Overview

Qhawarina is a real-time economic monitoring platform for Peru that provides nowcasts of key macroeconomic variables using dynamic factor models (DFM) and high-frequency data. The platform combines traditional monthly indicators with daily scraped data from supermarkets and news sources to provide up-to-date economic intelligence.

Key Features

GDP Nowcasting

  • Dynamic factor model with 35+ monthly indicators
  • Regional disaggregation for 25 departments
  • Rolling 7-year window to handle structural breaks
  • Ridge regression bridge equations (α=1.0)

Inflation Monitoring

  • Daily price index based on Billion Prices Project methodology
  • 42,000+ products from Plaza Vea, Metro, and Wong supermarkets
  • Jevons bilateral chain-linked index
  • 25 monthly series in DFM (BCRP, MIDAGRI, supermarket data)

Poverty Nowcasting

  • GradientBoosting regressor for 24 departments
  • Monthly frequency estimates
  • NTL (nighttime lights) data integration
  • District-level spatial disaggregation

Political Risk Index

  • Daily RSS feed classification using Claude API with keyword fallback
  • EPU-style severity-weighted methodology
  • Separate political and economic instability indices
  • Real-time news monitoring from 81 Peruvian sources

Technical Stack

Backend: Python (statsmodels, scikit-learn, pandas, anthropic) Scraping: VTEX API, RSS feeds, BCRP API, MIDAGRI bulletins Frontend: Next.js 14, TypeScript, Tailwind CSS, Recharts Deployment: GitHub Pages (data), Vercel (website) Automation: Windows Task Scheduler (daily updates)

Data Sources

  • BCRP: 58 national series + 233 departmental series
  • INEI: GDP (quarterly), CPI (monthly), Poverty (annual)
  • Supermarkets: Daily prices via VTEX API
  • MIDAGRI: Wholesale food prices, poultry prices
  • News: 81 RSS feeds (El Comercio, La República, Gestión, etc.)
  • Satellite: VIIRS nighttime lights (monthly, 2012-2024)

Methodology

DFM Specification: Based on Giannone et al. (2008), Stock & Watson (2002)

  • EM algorithm for factor extraction with PCA fallback
  • Handles ragged edge via truncation (50% threshold)
  • COVID-exclusion filter (2020-2021) for post-pandemic stability

Price Index: Cavallo & Rigobon (2016) Billion Prices Project

  • Geometric mean of price ratios (Jevons formula)
  • Chain-linking with daily base updates
  • Extreme ratio filter: 0.5 < ratio < 2.0

Poverty Model: Change-prediction approach

  • Predict Δpoverty_t = poverty_t - poverty_{t-1}
  • GBR with dept-specific features from panel
  • Beats AR(1) benchmark (Rel.RMSE=0.953)

Performance

GDP Nowcast: RMSE=1.41pp (pre-COVID), Rel.RMSE=0.69 vs AR(1) Inflation Nowcast: RMSE=0.319% vs AR(1)=0.322% (3-month MA target) Poverty Nowcast: RMSE=2.54pp vs AR(1)=2.65pp Daily Price Index: 12 days of data, -0.57% cumulative since Feb 10 Political Index: 417 days, 24,541 articles classified

Future Work

  • Incorporate commodity prices (copper, gold, zinc)
  • Add employment nowcast using job postings data
  • Implement MIDAS regression for mixed-frequency models
  • Expand to other Latin American countries
  • Add forecast evaluation dashboard

References

  • Cavallo & Rigobon (2016). “The Billion Prices Project”, MIT
  • Giannone, Reichlin & Small (2008). “Nowcasting”, ECB Working Paper
  • Stock & Watson (2002). “Forecasting Using Principal Components”, JASA