most successful ones, including the Kyrgyz Republic.
  - Monetary policy remains set on an adhoc basis due to instability of money demand and inadequate representation of the seasonal structure of the economy in inflation forecasts; unanticipated external shocks (for example, world oil price increases) also contributed.
  - Initial misspecification of seasonality in consumer prices and economic activity often yields monetary outcomes that amplify variations in prices and economic activity over the seasonal cycle.
- Objective and approach:
  - The paper suggests two approaches to integrating the seasonal structure of consumer prices into inflation forecasting and performs simulations to search for an optimal monetary policy stance where seasonality is pronounced.
  - Approach 1: Model seasonality in consumer prices using general autoregressive integrated moving average (ARIMA) models with stochastic seasonality — tracks seasonal pattern well and yields reliable forecasts but neglects the role of monetary policy in the inflation process.
  - Approach 2: Introduce the role of monetary policy in modeling and forecasting inflation; carry out tests for integration of consumer prices at all seasonal frequencies to establish constancy, or otherwise, of the seasonal structure over time.
- Model specification and forecasting framework:
  - Evidence suggests seasonal integration of consumer prices in the Kyrgyz Republic, indicating that seasonality could be modeled deterministically.
  - Based on seasonality and cointegration results, specify an error-correction model of inflation where:
    - The error-correcting equation is represented by a simple long-run inflation equation.
    - The model is augmented with autoregressive (AR(1)) data generation processes for the growth of money and exchange rate movements.
  - Develop a two-stage framework for forecasting inflation under two broad augmented error-correction models — including and excluding seasonal dummies.
  - Use response-surface-type simulation experiments under these models to yield quarterly and annual inflation forecasts and corresponding optimal monetary policy stances.

### Inflation and Monetary Policy in the Kyrgyz Republic
- Stylized facts (1995–2004):
  - Quarterly seasonal pattern: third-quarter inflation rates have generally been negative except before 1996 and after the Russian financial crisis of 1998; highest inflation in second and fourth quarters; relatively modest in first quarter and price reductions in third quarter.
  - Inflation trends:
    - Twelve-month consumer price inflation declined to 35 percent in 1996.
    - Inflation declined to 14 percent in 1997.
    - Inflation reached a little below 10 percent by mid-1998 (just before the Russian financial crisis).
  - Money and exchange rate dynamics:
    - Prior to the Russian financial crisis, money growth was on a steady downward path and the rate of depreciation of the som had declined from earlier high levels.
    - Following the Russian financial crisis of late 1998:
      - Inflation picked up to around 40 percent between June and December 1999, reflecting increased depreciation of the som in response to the fall of the Russian ruble.
      - In the first half of 1999, the som depreciated on average by close to 50 percent compared to the same period in the previous year.
      - Deposit dollarization reached 63 percent in June 1999.
      - Money growth increased to 34 percent by end-1999, from a low of 17 percent at end 1998.
    - Subsequent developments:
      - Inflation declined quickly after late-1999 as monetary policy was tightened, allowing a gradual decline in nominal depreciation of the som.
      - Since 2001, the som tended to appreciate amid increasing foreign inflows (especially for mining), contributing to further decline in headline inflation despite high money growth.
      - The increase in money supply was demand-determined and had little inflationary impact; reserve money increased with net foreign assets of the central bank in a pseudo currency board manner.
      - The pass-through of exchange rate appreciations to inflation overshadowed monetary expansion effects.

### Seasonal Characteristics of Consumer Prices
- Seasonal behavior and modeling implications:
  - Seasonal factors are important in explaining inflation dynamics over the year.
  - The paper identifies an ARIMA process for consumer price seasonality and investigates seasonal integration at all seasonal frequencies.
  - Test results and implications:
    - Tests for integration of consumer prices at all seasonal frequencies suggest seasonal integration in the Kyrgyz Republic.
    - This indicates the seasonality of consumer prices could be modeled deterministically.
- Augmented error-correction modeling and forecasting:
  - Given seasonal integration and cointegration evidence, the paper specifies an error-correction model of inflation augmented with AR(1) processes for money growth and exchange rate movements.
  - Two-stage forecasting framework:
    - Two augmented error-correction models are considered — one including seasonal dummies and one excluding seasonal dummies.
    - Response-surface-type simulation experiments under these models produce quarterly and annual inflation forecasts and determine consistent optimal monetary policy stances.
- Key empirical findings:
  - Simulations show empirical significance of seasonality and monetary policy stance in inflation forecasting.
  - The augmented error-correction model is empirically superior (in terms of information content and policymaking contributions) to single-equation, Box-Jenkins-type general autoregressive seasonal models of inflation.
  - The framework can inform short-term inflation targeting and be adapted as a toolkit for simultaneously determining medium-term money and inflation paths where seasonal fluctuations are pronounced.

*Source: _wp06175 - 1.     Augmented Error-Correction Mechanism (PDF chapter/section).*

### Section II emphasized the role of seasonal factors, monetary policy and exchange rate

### _wp06175 - Section II emphasized the role of seasonal factors, monetary policy and exchange rate

### Seasonal structure of consumer prices and stochastic seasonality tests
- ARIMA/SARMA estimation (Box-Jenkins approach) indicates significant seasonality in consumer price indices.
  - Monthly CPI: ARIMA(2,0,1) with a two-period lag seasonal autoregressive process and a one-month lag moving average process.
  - Quarterly CPI: ARIMA(1,0,1) with a seasonal AR(1) process.
  - "Almost all the autoregressive seasonal coefficients for the four consumer price series considered are significant at the 1 percent significance level."
- Canova-Hansen (CH) tests for seasonal integration (null: stationary seasonality) using ln(CPI) and ∆ln(CPI):
  - Absence of seasonal unit roots at the biannual and annual frequencies—nonsignificance of the l2 and lπ statistics relative to the critical value of 1.01 (generalized Von Mises distribution with 3 degrees of freedom) at the 5 percent level.
  - CH/joint tests indicate absence of seasonal instability.
  - Individual dummy stability tests (regressing inflation on quarterly seasonal dummies) show:
    - No seasonal inversion occurred.
    - Seasonal components remain significant in the inflation process.
    - Seasonal patterns have generally remained flat over the sample period (indicative of seasonal integration).
    - The third-quarter seasonal dummy remained negative for most of the sample, with its significance becoming stronger in later parts of the sample period.
    - The first and second quarters display the highest inflation rates; fourth quarters display relatively moderate rates.
- Selected numerical CH/Table 2 results (1995:01–2004:04):
  - Casanova-Hansen statistics (columns shown): 0.4800, 0.1300, 0.1300, 0.4200, 0.6600, 0.6500.
  - R squared: 0.9924 and 0.4887 (two dependent-variable specifications).
  - F-ratio: 1103.08* and 8.1249*.
  - Durbin's h-statistics: 0.4849 and 0.4982.
  - Note: An asterisk (*) indicates statistical significance at the 1 percent level.

### Monetary model of inflation and long-run equation
- Price aggregation and model structure:
  - Domestic price level p_t = θ p^T_t + (1−θ) p^NT_t, 0<θ<1.
  - Price of tradables p^T_t = q_t s_t (import price in foreign currency times nominal exchange rate).
  - Price of nontradables p^NT_t depends on excess money (s_t m_t − d_t m_t) and money demand (function of output and prices).
  - Log-linear long-run inflation specification (including gold price P_gold and trade-weighted foreign prices P_for):
    - ln p_t = constant + β1 ln P_gold_t + β2 ln P_for_t + β3 ln s_t + β4 ln m_t + μ_t.
- Long-run OLS estimates (1995–2004) — Table 3:
  - Coefficient estimates: 1.321, 0.402, 0.062, 0.141, -0.043, 0.058, 0.062.
  - Corresponding p-values (parentheses): (0.022), (0.000), (0.335), (0.038), (0.001), (0.014), (0.000).
  - R-squared: 0.995.
  - Regression F (6,31): 1,529.717; Significance of F: 0.000.
  - Durbin's h-statistics: -0.206.
- Interpretation: estimates are consistent with expectations—foreign trade partner price increases, monetary expansion, and exchange rate depreciation increase inflation.

### VAR identification of shocks and relative contributions
- Six-variable VAR (two lags) using Choleski decomposition with variables x' = [ln Pgold, ln Pfor, ln y, ln m, ln s, ln cpi] plus seasonal dummies yields:
  - A temporary 1 percent increase in Pgold → insignificant reductions in consumer prices throughout forecast horizon (gold price increases tend to be associated with appreciation of the domestic currency).
  - A temporary 1 percent increase in foreign prices → significant increases in consumer prices in the first five quarters following the shock.
  - A temporary 1 percent positive output shock → reduces consumer prices significantly in the first two quarters; effect tends to zero thereafter.
  - A temporary 1 percent increase in money growth → induces an increase in consumer prices two quarters after the shock.
  - Exchange rate depreciation passes through quickly to prices; effect peaks about four quarters after the shock.
- Contribution to forecast variation in consumer prices:
  - Monetary factors (changes in money and exchange rates) contribute two-thirds of the forecast variation in consumer prices over the forecast horizon after a two-period lag.
  - Monetary factors explain 36 percent of the variation in the immediate horizon; real factors explain 14 percent.
  - Effect of real factors increases to 43 percent of the second-quarter forecast error in consumer prices; monetary factors also increase slightly.
  - In the long run, monetary factors account for 77 percent of the 12-quarter forecast mean square error of consumer prices.

### Augmented error-correction models (ECMs) and short-run dynamics
- Preferred augmented ECM incorporates seasonal dummies, lagged inflation, money growth, exchange rate changes, and the lagged error-correction term EC_{t-1} derived from the long-run equation.
  - General short-run ECM form (equation (7)) includes seasonal dummies dum_j, lagged ∆ln cpi terms, ∆ln m (money), ∆ln s (exchange rate), and lagged EC term.
- Estimation details and findings:
  - Two broad specifications estimated: with and without seasonal dummies; both assume AR(1) data generation for money growth and exchange rate changes in baseline estimations (alternative random walk specifications also considered in forecasting).
  - Real output was statistically insignificant in the short-run dynamics when seasonal dummies are included and was excluded from short-run regressors.
  - Joint nonlinear systems estimation (BFGS with exact line search) indicates:
    - It takes between five and six months for deviations of inflation from the long-run path to be corrected under both model variants.
    - Symmetry in the adjustment mechanism (speed of adjustment upward similar to downward across model variants).
    - Instantaneous pass-through: exchange rate depreciation impact on inflation is faster and larger than instantaneous monetary expansion.
      - A 1 percent money growth increase raises inflation by 0.1 percent under both models (insignificant in model excluding seasonal components).
      - Instantaneous exchange rate depreciation increases inflation by about 0.3 percent under both models.
- Hypothesis tests on data-generation processes for money and exchange rate (at 5 percent significance level):
  - Null (i): change in money follows a random walk with no drift (γ10 = 0, γ11 = 1) — rejected.
  - Null (ii): change in exchange rate follows a random walk with no drift (δ10 = 0, δ11 = 1) — rejected.
  - Null (iii): levels of money and exchange rate series follow AR(1) processes with no constants (γ10 = δ10 = 0) — not rejected.
  - Forecasting exercise adopts an agnostic approach: considers AR(1) and random walk specifications for money and exchange rate changes, yielding four forecast models.

### Selected augmented ECM numerical results (Table 4 highlights)
- CPI equation and money/exchange rate equation parameter estimates reported under two specifications (Without Seasonal Dummies / With Seasonal Dummies). Selected reported statistics:
  - Log Likelihood: 82.862 and 92.883.
  - Durbin's h-statistics: 0.7043* and 0.3186* (asterisk denotes statistical significance at the 5 percent level in the table note).
  - Hypothesis test chi-square statistics: 39.4729*, 39.4729*; 18.1046*, 18.1046*; and 4.7006, 4.7006 (degrees of freedom 2, 2, and 1 respectively); asterisk (*) indicates 5 percent significance for the first two.
- Interpretation summary from ECM estimates:
  - Error-correction speed implies 5–6 months to correct deviations.
  - Instantaneous exchange rate pass-through ≈ 0.3 percent inflation per 1 percent depreciation.
  - Instantaneous money-growth pass-through ≈ 0.1 percent inflation per 1 percent money growth.

*Source: Author's estimations and analysis in Section II and IV of the provided content.*

### 2.         Forecast         Performance

### 2.         Forecast         Performance

### Forecast performance metrics and out-of-sample comparison
- Measures used: mean absolute forecast error, root mean square forecast error (RMSE), and Theil’s U inequality coefficient.
- Table 5 — Out-of-Sample Inflation Forecasts (as presented)
  - Quarterly inflation (benchmarks and forecasts shown in table):  
    13.04 1.20 3.63 3.16 1.68 3.57 2.61 2.50  
    20.87 0.69 3.08 4.96 3.21 2.63 1.40 1.80  
    3 -0.90 -0.91 2.84 -0.01 -2.33 3.27 1.50 -0.90  
    41.44 1.60 4.14 4.56 2.15 3.96 2.21 1.70
  - RMSE0.270.431.222.271.132.511.27
  - Mean Absolute Error1.532.518.787.774.318.533.62
  - Theil's U-statistics0.150.240.691.280.641.420.72
  - Source line in table: "Source: Author's estimates."
- Principal comparative findings (textual)
  - Models that include seasonal components outperformed those without seasonal components using all three measures.
  - A random walk process for exchange rates outperforms an AR(1) process in forecasting inflation.
  - Models with seasonal dummies track seasonality and duplicate turning points better than those without seasonal components.
  - Example: models with seasonal dummies project negative inflation in Q3 (consistent with historical data); models without seasonal dummies predict positive inflation between 1.5 percent and 3.3 percent in Q3.

### The role of seasonality and money: simulation design and first-stage results
- Two-stage simulation design:
  - Stage 1: simulate estimated augmented error-correction model under varying assumptions on the autoregressive coefficient for money growth while keeping the exchange rate data generation process fixed (either AR(1) or random walk).
  - Stage 2: simulate the system using the optimal money-growth AR(1) coefficients to determine the exchange rate data generation process consistent with the predetermined inflation forecast.
- First-stage simulation outcomes (optimal money-growth AR(1) coefficients to attain an ex ante annual inflation forecast of 4 percent):
  - Under the model with seasonal dummies: optimal AR(1) coefficient for the money growth process is 0.7 for both AR(1) and random walk specifications for exchange rate movements.
  - Under the model excluding seasonal dummies: corresponding AR(1) coefficient for money growth is about 0.3 for both exchange rate data generation processes.
- Interpretation:
  - Inflation-expanding effects in Q2 and Q4 are higher under the model without seasonal components than with seasonal components.
  - Inflation-dampening effects in Q1 and Q3 are far higher under the model with seasonal components.
  - Implication: model without seasonal components implies a higher annual inflation and the need for a relatively tighter monetary stance (consistent with a lower AR(1) coefficient for money growth).

### Second-stage simulations: exchange rate path, monetary policy stance, and seasonality
- Key simulation inferences:
  - Simulations under the augmented EC model without seasonal components suggest bunching of positive quarterly inflation forecasts that do not disappear at any level of the AR(1) coefficient for nominal exchange rate changes—these are inconsistent with the historical seasonal pattern of inflation in the Kyrgyz Republic.
  - When seasonal components are included, simulation results are consistent with the historical pattern of inflation and its relationship with money and the exchange rate.
  - Conditional on an optimal money supply process with an AR(1) coefficient of 0.7, an AR(1) coefficient of below 0.4 for nominal exchange rate depreciation would be consistent with the observed inflation dynamics in the Kyrgyz Republic (as illustrated in Figure 4).
- Policy-stance implications for achieving a given ex ante inflation forecast:
  - Attaining a given inflation forecast requires a tighter monetary policy stance under the model without seasonal components than under the model that includes seasonal components.
  - Example for a 3 percent annual inflation forecast:
    - Under the seasonal-components model, the required exchange rate change path corresponds to an AR(1) coefficient at 0.3 for DS (vertical line AA in Figure 4).
    - Achieving the same forecast under the model without seasonal components would require a tighter monetary policy stance with a lower exchange rate depreciation (a lower AR(1) coefficient for exchange rate changes along vertical line BB in Figure 4).
  - Quarterly inflation figures along the illustrative lines:
    - Along line AA (seasonal-components model, 3-percent inflation line): Q1 = 3.1 percent, Q2 = 4.9 percent, Q3 = -0.2 percent, Q4 = 4.5 percent.
    - Along line BB (no seasonal components): Q1 = 3.4 percent, Q2 = 2.4 percent, Q3 = 2.9 percent, Q4 = 3.6 percent.
- Additional note on exchange rate specification:
  - Random walk exchange rate specification in the second-stage simulations yields lower quarterly and annual inflation forecasts under both model variants, consistent with the Table 5 results.

### Policy recommendations and concluding remarks
- Framework and applicability:
  - Adopt an approach that accounts for seasonality, the monetary policy stance, and implied exchange rate movements in an encompassing setup to inform monetary policy consistent with short-term inflation forecasts or targets.
  - The paper proposes an econometric "toolkit" for inflation forecasting that yields annual and quarterly inflation forecasts, corresponding monetary policy stance, and exchange rate movements consistent with that stance.
  - The framework is generally applicable to countries with pronounced seasonal fluctuations in economic activity and prices where authorities wish to determine monetary policy stance for short-term inflation forecasts or targets.
- Caveats and implementation considerations:
  - Efficacy of the approach depends critically on stability of the seasonal structure of consumer prices.
  - The approach may be subject to the "velocity instability problem" (the relationship between inflation and money may change over time), which mainly affects long-term forecasts more than short-term exercises.
  - Continual empirical updates and tests for stability of the money–inflation relationship are necessary before and during implementation.
- Operational note:
  - In practice, the two-stage procedure could be implemented simultaneously by monetary authorities determining foreign exchange market intervention and/or domestic open market operations (sterilized or nonsterilized) to yield the required monetary policy stance and corresponding exchange rate change consistent with the ex ante inflation forecast or target.

*Source: Author's estimates and simulations as presented in the content unit.*

---


_Source: https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp06175.pdf_