Skip to content
Front pageRaman Lamba

Finance · ML

Gold Price Forecasting

Solo project

A probabilistic CNN-LSTM that forecasts gold under uncertainty and drives a risk-based trading strategy.

Key figures

By the numbers

  • 54-58%directional accuracy, 5-day
  • 18%backtested strategy return
  • 3,939trading days

The problem

Forecast short-term gold prices under uncertainty, then turn the forecasts into trading decisions.

The approach

A probabilistic CNN-LSTM trained on 3,939 trading days, 9 macroeconomic indicators and more than 45 engineered features, with quantile models and checks on calibration and uncertainty. A risk-based strategy turns the model outputs into buy and sell signals.

The result

54-58% directional accuracy on 5-day forecasts and an 18% backtested strategy return. Feature-importance analysis looked at interest rates, the USD index, inflation and crude oil to see which indicators were most predictive.

Built with

Figures

Four charts: residual distribution, residuals over time, a normal Q-Q plot and uncertainty width against actual error.
Residual diagnostics: error distribution, errors over time, a normal Q-Q plot, and predicted uncertainty width against actual error.
Five-day price targets from the model, from a bearish 10th percentile to a bullish 90th percentile.
Five-day price targets from the probabilistic model, from the 10th to the 90th percentile.

Source code on GitHub