Gold Return and Directional Predictability with Machine Learning

This repository contains the code used in the study Gold Return and Directional Predictability with Machine Learning.
The project investigates whether gold futures returns and price direction can be forecast over short- to medium-term horizons using machine learning models and a broad set of financial, macroeconomic, and geopolitical predictors.

All models are evaluated under a strict yearly walk-forward (fully out-of-sample) framework, ensuring realistic forecasting conditions and preventing information leakage.

---

Project Overview

The analysis evaluates gold futures return and directional predictability over 7-, 14-, 21-, and 28-day forecast horizons using a yearly walk-forward evaluation design.

This framework closely mimics real-time forecasting conditions by ensuring that:

• models are trained only on past information,
• test data are strictly out-of-sample,
• and information leakage is avoided throughout preprocessing and model selection.

Key Findings

• Gold return predictability is limited and highly state-dependent
• Directional predictability emerges primarily at the 28-day horizon
• Predictability is stronger during periods of market instability and frequent regime switching
• Ensemble methods and regime-aware modeling provide economically meaningful improvements

---

Methods

Models

The following classes of models are implemented and compared:

• Linear models

  • Ridge Regression

• Tree-based ensembles

  • Random Forest
  • Extremely Randomized Trees (Extra Trees)

• Boosting methods

  • Gradient Boosting
  • XGBoost
  • CatBoost

• Neural networks

  • Multi-Layer Perceptron (MLP)

Forecast Tasks and Evaluation

Two prediction tasks are considered at each forecast horizon:

1. Return prediction

   • Continuous cumulative returns
   • Evaluated using out-of-sample ( R^2 )

2. Directional prediction

   • Binary up/down price movement
   • Evaluated using Accuracy, Balanced Accuracy, and skill scores relative to naïve benchmarks

---

Data

• Gold futures: GC=F (Yahoo Finance)
• Financial and macroeconomic predictors: Yahoo Finance
• Geopolitical Risk (GPR) Index: Caldara and Iacoviello (2022)
• Frequency: Daily
• Sample period: January 2010 – October 2025

Note: Yahoo Finance data may be revised over time.
For exact replication, results should be generated using the datasets saved in the data/ directory.

---

Implementation Details

• Models are trained using a rolling yearly walk-forward procedure
• Each model is estimated using a four-year rolling training window
• Hyperparameter tuning and model comparison are conducted using the PyCaret framework with time-series-aware cross-validation
• Ensemble forecasts are constructed using:
  • Averaging for regression models
  • Majority voting for classification models
• Numerical results and figures are automatically exported for further analysis

---

Execution Order

The project is implemented as a single, self-contained Jupyter notebook.
All code must be executed sequentially from top to bottom to reproduce the results.

Step 1 — Environment Setup

• Import required Python libraries
• Set global configuration options and random seeds

---

Step 2 — Data Acquisition and Preprocessing

• Download gold futures (GC=F) and related financial series from Yahoo Finance
• Load the Geopolitical Risk (GPR) Index
• Align time series and handle missing values
• Save the final merged dataset to the data/ directory

---

Step 3 — Feature Construction and Target Definition

• Construct lagged financial, macroeconomic, and technical predictors
• Standardize and transform features where appropriate
• Construct return targets for 7-, 14-, 21-, and 28-day horizons
• Create binary directional labels
• Perform exploratory analysis (target–feature and feature–feature relationships)

---

Step 4 — Regression Analysis

4.1 Walk-Forward Evaluation

• Define yearly walk-forward splits
• Specify rolling four-year training windows
• Ensure all preprocessing is performed using training data only
• Train regression models using the PyCaret framework
• Perform time-series-aware three-fold cross-validation within each training window
• Generate out-of-sample return forecasts
• Save evaluation tables to the result/ directory
• Generate figures of model–target ( R^2 ) over the full test period

4.2 Ensemble Construction

• Combine individual model forecasts using averaging
• Evaluate all model combinations consisting of at least two models
• Identify ensemble combinations with improved predictive performance

4.3 Regime-Based Analysis

• Compute yearly out-of-sample ( R^2 ) for each model and forecast horizon
• Identify periods of elevated predictability (e.g., 2018, 2022)
• Apply K-means clustering to classify daily market regimes
• Compute normalized entropy as a measure of regime instability
• Relate high regime-switching frequency to forecast performance
• Determine the optimal entropy threshold using the F1 score

---

Step 5 — Classification Analysis

5.1 Walk-Forward Evaluation

• Define yearly walk-forward splits
• Specify rolling four-year training windows
• Ensure all preprocessing is performed using training data only
• Train classification models using the PyCaret framework
• Perform time-series-aware three-fold cross-validation
• Generate out-of-sample directional forecasts
• Evaluate Accuracy and Balanced Accuracy
• Compare model performance against naïve benchmarks
• Save evaluation tables to the result/ directory

5.2 Ensemble Construction

• Combine individual classification models using majority voting
• Evaluate all ensemble combinations with at least two models
• Observe performance improvements, particularly for the 28-day horizon

---

Important:
Cells must be run in order. Skipping or reordering cells may result in errors or inconsistent results.

---

Repository Structure

├── gold_prediction_code.ipynb   # Main notebook (full pipeline)
├── data/                        # Input datasets
├── result/                      # Exported tables and evaluation metrics
├── plot/                        # Exported figures
├── requirements.txt             # Python dependencies
└── README.md