Energy Futures Trading with Machine Learning & Event Detection

Executive Summary:

The commodity futures spectrum is an integral part of today’s financial markets. Specifically, trading energy futures such as crude oil, gasoline and natural gas, among many more, all react to the ebbs and flows of supply and demand.

Trading energy futures plays a crucial role in everyday life, as they fuel most of the world’s transportation system and they are the input to businesses across all the industrial sectors, hence they are inherently linked to the economic cycle.

From economic indicators such as gross domestic product and the unemployment rate to political upheaval and natural disasters, not to mention commodity-specific issues like oil and gas pipeline disruptions or mining accidents, they all contribute to the pricing of commodity futures.

Energy Futures Trading with Machine Learning Algorithms

For trading energy, we utilize five well-known machine learning algorithms to predict next day returns across a basket of energy commodities. The underlying signals of trading energy futures are driven by news events detected across thousands of sources. Each model is evaluated individually and as part of an ensemble strategy. We show how combining all models using machine-learning techniques produce solid risk-adjusted returns with lower average bias and without the need to select one particular model.

Energy Futures Trading: Findings

Specifically, we find that:

• Our ensemble portfolio provides an Information Ratio of 0.65 reducing the risk associated to model selection.
• By incorporating regimes which limit trading during high-volatility, we can improve on the out-of-sample return irrespective of whether we look at annualized, risk-adjusted, or per-trade return. IR climbs from 0.65 to 1.27 and annualized returns from 9.8% to 21.3% for the high-volatility strategy, despite the reduction in the number of trades. Moreover, we observe a more than 2x increase in per-trade returns from 3.88bp to 8.82bp.

White Paper: Energy Futures Trading with Machine Learning & Event Detection

Introduction

The commodity futures spectrum is an integral part of today’s financial markets. Specifically, energy related ones like crude oil, gasoline and natural gas, among many more, all react to the ebbs and flows of supply and demand. These commodities play a crucial role in everyday life, as they fuel most of the world’s transportation system and they are the input to businesses across all the industrial sectors, hence they are inherently linked to the economic cycle. From economic indicators such as gross domestic product and the unemployment rate to political upheaval and natural disasters, not to mention commodity-specific issues like oil and gas pipeline disruptions or mining accidents, they all contribute to the pricing of commodity futures.

In previous research Brandt and Gao took a novel approach by constructing supply and demand sentiment indices, using RavenPack data, to model the price impact of geopolitical and macroeconomic events and sentiments on crude oil. In particular, they found that news about macroeconomic fundamentals has a predictive ability over a monthly horizon, while geopolitical events sizably affected the price, but without sign predictability in the short term.

Rather than relying on a single commodity strategy, we seek to build predictive models for a group of commodities by means of RavenPack Analytics’ event detection capabilities. By utilizing RavenPack Analytics (RPA) 1.0, investors can benefit from the latest innovations in Natural Language Processing (NLP) technology to identify the information that matters for commodities. With the latest release, the RavenPack event taxonomy has grown to more than 6,800 event categories allowing the swift and precise identification of market-moving events across multiple asset classes and commodities. Events include supply increases, import/export guidance, inventory changes and more.

We select four commodity futures related to energy. We proceed to model the one-day ahead volatility-adjusted returns for the energy basket using an ensemble of machine learning techniques. Our results indicate that moving beyond linear models, including a wider spectrum of non-linear models, such as K-Nearest Neighbours (KNN) and Random Forest (RF) provides a way to achieve better performance and at the same time reduces the risk associated to model selection. Moreover, we demonstrate how return predictability at the basket level can be enhanced by conditioning on volatility regimes.

The paper is organized as follows: Section 2 discloses the different data sources used, in particular how the input variables from RPA 1.0 are constructed. Section 3 describes the modelling framework, which is based on five machine learning algorithms. Section 4 compares the performance of the various models introduced in Section 3. Finally, Section 5 presents the general conclusions.