How I Built a Silver Price Prediction Model and Shipped it Live

The Problem

If you've ever tried to check silver prices in India, you know the pain — most global APIs give prices in USD per troy ounce, which means nothing to an Indian buyer who thinks in ₹ per gram. I wanted to build something that not only predicts tomorrow's silver price but also shows accurate Indian market rates with GST included.

🚀 Try it live: Silver Price Prediction App

What I Built

A full-stack ML application that does two things:

1. Fetches real-time silver prices and converts them to Indian market rates (₹ per gram, per 10g, per kg) — including import duties and 3% GST
2. Predicts next-day prices using a Lasso Regression model trained on historical data

How the Architecture Works

The system follows a simple three-stage pipeline:

┌─────────────────┐     ┌──────────────────┐     ┌─────────────────┐
│  MetalpriceAPI  │────▶│   Flask App      │────▶│   ML Model      │
│  (Live Prices)  │     │   (Conversion)   │     │   (Prediction)  │
└─────────────────┘     └──────────────────┘     └─────────────────┘
         │                       │                        │
         ▼                       ▼                        ▼
   USD/oz prices         INR conversion          Next-day forecast
                         + GST (3%)

The price conversion pipeline is where it gets interesting. You can't just multiply USD × exchange rate — Indian silver prices include layers of additional costs:

1. Fetch USD price per troy ounce from MetalpriceAPI
2. Convert using live USD/INR exchange rate
3. Convert troy ounce to grams (÷ 31.1035)
4. Add Import Duty (+6%)
5. Add Local Premium (+10%)
6. Add GST (+3%)

The result? Prices that match GoodReturns.in (Hyderabad rates) within ₹5-10 accuracy.

The ML Model

I experimented with several regression models and landed on Lasso Regression for its simplicity and strong performance:

The model uses a rich set of features engineered from historical price data:

• Lag features: Closing prices from 1, 2, 3, 5, and 7 days ago
• Moving averages: 5-day, 10-day, and 20-day windows
• Technical indicators: RSI, MACD, and Bollinger Bands

An R² of 0.98 means the model explains 98% of the variance in price movements — which is excellent for a regression task on financial data.

Data Source Strategy

One lesson I learned early: never rely on a single data source. APIs go down, rate limits kick in, and free tiers expire. So I built a three-tier fallback system:

The 24-hour caching layer ensures I'm not burning through API credits on every request.

API Design

The Flask app exposes two clean endpoints:

GET /api/predict         → Next-day price prediction
GET /api/current-price   → Current market price with GST

Sample response:

{
    "success": true,
    "market": "India",
    "currency": "INR",
    "with_gst": {
        "per_10_grams": 3650,
        "per_kg": 365000
    },
    "gst_rate": "3%"
}

Shipping to Production

Getting the model from my laptop to a live URL involved deploying to Hugging Face Spaces. The deployment is straightforward — push the Flask app with the trained model artifacts, and HF Spaces handles the rest.

# Clone and run locally
git clone https://github.com/danishsyed-dev/Silver-Price-Prediction.git
cd Silver-Price-Prediction
pip install -r requirements.txt
python app.py

Key Takeaways

1. Domain knowledge matters more than model complexity — Understanding how Indian silver pricing works (import duty, GST, local premiums) was more valuable than trying fancier ML models
2. Build fallback systems — External APIs will fail; plan for it from day one
3. Cache aggressively — 24-hour caching reduced my API costs to nearly zero
4. Ship early — Deploying to Hugging Face Spaces took less than 10 minutes and gave me a shareable URL instantly

⚠️ Disclaimer

This project is for educational purposes only. Actual silver prices at jewellers may include making charges (8-20%), wastage charges, purity variations (925, 999), and local market premiums. Do not use for actual trading decisions.