## 1geoea2020002

## Source details

**Canonical URL:** [1geoea2020002](https://www.imf.org/-/media/files/publications/cr/2020/english/1geoea2020002.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/cr/2020/english/1geoea2020002.pdf.md)
- [Structured JSON version](/-/media/files/publications/cr/2020/english/1geoea2020002.pdf.json)

---

### Summary of mission outcomes and priority recommendations
- Mission purpose: assist National Statistics Office (Geostat) in compiling an experimental residential property price index (RPPI) for Tbilisi; mission completed on March 16, 2020.
- Coverage target: quarterly RPPI covering new flats and new detached houses for the capital city, Tbilisi; recommendation to obtain secondary market and rest-of-country data and store for future expansion.
- Data sources obtained and status:
  - Two web sites with advertised properties for sale; data from first site available since 4Q2018; web scraping improved after 2Q2019 yielding more observations and variables.
  - National Agency of Public Registry (NAPR) data automatically transmitted from October 2019 covering only Tbilisi (flats and land plots with unspecified buildings); cannot distinguish primary vs. secondary market in transmitted data; dwelling characteristics in NAPR data are very limited.
- Methods implemented: R programs based on IMF’s draft RPPI Practical Compilation Guide; three hedonic methods compiled and should continue being tested during 2019; Geostat staff capable of adapting R scripts.
- Exchange rate sensitivity: house prices are set in dollars although published mostly in local currency; 15 per cent depreciation of the Georgian currency to the dollar likely contributed to RPPI increases in 3Q2019.
- Priority recommendations and targets:
  - By December/2020: Compile the RPPI with data from the two websites and assess the results. (Responsible: Geostat)
  - By December/2020: Continue negotiations with NAPR to improve the survey. (Responsible: Geostat)
  - By October/2019: Obtain data for the secondary market. (Responsible: Geostat)

### Detailed technical assessment and recommendations — A. Governance
- Index structure and stratification:
  - RPPI breakdown recommended to include: Apartments (Prestigious, Non-Prestigious) and Detached houses (Prestigious, Non-Prestigious).
  - Location stratification by clustering 44 district areas into prestigious vs. non-prestigious using median price per square meter cut-off and coefficient of variation outlier detection.
- Observations and value share for Flats by quarter and stratum:
  - 4Q2018:
    - Prestigious — No. Obs 2091, Total value 661,362,304, Value Share 0.72
    - Non-Prestigious — No. Obs 1851, Total value 251,898,956, Value Share 0.28
  - 1Q2019:
    - Prestigious — No. Obs 2926, Total value 2,460,546,608, Value Share 0.89
    - Non-Prestigious — No. Obs 2326, Total value 311,423,711, Value Share 0.11
  - 2Q2019:
    - Prestigious — No. Obs 8061, Total value 3,458,924,804, Value Share 0.80
    - Non-Prestigious — No. Obs 6363, Total value 843,508,634, Value Share 0.20
  - 3Q2019:
    - Prestigious — No. Obs 10694, Total value 3,458,924,804, Value Share 0.76
    - Non-Prestigious — No. Obs 7241, Total value 1,104,995,366, Value Share 0.24
  - Total annual:
    - Prestigious — 10,039,758,520 Value Share 0.80
    - Non-Prestigious — 2,511,826,667 Value Share 0.20
- Data management and staff:
  - A new staff member added; four statisticians working on RPPI (head of prices unit + three staff who also work on CPI and PPI).
  - File management framework developed; multiple methods and sources generate multiple file versions — backups and comprehensive schema recommended (Annex 1 referenced).
- Data merging guidance:
  - Merge two web site datasets for flats (similar structure) and consider separate models or sub-indexes for detached houses if extra variables from second site are significant; aggregate sub-indexes with weighted average using total value of dwellings; remove duplicates.
- NAPR data concerns and action:
  - NAPR data cannot distinguish primary vs. secondary market; Geostat to contact NAPR to clarify content and negotiate improved variables.
- Methodology stability:
  - Model and methodology should be kept fixed for at least one year; recommended stability for five years.
  - Weights (average characteristics, base price, or dwellings sample depending on method) should be kept fixed for at least one year (ideally five) and updated every first quarter.
- Model selection diagnostics:
  - Indicator — Detached vs. Flats:
    - AIC: Detached 2007, Flats 3117
    - R^2: Detached 70, Flats 77
    - RRSE: Detached 0.54, Flats 0.28
  - Stratum Flats performs better due to much higher number of observations.

### Detailed technical assessment and recommendations — B. RPPI for Flats
- Data cleaning and exclusions:
  - Limits applied: number of rooms between 1 and 7; maximum number of floors set to 40.
  - Excluded observations: around 28.6 percent of observations were ruled out in the overall Flats cleaning process (specific by quarter below).
- Observations and variables for Flats:
  - 4Q2018: No. Obs before cleaning 3942; No. Obs after cleaning 3345; % of excluded Obs 15,1; Total no. of variables 11
  - 1Q2019: No. Obs before cleaning 5252; No. Obs after cleaning 4778; % of excluded Obs 9,0; Total no. of variables 11
  - 2Q2019: No. Obs before cleaning 14424; No. Obs after cleaning 12344; % of excluded Obs 14,4; Total no. of variables 20
  - 3Q2019: No. Obs before cleaning 17935; No. Obs after cleaning 12809; % of excluded Obs 28,6; Total no. of variables 20
- Thresholds and categorization for Flats:
  - Number of Rooms categories: [1,2], [3,4], [5,7], > 17
  - Number of Floors categories: [1,4], [5,9], [10,16], > 17
- Data diagnostics:
  - Histograms and outlier treatments performed for prestigious and non-prestigious sub-strata (4Q2018 examples shown in mission figures).
- RPPI movement for Flats:
  - RPPI for strata within Flats increased particularly in 3Q2019, likely reflecting 15 per cent depreciation of the Georgian currency to the dollar and dollar-denominated house prices.

### Detailed technical assessment and recommendations — C. RPPI for Detached Houses
- Data cleaning and exclusions:
  - Limits applied: number of rooms limited to 15; area per room limited to 200 sqm; land area between 15 and 6000.
  - Excluded observations: around 30 per cent of observations were ruled out overall, indicating data quality filtering.
- Observations and variables for Detached:
  - 4Q2018: No. Obs before cleaning 1045; No. Obs after cleaning 695; % of excluded Obs 33,5; Total no. of variables 10
  - 1Q2019: No. Obs before cleaning 1212; No. Obs after cleaning 857; % of excluded Obs 28,9; Total no. of variables 10
  - 2Q2019: No. Obs before cleaning 2335; No. Obs after cleaning 1556; % of excluded Obs 33,0; Total no. of variables 19
  - 3Q2019: No. Obs before cleaning 2522; No. Obs after cleaning 1741; % of excluded Obs 30,1; Total no. of variables 19
- Thresholds and categorization for Detached:
  - Number of Rooms categories: [1,5], [6,10], [10,15]
- RPPI movement for Detached:
  - RPPI for strata within Detached increased particularly in 3Q2019 (figures and sub-indices presented in mission materials).

### Detailed technical assessment and recommendations — D. Scanner Data
- Role: Mission provided guidance on use of scanner data (SD) for CPI compilation at Geostat Director’s request.
- Introduction approach:
  - Introduce SD on a step-wise approach to avoid huge impacts on CPI and to make the process manageable, reliable, and safe.
  - Prioritize food products as first group for SD coverage because electronic systems are well developed in food retailers and they cover a significant part of price surveys, avoiding complex quality adjustment issues.
- Engagement with data providers:
  - Negotiations can take a long time; first meeting aims to engage providers rather than rely immediately on statistical law.
  - On first meeting Geostat should: inform of progress in other countries, assure confidentiality, assure data used only for CPI and not shared with other authorities, assure that retailer-specific details and price levels will not be published. In return Geostat can offer a personalized report.
  - To guarantee confidentiality, involve at least two retailers.
- Data request and frequency:
  - Request two years of back data of all products.
  - Data should be received automatically weekly for at least the first two weeks of each month.
- Recommended SD data structure:
  - EAN/GTIN code
  - Retailer category code
  - Retailer category label
  - Item label
  - Sales/ turnover
  - Number of units sold/Quantity
  - Unit size (kg, lt)
  - Package size
  - Reference period (Year, month)
- Classification challenge:
  - Mapping scanner items to COICOP is most challenging; build a learning data set by manual classification of received retailer datasets.

### Section 2 — Methodology for use of scanner (SD) data in RPPI
- Proposed methodology: matched model approach with a dynamic sample.
- Implementation steps:
  - Merge current month data with previous month to identify products available in both months.
  - Sampling is made using the share of each item.
  - Apply a filter for outliers and dumping prices.
  - Impute these prices as well as the missing prices, for 14 months.
  - Compile sub-indexes by retailer at the lowest level of the COICOP, starting from monthly changes of items available in two consecutive months; then multiply by the index–based December—of the previous month; finally chain by multiplying by the index–based 2010—of the previous month.
  - Aggregate retailers with turnover weights to obtain an SD index per retailer per lower level COICOP that is afterward aggregated with the same level sub-index obtained by the field survey.
  - Products covered by scanner will no longer be included in the field price collection for those retailers.
- Implementation note:
  - Methodology is rather simple to implement and takes good advantage of the SD.
  - While a machine learning process for the use of semantic data is not developed, Geostat should focus on classifying items with high turnover.

### Priority action plan (one-year) to improve the RPPI — priority recommendations and target dates
- Outcome: Experimental RPPI is compiled
- Priority Action/Milestone — Target Completion Date (priority code as in source)
  - H Compile the RPPI with data from the two websites to assess the results. — December 2020
  - H Continue negotiations with NAPR to improve the survey — December 2020
  - H Obtain all data including secondary market and rest for the country — October 2019
  - M Built a comprehensive files management framework — October 2019
  - M Methodology should be kept fixed for at least one year — December 2020
  - M Merge data from the two websites or compile a sub-index for each — January 2020

### Officials met during the mission
- Gogita Todradze — Executive Director, Geostat
- Giorgi Tetrauli — Head of Price Statistics Department, Geostat
- Khatuna Aptsiauri — Head of Consumer Price Statistics Division
- Revaz Maisuradze — Senior Specialist, Consumer Price Statistics Division, Geostat

### Appendix 1. Files management (selected filenames and structure as presented)
- RPPI 2019
  - IndicesFlatsDetachedNAPR
  - Indices
  - All_indices_Flats_imputations.csv
  - All_indices_Detached_imputations.csv
  - All_indices_imputations.csv
- Flats (structure highlights)
  - Raw data: Q2019.xlsx (web scrape)
  - Output: Data_clean_excel; 1Q2019_Flats.csv
  - Weights.r
  - Data (inputs/intermediate outputs): Weights_2019.csv; Clean data (1Q2019_clean.csv); Results (All_indices_Flats_imputations.csv); Base_coef.csv; Base_Interp.csv; Base_price.csv
  - Sripts: Analysis.r; Imputations.r; Characteristics.r
- Detached (structure highlights)
  - Raw data: Q2019.xlsx (web scrape)
  - Output: Data_clean_excel; 1Q2019_Flats.csv
  - Weights.r
  - Data (inputs/intermediate outputs): Weights_2019.csv; Clean data (1Q2019_clean.csv); Results (All_indices_Detached_imputations.csv); Base_coef.csv; Base_Interp.csv; Base_price.csv
  - Sripts: Analysis.r; Imputations.r; Characteristics.r
- NAPR
  - Data (inputs/intermediate outputs): Clean data (1Q2019_clean.csv); Results (All_indices.csv)
  - Sripts: Analysis.r; Stratification.r

*Prepared by Vanda Guerreiro; Technical Assistance Report on Residential Property Price Indices—Georgia (mission September 23–October 4, 2019); report completed March 16, 2020.*

### Section 1

### 1geoea2020002 - Section 1

### Summary of mission outcomes and priority recommendations
- Mission purpose: assist National Statistics Office (Geostat) in compiling an experimental residential property price index (RPPI) for Tbilisi; mission completed on March 16, 2020.
- Coverage target: quarterly RPPI covering new flats and new detached houses for the capital city, Tbilisi; recommendation to obtain secondary market and rest-of-country data and store for future expansion.
- Data sources obtained and status:
  - Two web sites with advertised properties for sale; data from first site available since 4Q2018; web scraping improved after 2Q2019 yielding more observations and variables.
  - National Agency of Public Registry (NAPR) data automatically transmitted from October 2019 covering only Tbilisi (flats and land plots with unspecified buildings); cannot distinguish primary vs. secondary market in transmitted data; dwelling characteristics in NAPR data are very limited.
- Methods implemented: R programs based on IMF’s draft RPPI Practical Compilation Guide; three hedonic methods compiled and should continue being tested during 2019; Geostat staff capable of adapting R scripts.
- Exchange rate sensitivity: house prices are set in dollars although published mostly in local currency; 15 per cent depreciation of the Georgian currency to the dollar likely contributed to RPPI increases in 3Q2019.
- Priority recommendations and targets (Table 1):
  - By December/2020: Compile the RPPI with data from the two websites and assess the results. (Responsible: Geostat)
  - By December/2020: Continue negotiations with NAPR to improve the survey. (Responsible: Geostat)
  - By October/2019: Obtain data for the secondary market. (Responsible: Geostat)

### Detailed technical assessment and recommendations — A. Governance
- Index structure and stratification:
  - RPPI breakdown recommended to include: Apartments (Prestigious, Non-Prestigious) and Detached houses (Prestigious, Non-Prestigious).
  - Location stratification by clustering 44 district areas into prestigious vs. non-prestigious using median price per square meter cut-off and coefficient of variation outlier detection.
- Observations and value share for Flats by quarter and stratum (Table 2):
  - 4Q2018: Prestigious — No. Obs 2091, Total value 661,362,304, Value Share 0.72; Non-Prestigious — No. Obs 1851, Total value 251,898,956, Value Share 0.28
  - 1Q2019: Prestigious — No. Obs 2926, Total value 2,460,546,608, Value Share 0.89; Non-Prestigious — No. Obs 2326, Total value 311,423,711, Value Share 0.11
  - 2Q2019: Prestigious — No. Obs 8061, Total value 3,458,924,804, Value Share 0.80; Non-Prestigious — No. Obs 6363, Total value 843,508,634, Value Share 0.20
  - 3Q2019: Prestigious — No. Obs 10694, Total value 3,458,924,804, Value Share 0.76; Non-Prestigious — No. Obs 7241, Total value 1,104,995,366, Value Share 0.24
  - Total annual: Prestigious — 10,039,758,520 Value Share 0.80; Non-Prestigious — 2,511,826,667 Value Share 0.20
- Data management and staff:
  - A new staff member added; four statisticians working on RPPI (head of prices unit + three staff who also work on CPI and PPI).
  - File management framework developed; multiple methods and sources generate multiple file versions — backups and comprehensive schema recommended (Annex 1 referenced).
- Data merging guidance:
  - Merge two web site datasets for flats (similar structure) and consider separate models or sub-indexes for detached houses if extra variables from second site are significant; aggregate sub-indexes with weighted average using total value of dwellings; remove duplicates.
- NAPR data concerns and action:
  - NAPR data cannot distinguish primary vs. secondary market; Geostat to contact NAPR to clarify content and negotiate improved variables.
- Methodology stability:
  - Model and methodology should be kept fixed for at least one year; recommended stability for five years.
  - Weights (average characteristics, base price, or dwellings sample depending on method) should be kept fixed for at least one year (ideally five) and updated every first quarter.

- Model selection diagnostics (Table 3):
  - Indicator — Detached vs. Flats:
    - AIC: Detached 2007, Flats 3117
    - R^2: Detached 70, Flats 77
    - RRSE: Detached 0.54, Flats 0.28
  - Stratum Flats performs better due to much higher number of observations.

### Detailed technical assessment and recommendations — B. RPPI for Flats
- Data cleaning and exclusions:
  - Limits applied: number of rooms between 1 and 7; maximum number of floors set to 40.
  - Excluded observations: around 28.6 percent of observations were ruled out in the overall Flats cleaning process (specific by quarter in Table 4).
- Observations and variables for Flats (Table 4):
  - 4Q2018: No. Obs before cleaning 3942; No. Obs after cleaning 3345; % of excluded Obs 15,1; Total no. of variables 11
  - 1Q2019: No. Obs before cleaning 5252; No. Obs after cleaning 4778; % of excluded Obs 9,0; Total no. of variables 11
  - 2Q2019: No. Obs before cleaning 14424; No. Obs after cleaning 12344; % of excluded Obs 14,4; Total no. of variables 20
  - 3Q2019: No. Obs before cleaning 17935; No. Obs after cleaning 12809; % of excluded Obs 28,6; Total no. of variables 20
- Thresholds and categorization for Flats (Table 5):
  - Number of Rooms categories: [1,2], [3,4], [5,7], > 17
  - Number of Floors categories: [1,4], [5,9], [10,16], > 17
- Data diagnostics:
  - Histograms and outlier treatments performed for prestigious and non-prestigious sub-strata (4Q2018 examples shown in mission figures).
- RPPI movement for Flats:
  - RPPI for strata within Flats increased particularly in 3Q2019, likely reflecting 15 per cent depreciation of the Georgian currency to the dollar and dollar-denominated house prices.

### Detailed technical assessment and recommendations — C. RPPI for Detached Houses
- Data cleaning and exclusions:
  - Limits applied: number of rooms limited to 15; area per room limited to 200 sqm; land area between 15 and 6000.
  - Excluded observations: around 30 per cent of observations were ruled out overall, indicating data quality filtering.
- Observations and variables for Detached (Table 6):
  - 4Q2018: No. Obs before cleaning 1045; No. Obs after cleaning 695; % of excluded Obs 33,5; Total no. of variables 10
  - 1Q2019: No. Obs before cleaning 1212; No. Obs after cleaning 857; % of excluded Obs 28,9; Total no. of variables 10
  - 2Q2019: No. Obs before cleaning 2335; No. Obs after cleaning 1556; % of excluded Obs 33,0; Total no. of variables 19
  - 3Q2019: No. Obs before cleaning 2522; No. Obs after cleaning 1741; % of excluded Obs 30,1; Total no. of variables 19
- Thresholds and categorization for Detached (Table 7):
  - Number of Rooms categories: [1,5], [6,10], [10,15]
- RPPI movement for Detached:
  - RPPI for strata within Detached increased particularly in 3Q2019 (figures and sub-indices presented in mission materials).

### Detailed technical assessment and recommendations — D. Scanner Data
- Role: Mission provided guidance on use of scanner data (SD) for CPI compilation at Geostat Director’s request.
- Introduction approach:
  - Introduce SD on a step-wise approach to avoid huge impacts on CPI and to make the process manageable, reliable, and safe.
  - Prioritize food products as first group for SD coverage because electronic systems are well developed in food retailers and they cover a significant part of price surveys, avoiding complex quality adjustment issues.
- Engagement with data providers:
  - Negotiations can take a long time; first meeting aims to engage providers rather than rely immediately on statistical law.
  - On first meeting Geostat should: inform of progress in other countries, assure confidentiality, assure data used only for CPI and not shared with other authorities, assure that retailer-specific details and price levels will not be published. In return Geostat can offer a personalized report.
  - To guarantee confidentiality, involve at least two retailers.
- Data request and frequency:
  - Request two years of back data of all products.
  - Data should be received automatically weekly for at least the first two weeks of each month.
- Recommended SD data structure:
  - EAN/GTIN code
  - Retailer category code
  - Retailer category label
  - Item label
  - Sales/ turnover
  - Number of units sold/Quantity
  - Unit size (kg, lt)
  - Package size
  - Reference period (Year, month)
- Classification challenge:
  - Mapping scanner items to COICOP is most challenging; build a learning data set by manual classification of received retailer datasets.

*Prepared by Vanda Guerreiro; Technical Assistance Report on Residential Property Price Indices—Georgia (mission September 23–October 4, 2019); report completed March 16, 2020.*

### Section 2

### 1geoea2020002 - Section 2

### Methodology for use of scanner (SD) data in RPPI
- The methodology proposed by the mission is a matched model approach with a dynamic sample.
- Implementation steps as described:
  - The data of the current month is merged with the previous month to identify the products available in both months.
  - Sampling is made using the share of each item.
  - A filter for outliers and dumping prices is applied.
  - These prices are imputed as well as the missing prices, for 14 months.
  - Sub-indexes by retailer are compiled, at the lowest level of the COICOP, starting from monthly changes of the items available in two consecutive months; then multiply by the index–based December—of the previous month; finally chain by multiplying by the index–based 2010—of the previous month.
  - The retailers are aggregated with the turnover weights to obtain an SD index per retailer per lower level COICOP that is afterward aggregated with the same level sub-index obtained by the field survey.
  - The products covered by scanner will no longer be included in the field price collection for those retailers.
- The methodology is noted as rather simple to implement and takes good advantage of the SD.
- While a machine learning process for the use of semantic data is not developed, Geostat should focus on classifying items with high turnover.

### Priority action plan (one-year) to improve the RPPI — priority recommendations and target dates
- Outcome: Experimental RPPI is compiled
- Priority Action/Milestone — Target Completion Date (priority code as in source)
  - H Compile the RPPI with data from the two websites to assess the results. — December 2020
  - H Continue negotiations with NAPR to improve the survey — December 2020
  - H Obtain all data including secondary market and rest for the country — October 2019
  - M Built a comprehensive files management framework — October 2019
  - M Methodology should be kept fixed for at least one year — December 2020
  - M Merge data from the two websites or compile a sub-index for each — January 2020

### Officials met during the mission
- Gogita Todradze — Executive Director, Geostat
- Giorgi Tetrauli — Head of Price Statistics Department, Geostat
- Khatuna Aptsiauri — Head of Consumer Price Statistics Division
- Revaz Maisuradze — Senior Specialist, Consumer Price Statistics Division, Geostat

### Appendix 1. Files management (selected filenames and structure as presented)
- RPPI 2019
  - IndicesFlatsDetachedNAPR
  - Indices
  - All_indices_Flats_imputations.csv
  - All_indices_Detached_imputations.csv
  - All_indices_imputations.csv
- Flats
  - Raw data
    - Q2019.xlsx (web scrape)
    - (...)
  - Output
    - Data_clean_excel
    - 1Q2019_Flats.csv
    - (....csv)
  - Weights.r
  - Few_variables
  - Data (inputs/intermediate outputs)
    - Weights_2019.csv
    - Clean data (1Q2019_clean.csv)
    - Results (All_indices_Flats_imputations.csv)
    - Base_coef.csv
    - Base_Interp.csv
    - Base_price.csv
  - Sripts
    - Analysis.r
    - Imputations.r
    - Characteristics.r
    - (....r)
  - All_variables
    - Data (inputs/intermediate outputs)
    - Weights_2019.csv
    - Clean data (1Q2019_clean.csv)
    - Results (All_indices_Flats_imputations.csv)
    - Base_coef.csv
    - Base_Interp.csv
    - Base_price.csv
    - Sripts
      - Analysis.r
      - Imputations.r
      - Characteristics.r
      - (....r)
- Detached
  - Raw data
    - Q2019.xlsx (web scrape)
    - (...)
  - Output
    - Data_clean_excel
    - 1Q2019_Flats.csv
    - (....csv)
  - Weights.r
  - Few_variables
  - Data (inputs/intermediate outputs)
    - Weights_2019.csv
    - Clean data (1Q2019_clean.csv)
    - Results (All_indices_Detached_imputations.csv)
    - Base_coef.csv
    - Base_Interp.csv
    - Base_price.csv
  - Sripts
    - Analysis.r
    - Imputations.r
    - Characteristics.r
    - (....r)
  - All_variables
    - Data (inputs/intermediate outputs)
    - Weights_2019.csv
    - Clean data (1Q2019_clean.csv)
    - Results (All_indices_Detached_imputations.csv)
    - Base_coef.csv
    - Base_Interp.csv
    - Base_price.csv
    - Sripts
      - Analysis.r
      - Imputations.r
      - Characteristics.r
      - (....r)
- NAPR
  - Data (inputs/intermediate outputs)
    - Clean data (1Q2019_clean.csv)
    - Results (All_indices.csv)
  - Sripts
    - Analysis.r
    - Stratification.r

*Source: 1geoea2020002 - Section 2*

---


_Source: https://www.imf.org/-/media/files/publications/cr/2020/english/1geoea2020002.pdf_