An end-to-end analytics and ML platform that tracks central bank gold accumulation, measures USD dominance, quantifies geopolitical risk, and predicts which countries are likely to increase gold reserves next — combining structured macroeconomic data with NLP-derived financial narratives across 182 countries from 2000 to 2025.
→ Open Live Dashboard · → Case Study (PDF)
- $2.99 trillion in central bank gold holdings globally (2025) — up 21% in one year
- USD share of global reserves has declined from 71.1% (2001) → 56.9% (2025), falling every year since 2015
- Countries under OFAC sanctions hold proportionally more gold on average (25.4% vs 13.3% for non-sanctioned peers)
- Poland added 102 tonnes in 2025 alone — gold share jumped from 16.9% → 30.1% in 12 months
- Top 2026 predicted buyers: Belarus, Iraq, Libya, Uzbekistan, Qatar — scored on buying momentum, streak consistency, and geopolitical motivation
The platform follows a Bronze → Silver → Gold data engineering pattern with a separate NLP layer feeding into the final master panel.
| Layer | Path | Description |
|---|---|---|
| Bronze | data/raw/ |
Raw source files — WGC, IMF, World Bank, OFAC, UN |
| Silver | data/staging/ |
Cleaned, ISO3-coded, quarterly-aligned |
| Gold | data/curated/ |
Feature-engineered master panel (4,376 rows × 49 cols) |
| ML | data/curated/ml_* |
Scores, predictions, model outputs |
The PostgreSQL schema is documented in sql/schema.sql — a star schema with 5 tables, 1 denormalized analytical fact table, and a ranked accumulator view.
Goal: Predict which countries will increase gold reserves in 2026.
Approach: 4-Pillar Rule-Based Scoring Model, validated against Gradient Boosting and Logistic Regression baselines.
| Pillar | Weight | Signals |
|---|---|---|
| Physical Buying Momentum | 30% | Actual tonne changes YoY (price-neutral, from WGC data) |
| Buying Consistency | 25% | Accumulation streak + 5-year buying frequency |
| Geopolitical Motivation | 25% | UN divergence from US + sanctions exposure |
| Strategic Allocation Gap | 20% | Room to grow gold share relative to peers |
Coverage: 182 countries in the master panel → 72 scored (filters: ≥ $500M gold holdings, sufficient history, non-null geo data)
Model Performance (test set: 2020–2025)
| Model | Accuracy | Precision | Recall | F1 | AUC-ROC |
|---|---|---|---|---|---|
| Logistic Regression | 0.264 | 0.787 | 0.218 | 0.342 | 0.529 |
| Gradient Boosting | 0.402 | 0.861 | 0.378 | 0.525 | 0.561 |
| Ensemble | 0.256 | 0.800 | 0.201 | 0.321 | 0.528 |
Gradient Boosting achieves 0.86 precision — when it predicts a country will buy gold, it is correct 86% of the time. Training: 2001–2019 | Test: 2020–2025 | Predict: 2026.
| # | Country | Score | Key Driver |
|---|---|---|---|
| 1 | Belarus | 100.0 | 10-yr streak + max sanctions (level 2) + high geo risk |
| 2 | Iraq | 97.9 | +12t buying + 10-yr streak + UN divergence |
| 3 | Libya | 88.8 | 10-yr streak + sanctions + high geo risk |
| 4 | Uzbekistan | 86.3 | +7.8t buying + 12-yr streak + 86% gold share |
| 5 | Qatar | 86.0 | +4.4t buying + 11-yr streak |
| 6 | Algeria | 83.1 | 10-yr streak + high UN divergence score |
| 7 | China | 83.0 | +26.7t buying + 11-yr streak + high geo risk |
| 8 | Egypt | 77.1 | +2.5t buying + 10-yr streak |
| 9 | India | 71.0 | +4.2t buying + 10-yr streak |
| 10 | Lebanon | 70.4 | 10-yr streak + 82% gold share |
| Page | What it shows |
|---|---|
| Overview | Global gold map, top 15 holders, KPI cards — macro context at a glance |
| Gold vs USD | Gold trend vs USD dominance (dual-axis), Pearson correlation, OLS trendline |
| Geopolitics | Sanctions scoring, UN alignment scatter, geo bloc analysis |
| Sentiment | Live NLP feed — gold & USD sentiment signals from GDELT/Reuters |
| ML Predictions | 2026 country-level scores — 4-pillar breakdown, model metrics, coverage funnel |
| Source | Data | Coverage |
|---|---|---|
| World Gold Council / IMF IFS | Gold holdings in tonnes | 182 countries |
| IMF COFER | USD share of global reserves | Global aggregate |
| World Bank API | Total reserves, GDP, macro indicators | 182 countries |
| OFAC | Sanctions severity (0–3) | All countries |
| UN General Assembly | Voting divergence from US | 182 countries |
| GDELT / News APIs | Financial news sentiment (NLP) | Live feed |
Run the refresh script to pull the latest data from all automated sources and re-run the full pipeline:
python scripts/refresh_data.pyWhat it does automatically:
| Source | Frequency | Method |
|---|---|---|
| World Bank API | Monthly | Auto-fetched (REST API) |
| IMF COFER | Quarterly | Auto-fetched (API + curated fallback) |
| UN Votes | Annual | Auto-fetched (Harvard Dataverse) |
| OFAC Sanctions | As needed | Auto-fetched (US Treasury CSV) |
| GDELT News | Daily | Auto-fetched (GDELT DOC 2.0 API) |
| World Gold Council | Monthly | Manual — download from gold.org |
Useful flags:
python scripts/refresh_data.py --check # staleness check only, no fetching
python scripts/refresh_data.py --no-pipeline # fetch data but skip pipeline re-run
python scripts/refresh_data.py --source gdelt # refresh a single source (wb|cofer|un|ofac|gdelt)The script tracks last-fetch timestamps in data/.refresh_meta.json and warns when any source is stale.
Python · pandas · numpy · scikit-learn · XGBoost · HuggingFace FinBERT
PostgreSQL · SQLAlchemy · psycopg2
Streamlit · Plotly
Docker · docker-compose
# Clone and install
git clone https://github.com/sathwikarr/gold-reserve-platform.git
cd gold-reserve-platform
pip install -r requirements.txt
# Run the full pipeline (re-processes all data + ML)
python run_pipeline.py
# Launch the Streamlit app
streamlit run app.pycp .env.example .env # fill in your settings
docker-compose up --build # starts app + postgres at localhost:8501export DATABASE_URL=postgresql://user:pass@localhost:5432/gold_reserve_db
psql $DATABASE_URL -f sql/schema.sql # create tables and view
python src/db/load_to_postgres.py # load curated data
psql $DATABASE_URL -f sql/queries.sql # run analytical queriesgold-reserve-platform/
├── app.py # Streamlit app (5 analytical pages)
├── run_pipeline.py # Single command re-runs full pipeline
│
├── src/
│ ├── ingestion/ # Raw data fetchers (WGC, IMF, World Bank)
│ ├── cleaning/ # Standardisation and ISO3 alignment
│ ├── features/ # Feature engineering and master panel build
│ ├── nlp/ # NLP sentiment merge
│ ├── ml/ # Scoring, training, and 2026 predictions
│ └── db/ # PostgreSQL ETL loader
│
├── sql/
│ ├── schema.sql # Star schema — 5 tables, 1 view, indexes
│ └── queries.sql # 10 analytical queries
│
├── data/
│ ├── raw/ # Bronze: original source files
│ ├── staging/ # Silver: cleaned, standardised
│ └── curated/ # Gold: master panel + ML outputs
│ ├── master_panel_nlp.csv # 4,376 rows × 49 cols, 182 countries
│ ├── ml_country_scores.csv # 72 scored countries with pillar breakdown
│ └── ml_top10_predictions.csv
│
├── notebooks/
│ ├── 01_eda_gold_reserves.ipynb
│ ├── 02_eda_usd_vs_gold.ipynb
│ ├── 03_eda_geopolitical.ipynb
│ └── 04_ml_model_training.ipynb
│
├── docs/
│ ├── architecture.png
│ ├── ml_top10_predictions.png
│ ├── ml_feature_importance.png
│ └── ml_roc_curves.png
│
├── Dockerfile
├── docker-compose.yml
└── requirements.txt
MIT © 2026 Sathwik Arroju