evidoria

← Back to browse

Good practice

Food Inflation Nowcasting with Web-Scraped Prices — Narodowy Bank Polski's Big-Data Model for Real-Time CPI Estimates

Poland · Warsaw · See the Poland profile

Narodowy Bank Polski scrapes millions of online retail prices to nowcast food inflation ahead of official CPI releases. Peer-reviewed tests cut forecast error by up to 30% against standard models, with gains re-confirmed during COVID-19 and the 2022 food-price shock.

~30 %
Reduction in root-mean-square forecast error vs. best ARMA benchmark (combined forecasts) (backtest Jan 2014 - Jun 2018)
~27 %
Reduction in root-mean-square forecast error vs. best ARMA benchmark (ADL model alone) (backtest Jan 2014 - Jun 2018)
~159 million
Web-scraped prices in NBP eCPI database (end of 2020)
~250 million
Web-scraped prices in NBP eCPI database (2025 update)
~640,000
Products covered by NBP eCPI database (end of 2020)

Details

Maturity
Established
Promoter
Narodowy Bank Polski (National Bank of Poland), Economic Analysis and Research Department
Period
Method piloted 2019 (NBP WP 302); re-tested through the COVID-19 and 2022 food-price shocks; latest peer-reviewed update 2025
Keywords
monetary policy, inflation forecasting, big data, machine learning

Context

Narodowy Bank Polski (NBP), Poland's central bank, has since around 2010 collected high-frequency prices scraped from major online retailers, building an 'eCPI' database that held close to 159 million observed prices for roughly 640,000 products by the end of 2020, growing to around 250 million web-scraped prices by a 2025 peer-reviewed update. NBP's Economic Analysis and Research Department uses this data to nowcast food inflation ahead of the official Consumer Price Index published monthly by Statistics Poland (GUS).

Objectives

Test whether a web-scraped price index can improve the accuracy of short-term food-inflation forecasts compared with standard time-series benchmarks, and use it to nowcast food inflation before the official CPI release.

Activities

NBP Working Paper No. 302 (Macias & Stelmasiak, 2019) backtested the method against random-walk, seasonal random-walk and ARMA/SARMA benchmarks over a pseudo real-time window from January 2014 to June 2018, across 84 elementary COICOP groups and 10 food subaggregates. A 2023 follow-up in the International Journal of Forecasting and a 2025 study in the Journal of Forecasting re-tested the framework through the COVID-19 pandemic and the 2022 food-price shock following Russia's invasion of Ukraine.

Results

Combining the raw web-scraped price index with disaggregated autoregressive distributed-lag (ADL) models cut the root-mean-square forecast error by roughly 27% relative to the best ARMA benchmark, rising to about 30% when forecasts were combined — a reduction found statistically significant using the Diebold-Mariano test. The web-scraped index delivered its largest accuracy gains during the high-volatility COVID-19 and 2022 food-price-shock periods, with only marginal benefit in calmer times.

Conclusions

The published record documents a real, quantified and independently replicated improvement in short-term forecasting accuracy for one narrow but policy-relevant indicator. It does not establish that the eCPI nowcast has itself changed a specific interest-rate decision, and NBP has not published a citizen-facing version of the series.

Implementation

Indicative cost
Medium (€50k–€500k) — Not itemised in the source material; the effort represents an ongoing, multi-year research-department activity combining web-scraping infrastructure and economist time rather than a large capital programme.
Time to results
Long (> 3 years) — Price collection began around 2010; the nowcasting methodology was formalised and backtested by 2019 (NBP WP 302), then independently re-tested in 2023 and 2025 studies.
Staffing & skills
NBP Economic Analysis and Research Department economists and researchers, Data engineers maintaining the web-scraping pipeline

Conditions for success

  • Continuous access to a large volume of online retail price data
  • Rigorous benchmarking against standard forecasting models (random-walk, ARMA/SARMA)
  • Peer review and publication of methodology for external scrutiny

Common failure modes

  • Accuracy gains are only marginal during calmer, low-volatility periods
  • No published evidence that the nowcast has changed a specific policy decision

Where it fits

Governance type
independent central bank
Scale
national
Income level
high-income

Data sources

Where this practice's information was retrieved from, and when.

Similar practices you may find useful