## tarea2025049-print-pdf

## Source details

**Canonical URL:** [tarea2025049-print-pdf](https://www.imf.org/-/media/files/publications/tar/2025/english/tarea2025049-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/tar/2025/english/tarea2025049-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/tar/2025/english/tarea2025049-print-pdf.pdf.json)

---

### A. Introduction
- TA provided to the National Statistics Office of Georgia (Geostat) during December 2–6, 2024 to extend scope and coverage of the residential property price index (RPPI) using web-scraped data.
- Mission objectives and outcomes:
  - Modernize index calculation process and update methods for index calculation.
  - Deploy a fully automated web-scraper and a new production system based on web-scraping www.ss.ge covering new and existing dwellings (flats and detached houses) for sale in Tbilisi.
  - Assist Geostat in integrating additional data sources and in developing hedonic imputation methods.
- Data-sharing and legal arrangements:
  - Geostat should establish a Memorandum of Understanding (MoU) with myhome.ge to agree data delivery and data structure (fields, names, temporal coverage, etc.).
  - Geostat should inquire with ss.ge whether their database can be shared directly.

### B. Tidying the Scraped Data
- Exchange rate treatment:
  - Scraped prices denominated in Georgian lari, or US dollar amounts converted using the end-of-month exchange rate from the National Bank of Georgia.
  - Decision: do not use daily exchange rates to avoid inducing artificial variability.
- Filtering and scope:
  - Houses and flats filtered where the condition indicates they are old with repairs.
  - Completion status filtered in the web-scraper to newly built or under construction.
  - For flats, project type restricted to “non-standard”. For houses, project type mostly missing and not used for filtering.
- Duplicate handling and outcomes (October and November 2024 / first two months of Q4 2024):
  - Perfect duplicates (exact matches across every field): always keep the newest record; removes 1,722 of 7,417 houses (23 percent) and 8,246 of 82,439 flats (10 percent).
  - Partial duplicates (two sets tested: numerical-focused; price + numerical + categorical-focused): further removes 1,529 of the remaining 5,695 houses (27 percent) and 37,866 of 74,193 flats (51 percent).
- Data relabeling and augmentation:
  - Re-label condition, completion status, project type, and district from Georgian to English.
  - Stratify districts into prestigious and non-prestigious, separately for houses and flats.
  - Add price per square meter variable.
  - Save tidied quarterly file for subsequent processing.

### C. Removing Outliers (and Errors)
- Definitions and approach:
  - Error: mistake or inaccuracy in data collection, entry, or processing; must be corrected or removed.
  - Outlier: observation significantly different from others; not all outliers are errors.
  - Use empirical, distribution-based thresholds (interquartile range method) per stratum and period to automate outlier detection.
- Method specifics:
  - Compute lower quartile and upper quartile; drop observations outside [lower quartile − multiple × IQR, upper quartile + multiple × IQR].
  - All calculations performed on the logarithmic scale.
  - Threshold multiples established experimentally:
    - Houses: multiple = 1.0 in prestigious districts; multiple = 2.0 in non-prestigious districts.
    - Flats: multiple = 3.0 in both strata.
  - Only gross outliers removed at this stage; hedonic regression step to follow with Cook’s distance for further influence diagnostics.
- Outlier removal results (quarterly level):
  - Houses, prestigious districts: out of 1,135 houses, 60 removed (five percent).
  - Houses, non-prestigious districts: out of 3,031 houses, 63 removed (two percent).
  - Flats, prestigious districts: out of 14,706 flats, 28 removed (0.2 percent).
  - Flats, non-prestigious districts: out of 21,620 flats, 137 removed (zero-point six percent).
  - Majority of outliers are lower outliers (price < lower whisker), likely data errors.
- Save dataset with outliers removed for next steps.

### D. Model Selection, Weights, and Medians
- Strata defined for hedonic models:
  - (i) houses in prestigious districts
  - (ii) houses in non-prestigious districts
  - (iii) flats in prestigious districts
  - (iv) flats in non-prestigious districts
- Pre-processing and variable handling:
  - Group condition into three frame types: “black frame” (need redecoration), “white frame” (currently being redecorated), and “green frame” ((newly) redecorated).
  - Filter numerical variables to allowable ranges using experimentally developed thresholds (e.g., floor area, number of rooms, area per room).
  - Address complications such as plot area for houses potentially referring to garden only.
  - Aggregate number of rooms for greater reliability.
  - Create new variable aggregating administrative districts into groups.
  - Note: R codes are separate for model selection, weights, medians, and hedonic imputation; calculations are repeated in each code.
- Model-selection considerations:
  - No single model dominates across all datasets; best model can vary by property type, district stratum, and data source.
  - If multiple models perform similarly, choose the simplest model (smallest number of predictors).
- Selected model characteristics (based on adjusted R2 and stepwise contribution):
  - Houses (both prestigious and non-prestigious): final model includes (i) floor area, (ii) plot area, (iii) administrative district, and (iv) condition. Time dummies will not be included in the model used for producing the RPPI.
  - Flats:
    - Prestigious districts: floor area plus (in order of relevance) administrative district, number of floors, and condition.
    - Non-prestigious districts: floor area plus (in order of relevance) condition, administrative district, and whether the flat has central heating.
  - Adjusted R2s of models using more characteristics were quite similar; simplicity preferred.
- Medians and publication:
  - Geostat publishes quarterly indices and median prices per square meter.
    - Newly built houses: medians broken down by ten administrative districts.
    - Newly built flats: medians broken down by ten administrative districts and by three frame types.
  - Improved methodology for calculating median prices will be explicitly documented; the previous time series will remain unchanged (no revisions).

### E. Index Compilation — Hedonic Approach, Weights, Aggregation, and Chaining
- Hedonic approach rationale and implementation:
  - Adjusts for varying mix of properties in the sample from period to period.
  - Allows calculation of a constant quality price index.
  - Applied hedonic methods to web-scraped data to adjust for property mix using a wide range of characteristics.
  - Implemented hedonic imputation approach (formulas and implementation detailed in Annex 2 of the source).
  - Finalized a fully automated production system that requires no manual cleaning or adjustment of scraped results.
  - The new index production system will be used for the first time with the publication of the RPPI for the first quarter of 2025.
- Hedonic regression specification and interpretation:
  - Log-price hedonic specification: ln p_i = β_0 + Σ_{k=1}^K z_{k,i} · β_k + ε_i
  - Use Cook’s distance as a second-layer detection of unusual observations to avoid undue influence from outliers or high-leverage observations.
  - Interpretation examples:
    - Log of floor/plot area: elasticity of price. Example: coefficient = 0.7 implies prices increase by 0.7 percent if floor/plot area increases by 1 percent.
    - Dummy coefficient example: coefficient = 0.2 implies prices are, on average, 100 × (exp(0.2) − 1) = 22 percent higher than the reference group.
- Hedonic Imputation (HI) index definition (period-specific regressions and imputation):
  - Period-specific hedonic regression: ln p_{i,t} = β_{0,t} + Σ_{k=1}^K z_{k,i,t} · β_{k,t} + ε_{i,t}
  - HI index I_HI 0:t = exp[ (1 / N_0) Σ_{i=1}^{N_0} ln p̂_i | z_0,t ] / exp[ (1 / N_0) Σ_{i=1}^{N_0} ln p̂_i | z_0,0 ]
  - N_0 refers to observations kept in the final hedonic regressions after Cook’s distance filtering.
- Weights, flows, and aggregation:
  - Explicit weights: calculated on the basis of the value of the houses and flats and reflect the relative importance of each stratum.
  - Flow weights: reflect changes from one period to the next (the value of all houses and flats purchased during a certain period); can be derived from the same source used for prices if reflective of overall market activity.
  - Best practice: update the weights annually and chain-link the indices.
  - Flow weights based on values can be derived by summing purchase prices (equivalent to number of purchases × average price).
  - Aggregation from strata to all-dwellings uses the Laspeyres formula with previous-year flow weights: I_L 4,y−1 : q,y = (Σ_{j=1}^4 w_{j,y−1} I_HI 4,y−1 : q,y) / (Σ_{j=1}^4 w_{j,y−1})
  - Chaining: Laspeyres index links in 2025 will be chained on the index last compiled using the old methodology in the fourth quarter of 2024; the index keeps the reference period 2020 = 100: I_LLL 2020 : q,2025 = I_LLL 2020 : 4,2024 I_L 4,2024 : q,2025
  - Chain index construction operational detail: for each year a series of five index numbers runs from the fourth quarter of the previous year (index = 100) to the fourth quarter of the current year; link these five-quarter series into a single long-term series.
- Operational workflow:
  - R code produces hedonic imputation indices for the four strata, weights, and medians per administrative district.
  - Aggregation and chaining steps are performed in Microsoft Excel; the updated spreadsheet continues this practice.

### F. Outlook and Next Steps
- Data source expansion and integration:
  - Integrate myhome.ge as a second source once MoU is in place; perform in-depth data analysis analogous to ss.ge steps.
  - Inquire with ss.ge about direct database sharing to simplify collection.
- Additional indices and indicators:
  - Analyze collected rental price information to determine if a rental price index can be established (mid-2025 priority).
  - Build valuation indicator price-to-rent ratio once rental price index is available.
  - Test inclusion of second-hand (old) properties in RPPI (end-2025 priority).
- Further research areas:
  - Integrate myhome.ge data into the RPPI once the MoU is in place and analyze it for inclusion as a second source.
  - Use scraped rental-price information to establish a rental price index and derive the valuation indicator price-to-rent ratio to help detect housing bubbles.
  - Investigate inclusion of old properties in the RPPI: address non-representative sample size, missing variables (e.g., age), and higher variability leading to unreliable results.
- Publication and documentation timeline:
  - Publish the RPPI based on the new index production system and update the technical documentation — target: April 2025.

### Priority Recommendations
- April 2025
  - Publish the RPPI based on the new index production system and update the technical documentation — Responsible: Geostat
- Early 2025
  - Negotiate an MoU with myhome.ge on the data delivery and data structure and enter into negotiations with ss.ge — Responsible: Geostat
- Mid-2025
  - Analyze the collected information on rental prices whether a rental price index can be established — Responsible: Geostat
- End-2025
  - Test the available information on old properties whether the data can be integrated in the RPPI — Responsible: Geostat

*IMF | Technical Assistance Report – Georgia Residential Property Price Index Mission — Section I. Detailed Technical Assessment and Recommendations*

### Section I. Detailed Technical Assessment and Recommendations ...................................................... 4

### Section I. Detailed Technical Assessment and Recommendations

### A. Introduction
- TA provided to the National Statistics Office of Georgia (Geostat) during December 2–6, 2024 to extend scope and coverage of the residential property price index (RPPI) using web-scraped data.
- Mission objectives and outcomes:
  - Modernize index calculation process and update methods for index calculation.
  - Deploy a fully automated web-scraper and a new production system based on web-scraping www.ss.ge covering new and existing dwellings (flats and detached houses) for sale in Tbilisi.
  - Assist Geostat in integrating additional data sources and in developing hedonic imputation methods.
- Data-sharing and legal arrangements:
  - Geostat should establish a Memorandum of Understanding (MoU) with myhome.ge to agree data delivery and data structure (fields, names, temporal coverage, etc.).
  - Geostat should inquire with ss.ge whether their database can be shared directly.

### B. Tidying the Scraped Data
- Exchange rate treatment:
  - Scraped prices denominated in Georgian lari, or US dollar amounts converted using the end-of-month exchange rate from the National Bank of Georgia.
  - Decision: do not use daily exchange rates to avoid inducing artificial variability.
- Filtering and scope:
  - Houses and flats filtered where the condition indicates they are old with repairs.
  - Completion status filtered in the web-scraper to newly built or under construction.
  - For flats, project type restricted to “non-standard”. For houses, project type mostly missing and not used for filtering.
- Duplicate handling and results (based on October and November 2024 / first two months of Q4 2024):
  - Perfect duplicates (exact matches across every field): always keep the newest record; removes 1,722 of 7,417 houses (23 percent) and 8,246 of 82,439 flats (10 percent).
  - Partial duplicates (two sets tested: numerical-focused; price + numerical + categorical-focused): further removes 1,529 of the remaining 5,695 houses (27 percent) and 37,866 of 74,193 flats (51 percent).
- Data relabeling and augmentation:
  - Re-label condition, completion status, project type, and district from Georgian to English.
  - Stratify districts into prestigious and non-prestigious, separately for houses and flats.
  - Add price per square meter variable.
  - Save tidied quarterly file for subsequent processing.

### C. Removing Outliers (and Errors)
- Definitions and approach:
  - Error: mistake or inaccuracy in data collection, entry, or processing; must be corrected or removed.
  - Outlier: observation significantly different from others; not all outliers are errors.
  - Use empirical, distribution-based thresholds (interquartile range method) per stratum and period to automate outlier detection.
- Method specifics:
  - Compute lower quartile and upper quartile; drop observations outside [lower quartile − multiple × IQR, upper quartile + multiple × IQR].
  - All calculations performed on the logarithmic scale.
  - Threshold multiples established experimentally:
    - Houses: multiple = 1.0 in prestigious districts; multiple = 2.0 in non-prestigious districts.
    - Flats: multiple = 3.0 in both strata.
  - Only gross outliers removed at this stage; hedonic regression step to follow with Cook’s distance for further influence diagnostics.
- Outlier removal results (quarterly level):
  - Houses, prestigious districts: out of 1,135 houses, 60 removed (five percent).
  - Houses, non-prestigious districts: out of 3,031 houses, 63 removed (two percent).
  - Flats, prestigious districts: out of 14,706 flats, 28 removed (0.2 percent).
  - Flats, non-prestigious districts: out of 21,620 flats, 137 removed (zero-point six percent).
  - Majority of outliers are lower outliers (price < lower whisker), likely data errors.
- Save dataset with outliers removed for next steps.

### D. Model Selection, Weights, and Medians
- Strata defined for hedonic models:
  - (i) houses in prestigious districts
  - (ii) houses in non-prestigious districts
  - (iii) flats in prestigious districts
  - (iv) flats in non-prestigious districts
- Pre-processing for model selection:
  - Group condition into three frame types: “black frame” (need redecoration), “white frame” (currently being redecorated), and “green frame” ((newly) redecorated).
  - Filter numerical variables to allowable ranges using experimentally developed thresholds (e.g., floor area, number of rooms, area per room).
  - Address complications such as plot area for houses potentially referring to garden only.
  - Aggregate number of rooms for greater reliability.
  - Create new variable aggregating administrative districts into groups.
  - Note: R codes are separate for model selection, weights, medians, and hedonic imputation; calculations are repeated in each code.
- Model-selection considerations:
  - No single model dominates across all datasets; best model can vary by property type, district stratum, and data source.
  - If multiple models perform similarly, choose the simplest model (smallest number of predictors).
- Selected model characteristics (based on adjusted R2 and stepwise contribution):
  - Houses (both prestigious and non-prestigious): final model includes (i) floor area, (ii) plot area, (iii) administrative district, and (iv) condition. Time dummies will not be included in the model used for producing the RPPI.
  - Flats:
    - Prestigious districts: floor area plus (in order of relevance) administrative district, number of floors, and condition.
    - Non-prestigious districts: floor area plus (in order of relevance) condition, administrative district, and whether the flat has central heating.
  - Adjusted R2s of models using more characteristics were quite similar; simplicity preferred.

### E. Index Compilation–Hedonic Approach
- Hedonic approach rationale:
  - Adjusts for varying mix of properties in the sample from period to period.
  - Allows calculation of a constant quality price index.
- Implementation highlights from mission:
  - Applied hedonic methods to web-scraped data to adjust for property mix using a wide range of characteristics.
  - Implemented hedonic imputation approach (details and code separated; see Annex 2 referenced in source).
  - Finalized a fully automated production system that requires no manual cleaning or adjustment of scraped results.
- Operational note:
  - The new index production system will be used for the first time with the publication of the RPPI for the first quarter of 2025.

### F. Outlook and Next Steps
- Data source expansion and integration:
  - Integrate myhome.ge as a second source once MoU is in place; perform in-depth data analysis analogous to ss.ge steps.
  - Inquire with ss.ge about direct database sharing to simplify collection.
- Additional indices and indicators:
  - Analyze collected rental price information to determine if a rental price index can be established (mid-2025 priority).
  - Build valuation indicator price-to-rent ratio once rental price index is available.
  - Test inclusion of second-hand (old) properties in RPPI (end-2025 priority).
- Publication and documentation:
  - Publish the RPPI based on the new index production system and update the technical documentation (target: April 2025).

### Priority Recommendations (as summarized)
- April 2025
  - Publish the RPPI based on the new index production system and update the technical documentation — Responsible: Geostat
- Early 2025
  - Negotiate an MoU with myhome.ge on the data delivery and data structure and enter into negotiations with ss.ge — Responsible: Geostat
- Mid-2025
  - Analyze the collected information on rental prices whether a rental price index can be established — Responsible: Geostat
- End-2025
  - Test the available information on old properties whether the data can be integrated in the RPPI — Responsible: Geostat

### G. Officials Met During The Mission
- (List of officials referenced in source; details provided in the original report section.)

*IMF | Technical Assistance Report – Georgia Residential Property Price Index Mission — Section I. Detailed Technical Assessment and Recommendations*

### 27. Before turning to the index calculation based on these model selection results, weights

### tarea2025049-print-pdf - 27. Before turning to the index calculation based on these model selection results, weights

### Weights, flows, and aggregation
- Explicit weights: calculated on the basis of the value of the houses and flats and reflect the relative importance of each stratum.
- Flow weights: reflect changes from one period to the next (the value of all houses and flats purchased during a certain period); can be derived from the same source used for prices if reflective of overall market activity.
- Best practice: update the weights annually and chain-link the indices.
- Flow weights based on values can be derived by summing purchase prices (equivalent to number of purchases × average price).
- Flow weights are typically not restricted to one year of data; recommended to update annually because the relative importance of strata varies over time.
- Aggregation from strata to all-dwellings uses the Laspeyres formula with previous-year flow weights:
  - I_L 4,y−1 : q,y = (Σ_{j=1}^4 w_{j,y−1} I_HI 4,y−1 : q,y) / (Σ_{j=1}^4 w_{j,y−1})
- Chaining: Laspeyres index links in 2025 will be chained on the index last compiled using the old methodology in the fourth quarter of 2024; the index keeps the reference period 2020 = 100:
  - I_LLL 2020 : q,2025 = I_LLL 2020 : 4,2024 I_L 4,2024 : q,2025

### Medians, outliers, and data handling
- Geostat publishes quarterly indices and median prices per square meter.
  - Newly built houses: medians broken down by ten administrative districts.
  - Newly built flats: medians broken down by ten administrative districts and by three frame types.
- Outlier identification:
  - Previously performed on the linear scale; now performed on the logarithmic scale for houses and flats by prestigious and non-prestigious districts.
  - Due to positive skew of prices per square meter, linear-scale IQR filtering can disproportionately remove high-value observations and affect medians.
  - The effect of change in outlier identification is exemplified using an IMF synthetic dataset for training/education.
- Improved methodology for calculating median prices will be explicitly documented; the previous time series will remain unchanged (no revisions). The enhanced index method will be detailed in the updated methodological manual to be made available on the Geostat website.

### Index compilation — hedonic approach (policy recommendation)
- Price indices are generally measured on a matched-model basis; quality adjustments are used to obtain constant-quality property price indices despite heterogeneity and infrequent sales.
- Recommendation: Authorities should use the Hedonic Imputation Approach for index compilation (method outlined in Annex 2).
- Hedonic regressions treat a dwelling as a bundle of characteristics and estimate marginal valuations (shadow prices) for characteristics.
- Log-price hedonic specification (as presented):
  - ln p_i = β_0 + Σ_{k=1}^K z_{k,i} · β_k + ε_i
  - β_k are estimates of shadow prices (marginal valuations).
- After fitting hedonic regressions, use Cook’s distance as a second-layer detection of unusual observations to avoid undue influence from outliers or high-leverage observations.

### Model selection (Annex 1) — approach and operational choices
- Model-fit considerations:
  - R^2 increases with added variables; adjusted R^2 penalizes unnecessary variables.
  - Adjusted R^2 formula used:
    - Adjusted R^2 = 1 − [RSS / (N − K − 1)] / [TSS / (N − 1)]
  - Forward selection (stepwise) is a classical approach; forward stepwise selection fits 1 + K × (K + 1) / 2 models.
- Practical adaptation for categorical characteristics:
  - Characteristics that are represented by multiple dummy variables (e.g., frame types represented by two dummies) are treated as a block: include all dummies for a characteristic or exclude the characteristic entirely.
  - Adjusted R^2 used to select the best model when adding one characteristic (accounts for differing numbers of variables added).
- Parsimony guidance: where adjusted R^2 curves are flat, choose the simplest model (the “elbow” rule) though it is ad hoc.
- Baseline inclusion:
  - Log of floor area is always included first in all models; for houses, log of plot area is also always included.
  - Time dummy parameters included to measure the effect of “time” on log price.
  - As a consequence, baseline model for houses includes three characteristics, baseline model for flats includes two.

### Hedonic imputation approach (Annex 2) — formulas and implementation
- Period-specific hedonic regression form:
  - ln p_{i,t} = β_{0,t} + Σ_{k=1}^K z_{k,i,t} · β_{k,t} + ε_{i,t}
- Hedonic Imputation (HI) index from period 0 to t defined as ratio of average predicted prices in periods t and 0 using period 0 characteristics (preserving the source equations):
  - I_HI 0:t = exp[ (1 / N_0) Σ_{i=1}^{N_0} ln p̂_i | z_0,t ] / exp[ (1 / N_0) Σ_{i=1}^{N_0} ln p̂_i | z_0,0 ]
  - Expanded in the source text showing substitution with estimated coefficients β̂_{0,t}, β̂_{k,t}, and the identity linking average predicted prices to average actual prices when using same-period characteristics.
- Notes on sample adjustment:
  - Because Cook’s distance was used to filter influential observations, N_0 refers to observations kept in the final hedonic regressions and the corresponding estimated coefficients.
- Interpretation of regression parameters:
  - Log of floor/plot area: elasticity of price. Example: coefficient = 0.7 implies prices increase by 0.7 percent if floor/plot area increases by 1 percent (and vice versa); price per square foot decreases by 0.3 percent in this scenario.
  - Dummies on administrative districts / condition / number of floors / central heating: average price level difference versus the reference group. Example: coefficient = 0.2 implies prices are, on average, 100 × (exp(0.2) − 1) = 22 percent higher than the reference group.

### Chaining, indices, and operational workflow
- Chain index construction:
  - Update weights annually and link overlapping shorter period sequences (one-quarter overlap technique): for each year a series of five index numbers runs from the fourth quarter of the previous year (index = 100) to the fourth quarter of the current year; link these five-quarter series into a single long-term series.
  - Imputation for year-specific series uses the fourth quarter of the previous year as reference period; imputation sample z_{4,y−1} is adjusted for influential observations using Cook’s distance.
  - The mission calculated for each year a series of five index numbers using the fourth quarter of the previous year as reference (I_HI 4,y−1 : q,y).
- Operational implementation:
  - R code produces hedonic imputation indices for the four strata, weights, and medians per administrative district.
  - Aggregation and chaining steps are performed in Microsoft Excel; the updated spreadsheet continues this practice.

### Outlook and next steps (policy/action items)
- Three areas for further research identified:
  - Integrate myhome.ge data into the RPPI once the MoU is in place and analyze it for inclusion as a second source.
  - Use scraped rental-price information to establish a rental price index and derive the valuation indicator price-to-rent ratio to help detect housing bubbles.
  - Investigate inclusion of old properties in the RPPI: address non-representative sample size, missing variables (e.g., age), and higher variability leading to unreliable results.

*IMF | Technical Assistance Report –   Georgia Residential Property Price Index Mission*

---


_Source: https://www.imf.org/-/media/files/publications/tar/2025/english/tarea2025049-print-pdf.pdf_